{"site":"https://rsiratchet.com","generated_at":"2026-10-24","license":{"code":"MIT","content":"CC BY 4.0"},"ref_lookup":"https://rsiratchet.com/id/<ref>","counts":{"capabilities":50,"techniques":340,"claims":657,"sources":686,"heuristics":25},"capabilities":[{"ref":"CAP-0026","id":"knowledge-updating","label":"Absorbing and flexibly using new knowledge","summary":"Covers whether newly injected or post-training facts can be recalled under paraphrase, combined, and used to override stale beliefs; excludes retrieval-based grounding at inference time.","group":"knowledge","status":"active","aliases":[],"tags":["general"],"parent":null,"counts":{"claims":6,"contested":0,"techniques":4,"sources":7},"url":"https://rsiratchet.com/capabilities/knowledge-updating"},{"ref":"CAP-0025","id":"judge-position-bias","label":"Biased when judging other outputs","summary":"A model used as a judge favors whichever answer is shown first, longer answers, and its own outputs.","group":"verification","status":"active","aliases":["LLM-as-judge bias","position bias","verbosity bias","self-enhancement bias"],"tags":["evaluation","llm-as-judge"],"parent":null,"counts":{"claims":3,"contested":0,"techniques":2,"sources":2},"url":"https://rsiratchet.com/capabilities/judge-position-bias"},{"ref":"CAP-0032","id":"multi-step-reasoning","label":"Carrying a chain of inference through to its conclusion","summary":"Whether a model can hold several dependent inference steps and reach the conclusion they entail, as opposed to pattern-matching the answer; failures show as correct steps with a wrong join, or shortcuts.","group":"reasoning","status":"active","aliases":[],"tags":["reasoning","general"],"parent":null,"counts":{"claims":11,"contested":0,"techniques":8,"sources":14},"url":"https://rsiratchet.com/capabilities/multi-step-reasoning"},{"ref":"CAP-0035","id":"persuasion","label":"Changing what a person believes or decides","summary":"How effectively a model shifts a person's beliefs or choices in conversation, relative to a human doing the same; the risk is the flip side of the capability, a system that can talk its operators out of constraining it.","group":"behavior","status":"active","aliases":["persuasiveness","influence","manipulation"],"tags":["behavior","chat-assistant"],"parent":null,"counts":{"claims":7,"contested":0,"techniques":2,"sources":8},"url":"https://rsiratchet.com/capabilities/persuasion"},{"ref":"CAP-0016","id":"fact-verification","label":"Checking claims against evidence","summary":"Models accept confident false premises and resolve conflicts between sources by majority or recency rather than by reliability.","group":"knowledge","status":"active","aliases":["epistemic verification","truth adjudication","knowledge conflict"],"tags":["chat-assistant","data-analysis","knowledge","rag-qa"],"parent":null,"counts":{"claims":10,"contested":0,"techniques":4,"sources":9},"url":"https://rsiratchet.com/capabilities/fact-verification"},{"ref":"CAP-0010","id":"domain-knowledge","label":"Competence in a specialist domain","summary":"Whether a general model performs like a practitioner in a specific field (medicine, law, hardware design, a scientific subfield), and where the general-purpose competence stops.","group":"knowledge","status":"active","aliases":[],"tags":["knowledge"],"parent":null,"counts":{"claims":10,"contested":0,"techniques":5,"sources":10},"url":"https://rsiratchet.com/capabilities/domain-knowledge"},{"ref":"CAP-0029","id":"long-horizon-execution","label":"Completing tasks that take many steps or many hours","summary":"Whether an agent can carry a task through a long sequence of actions without losing the thread, compounding small errors, or stopping early; measured by how long a task, in human time or in steps, a model can finish reliably.","group":"agentic","status":"active","aliases":[],"tags":["agentic","autonomous-agent","coding-agent"],"parent":null,"counts":{"claims":23,"contested":0,"techniques":17,"sources":26},"url":"https://rsiratchet.com/capabilities/long-horizon-execution"},{"ref":"CAP-0053","id":"cutting-to-the-chase","label":"Cutting to the chase","summary":"Giving the answer directly when that's what is asked, without preamble, restatement or visible working the reader didn't request; the failure is padding the model was told to skip, or skipping working the task needed.","group":"behavior","status":"proposed","aliases":["answer-only compliance","no-thinking","concision","verbosity","overthinking"],"tags":["behavior","chat-assistant"],"parent":null,"counts":{"claims":4,"contested":0,"techniques":3,"sources":3},"url":"https://rsiratchet.com/capabilities/cutting-to-the-chase"},{"ref":"CAP-0027","id":"logical-deduction-robustness","label":"Deducing validly from logical structure","summary":"Covers strict rule-based deduction, stability under label-preserving rephrasings, and retracting conclusions when evidence changes; excludes arithmetic and factual QA.","group":"reasoning","status":"active","aliases":[],"tags":["general"],"parent":null,"counts":{"claims":3,"contested":0,"techniques":1,"sources":3},"url":"https://rsiratchet.com/capabilities/logical-deduction-robustness"},{"ref":"CAP-0002","id":"arithmetic","label":"Digit-level arithmetic","summary":"Multi-digit arithmetic, counting, and symbol manipulation fail in ways tied to tokenization and memorization.","group":"reasoning","status":"active","aliases":["numerical reasoning","token-level math"],"tags":["data-analysis","general","math","reasoning"],"parent":null,"counts":{"claims":11,"contested":1,"techniques":3,"sources":9},"url":"https://rsiratchet.com/capabilities/arithmetic"},{"ref":"CAP-0008","id":"debugging","label":"Finding and fixing a fault in existing code","summary":"Whether a model can localise the cause of a failing behaviour in code it did not just write and repair it without breaking something else; distinct from generating code and from self-repair of its own output.","group":"coding","status":"active","aliases":[],"tags":["coding-agent","agentic"],"parent":null,"counts":{"claims":4,"contested":0,"techniques":1,"sources":6},"url":"https://rsiratchet.com/capabilities/debugging"},{"ref":"CAP-0039","id":"repository-navigation","label":"Finding what matters in a large codebase","summary":"Whether an agent can locate the files, symbols and call paths relevant to a task in a repository too large to read, and avoid editing the wrong place; the failure is a plausible patch in the wrong file.","group":"coding","status":"active","aliases":["issue localization","fault localization","code localization","bug localization"],"tags":["coding-agent","context"],"parent":null,"counts":{"claims":6,"contested":0,"techniques":6,"sources":6},"url":"https://rsiratchet.com/capabilities/repository-navigation"},{"ref":"CAP-0044","id":"self-repair","label":"Fixing its own mistakes","summary":"Whether and how a model can diagnose and correct a wrong answer once it's already made.","group":"agentic","status":"active","aliases":[],"tags":["agentic","coding-agent","math"],"parent":null,"counts":{"claims":15,"contested":2,"techniques":6,"sources":17},"url":"https://rsiratchet.com/capabilities/self-repair"},{"ref":"CAP-0037","id":"procedure-following","label":"Following an unfamiliar procedure","summary":"Models drift from a new multi-step procedure or set of constraints back to familiar routines.","group":"agentic","status":"active","aliases":["instruction following","constraint adherence","workflow compliance"],"tags":["agentic","autonomous-agent","chat-assistant","coding-agent","customer-support"],"parent":null,"counts":{"claims":16,"contested":1,"techniques":3,"sources":18},"url":"https://rsiratchet.com/capabilities/procedure-following"},{"ref":"CAP-0038","id":"prompt-injection","label":"Following instructions hidden in data","summary":"Models treat instructions found in retrieved documents, tool outputs, or web pages as if they came from the user.","group":"security","status":"active","aliases":["indirect prompt injection","jailbreak via content"],"tags":["autonomous-agent","coding-agent","customer-support","rag-qa","security"],"parent":null,"counts":{"claims":19,"contested":0,"techniques":6,"sources":20},"url":"https://rsiratchet.com/capabilities/prompt-injection"},{"ref":"CAP-0020","id":"formal-math","label":"Formalizing and machine-checking mathematics","summary":"Covers autoformalization into a solver-checkable language, producing proof-assistant-verified proofs, and localizing the first flawed step in a proof; excludes informal math word problems.","group":"reasoning","status":"active","aliases":[],"tags":["general"],"parent":null,"counts":{"claims":5,"contested":0,"techniques":2,"sources":5},"url":"https://rsiratchet.com/capabilities/formal-math"},{"ref":"CAP-0054","id":"reward-hacking","label":"Gaming the measure instead of doing the task","summary":"Whether a model trained against a reward or a monitor learns the intended behaviour or learns to score well some other way, including ways the monitor cannot see.","group":"behavior","status":"proposed","aliases":["reward hacking","specification gaming","reward tampering","Goodhart"],"tags":["behavior","security"],"parent":null,"counts":{"claims":3,"contested":0,"techniques":2,"sources":2},"url":"https://rsiratchet.com/capabilities/reward-hacking"},{"ref":"CAP-0004","id":"code-generation","label":"Generating and editing working code","summary":"Whether a model can write correct code from a specification, and go beyond isolated functions to real repository-scale editing and debugging.","group":"coding","status":"active","aliases":[],"tags":["reasoning","agentic"],"parent":null,"counts":{"claims":21,"contested":0,"techniques":9,"sources":26},"url":"https://rsiratchet.com/capabilities/code-generation"},{"ref":"CAP-0031","id":"multi-model-coordination","label":"Getting more from several models than from one","summary":"Whether a system of several models or agents — debating, voting, dividing work, or sharing state — does better than its best single member; the failure is a quorum of copies that agree on the same mistake.","group":"agentic","status":"active","aliases":["multi-agent systems","ensembles","model quorum","mixture of agents","LLM debate"],"tags":["agentic","autonomous-agent"],"parent":null,"counts":{"claims":38,"contested":0,"techniques":32,"sources":42},"url":"https://rsiratchet.com/capabilities/multi-model-coordination"},{"ref":"CAP-0051","id":"visual-grounding","label":"Grounding answers in the image","summary":"Saying only what the image supports, rather than describing plausible objects that are not there.","group":"perception","status":"active","aliases":[],"tags":["knowledge","behavior"],"parent":null,"counts":{"claims":29,"contested":0,"techniques":16,"sources":26},"url":"https://rsiratchet.com/capabilities/visual-grounding"},{"ref":"CAP-0007","id":"cultural-competence","label":"Handling culture-specific norms and viewpoints","summary":"Covers implicit cultural norms, dialects, local pragmatics, and representing the spread of a population's views; excludes general multilingual fluency or factual trivia about cultures.","group":"behavior","status":"active","aliases":[],"tags":["general"],"parent":null,"counts":{"claims":5,"contested":0,"techniques":2,"sources":6},"url":"https://rsiratchet.com/capabilities/cultural-competence"},{"ref":"CAP-0023","id":"harness-self-improvement","label":"Improving its own scaffolding from failures","summary":"Whether an agent system can read its own failure traces and rewrite the prompts, tools and policies it runs inside so later runs do better, with no person making the edit.","group":"agentic","status":"active","aliases":["self-improving agents","harness optimization","scaffold evolution","prompt optimization"],"tags":["agentic","autonomous-agent","coding-agent"],"parent":null,"counts":{"claims":29,"contested":1,"techniques":24,"sources":35},"url":"https://rsiratchet.com/capabilities/harness-self-improvement"},{"ref":"CAP-0006","id":"critique","label":"Judging another's work as well as a careful reviewer would","summary":"The model as critic or judge: whether its evaluation of an artifact agrees with careful human judgement and is actionable; the failures are bias toward surface features, agreement with the author, and vague critique that sounds specific.","group":"verification","status":"active","aliases":["LLM-as-a-judge","model as critic","critique generation","reward modelling"],"tags":["evaluation","llm-as-judge"],"parent":null,"counts":{"claims":12,"contested":0,"techniques":8,"sources":13},"url":"https://rsiratchet.com/capabilities/critique"},{"ref":"CAP-0005","id":"context-hygiene","label":"Keeping its own context clean","summary":"Agents fill their working context with re-read files, restated notes, and excess documentation until quality degrades.","group":"context","status":"active","aliases":["context management","over-documentation","context bloat"],"tags":["autonomous-agent","coding-agent","context"],"parent":"long-context-degradation","counts":{"claims":4,"contested":0,"techniques":2,"sources":4},"url":"https://rsiratchet.com/capabilities/context-hygiene"},{"ref":"CAP-0028","id":"long-context-degradation","label":"Losing information in long inputs","summary":"Accuracy drops as inputs grow and is worst for information placed in the middle of the context.","group":"context","status":"active","aliases":["lost in the middle","needle in a haystack","effective context length"],"tags":["coding-agent","context","long-document","rag-qa"],"parent":null,"counts":{"claims":14,"contested":1,"techniques":6,"sources":14},"url":"https://rsiratchet.com/capabilities/long-context-degradation"},{"ref":"CAP-0045","id":"software-architecture","label":"Making design decisions that hold up as a system grows","summary":"Whether a model can choose structure — module boundaries, interfaces, data flow — that stays maintainable, as opposed to code that passes the tests today; the failure is locally correct code that raises the cost of every later change.","group":"coding","status":"active","aliases":[],"tags":["coding-agent"],"parent":null,"counts":{"claims":3,"contested":0,"techniques":0,"sources":3},"url":"https://rsiratchet.com/capabilities/software-architecture"},{"ref":"CAP-0041","id":"reversal-curse","label":"Not generalizing \"A is B\" to \"B is A\"","summary":"A model that learned a fact in one direction often cannot recall it when asked in the reverse direction.","group":"knowledge","status":"active","aliases":["directional recall","asymmetric knowledge"],"tags":["general","knowledge"],"parent":null,"counts":{"claims":3,"contested":0,"techniques":1,"sources":2},"url":"https://rsiratchet.com/capabilities/reversal-curse"},{"ref":"CAP-0033","id":"multilingual-disparity","label":"Performing evenly across languages","summary":"Covers gaps between high- and low-resource languages on identical tasks, including uneven parametric knowledge coverage by region; excludes translation quality per se.","group":"behavior","status":"active","aliases":[],"tags":["general"],"parent":null,"counts":{"claims":1,"contested":0,"techniques":0,"sources":1},"url":"https://rsiratchet.com/capabilities/multilingual-disparity"},{"ref":"CAP-0036","id":"planning","label":"Planning a sequence of actions before acting","summary":"Whether a model can lay out an ordered set of steps toward a goal, with the dependencies right, before or instead of executing them; distinct from carrying the plan out.","group":"reasoning","status":"active","aliases":[],"tags":["reasoning","autonomous-agent"],"parent":null,"counts":{"claims":13,"contested":0,"techniques":13,"sources":16},"url":"https://rsiratchet.com/capabilities/planning"},{"ref":"CAP-0019","id":"forecasting","label":"Predicting future events","summary":"Whether a model can make well-calibrated predictions about events that haven't happened yet, not just recall or verify existing facts.","group":"reasoning","status":"active","aliases":[],"tags":["reasoning","knowledge"],"parent":null,"counts":{"claims":16,"contested":0,"techniques":9,"sources":16},"url":"https://rsiratchet.com/capabilities/forecasting"},{"ref":"CAP-0021","id":"goal-conflict-safety","label":"Prioritizing safety under conflicting goals","summary":"When a task goal competes with a safety rule or a required check, models sometimes drop the check to finish the task.","group":"agentic","status":"active","aliases":["critical task prioritization","oversight subversion","guardrail bypass"],"tags":["agentic","autonomous-agent","coding-agent","customer-support"],"parent":null,"counts":{"claims":8,"contested":0,"techniques":3,"sources":7},"url":"https://rsiratchet.com/capabilities/goal-conflict-safety"},{"ref":"CAP-0050","id":"video-temporal-reasoning","label":"Reasoning about time in video","summary":"Tracking motion, ordering, causality, and change across video frames rather than treating video as isolated stills.","group":"perception","status":"active","aliases":[],"tags":["reasoning","context"],"parent":null,"counts":{"claims":19,"contested":0,"techniques":10,"sources":20},"url":"https://rsiratchet.com/capabilities/video-temporal-reasoning"},{"ref":"CAP-0017","id":"factual-recall","label":"Recalling facts it was trained on","summary":"Whether a model retrieves a fact from its weights when asked, especially long-tail facts; the failure is a wrong or missing answer, measured against ground truth, not the confidence of a made-up one.","group":"knowledge","status":"active","aliases":[],"tags":["knowledge","general"],"parent":null,"counts":{"claims":4,"contested":0,"techniques":1,"sources":5},"url":"https://rsiratchet.com/capabilities/factual-recall"},{"ref":"CAP-0024","id":"jailbreak-robustness","label":"Refusing harmful requests under obfuscation","summary":"Covers maintaining refusal when harmful intent is wrapped in roleplay or encodings, harm that emerges only from combined segments, and unsafe content inside reasoning traces; excludes instructions injected via retrieved data.","group":"security","status":"active","aliases":[],"tags":["general"],"parent":null,"counts":{"claims":3,"contested":0,"techniques":1,"sources":4},"url":"https://rsiratchet.com/capabilities/jailbreak-robustness"},{"ref":"CAP-0030","id":"long-term-memory","label":"Remembering across sessions","summary":"Assistants fail to recall and apply facts, preferences, and decisions from earlier sessions.","group":"context","status":"active","aliases":["persistent memory","cross-session memory","personalization"],"tags":["autonomous-agent","chat-assistant","context","customer-support"],"parent":null,"counts":{"claims":11,"contested":0,"techniques":5,"sources":12},"url":"https://rsiratchet.com/capabilities/long-term-memory"},{"ref":"CAP-0018","id":"fine-tuning-retention","label":"Retaining skills and safety after fine-tuning","summary":"Covers loss of general capability, out-of-domain skill, refusal behavior, or output diversity caused by domain, multimodal, alignment or RL post-training; excludes in-domain accuracy gains and inference-time behavior shifts.","group":"behavior","status":"active","aliases":["catastrophic forgetting","alignment tax","mode collapse","output diversity","exploration collapse"],"tags":["general"],"parent":null,"counts":{"claims":8,"contested":0,"techniques":7,"sources":11},"url":"https://rsiratchet.com/capabilities/fine-tuning-retention"},{"ref":"CAP-0001","id":"abstention-and-self-knowledge","label":"Saying it does not know","summary":"Covers detecting unanswerable or underspecified problems and gaps between what hidden states encode and what the model asserts; excludes ordinary accuracy on solvable problems.","group":"verification","status":"active","aliases":["abstention","knowing what it does not know","unanswerable questions","selective prediction","calibration"],"tags":["general"],"parent":null,"counts":{"claims":25,"contested":0,"techniques":19,"sources":23},"url":"https://rsiratchet.com/capabilities/abstention-and-self-knowledge"},{"ref":"CAP-0015","id":"explanation-faithfulness","label":"Saying what actually drove the answer","summary":"Whether a model's stated reasoning reflects the computation that produced its answer, rather than a plausible story told after the fact.","group":"behavior","status":"active","aliases":[],"tags":["behavior","evaluation"],"parent":null,"counts":{"claims":8,"contested":0,"techniques":4,"sources":7},"url":"https://rsiratchet.com/capabilities/explanation-faithfulness"},{"ref":"CAP-0013","id":"error-detection","label":"Spotting an error in a given artifact","summary":"Whether a model can find the mistake in something put in front of it — a proof, a program, an argument, a table — when told there may be one; distinct from fixing it, and from judging overall quality.","group":"verification","status":"active","aliases":[],"tags":["evaluation","llm-as-judge"],"parent":null,"counts":{"claims":8,"contested":0,"techniques":5,"sources":8},"url":"https://rsiratchet.com/capabilities/error-detection"},{"ref":"CAP-0022","id":"hallucination","label":"Stating false facts confidently","summary":"Models generate plausible but false statements, citations, and details, with no signal that they are unsure.","group":"knowledge","status":"active","aliases":["confabulation","fabrication","factual error","making things up","invented citations"],"tags":["chat-assistant","general","knowledge","rag-qa"],"parent":null,"counts":{"claims":57,"contested":0,"techniques":21,"sources":83},"url":"https://rsiratchet.com/capabilities/hallucination"},{"ref":"CAP-0047","id":"strategic-deception","label":"Strategic deception and detecting it","summary":"Whether a model can deliberately induce false beliefs in others when a task calls for it, and detect deception from others.","group":"behavior","status":"active","aliases":[],"tags":["behavior"],"parent":null,"counts":{"claims":7,"contested":0,"techniques":3,"sources":6},"url":"https://rsiratchet.com/capabilities/strategic-deception"},{"ref":"CAP-0048","id":"sycophancy","label":"Telling the user what they want to hear","summary":"Models shift answers toward the user's stated view, cave to pushback, and mirror mistaken premises.","group":"behavior","status":"active","aliases":["flattery","agreement bias","caving to pushback"],"tags":["behavior","chat-assistant","customer-support","llm-as-judge"],"parent":null,"counts":{"claims":15,"contested":0,"techniques":3,"sources":11},"url":"https://rsiratchet.com/capabilities/sycophancy"},{"ref":"CAP-0046","id":"state-tracking","label":"Tracking state through a long task","summary":"Models lose track of the current state of entities, files, or an environment after many steps.","group":"context","status":"active","aliases":["entity tracking","world-model consistency","stale state"],"tags":["autonomous-agent","coding-agent","context","data-analysis"],"parent":null,"counts":{"claims":5,"contested":0,"techniques":3,"sources":5},"url":"https://rsiratchet.com/capabilities/state-tracking"},{"ref":"CAP-0011","id":"embodied-control","label":"Turning plans into physical action","summary":"Closing the loop from perception to action in a body or simulator, where mistakes have physical consequences and cannot be retried freely.","group":"agentic","status":"active","aliases":[],"tags":["agentic","reasoning"],"parent":null,"counts":{"claims":37,"contested":0,"techniques":24,"sources":37},"url":"https://rsiratchet.com/capabilities/embodied-control"},{"ref":"CAP-0003","id":"audio-understanding","label":"Understanding speech and audio","summary":"Hearing what is said and how it is said — tone, emotion, speaker identity, and events in a soundscape.","group":"perception","status":"active","aliases":[],"tags":["knowledge","reasoning"],"parent":null,"counts":{"claims":10,"contested":0,"techniques":4,"sources":11},"url":"https://rsiratchet.com/capabilities/audio-understanding"},{"ref":"CAP-0049","id":"tool-use","label":"Using the tools it is given","summary":"Models often fail to reach for provided tools, skills, or documented procedures, call them with wrong arguments, or act on a stale or misread result.","group":"agentic","status":"active","aliases":["tool calling","function calling","API use","skill utilization","computer use","environment interaction","acting on an environment and reading what came back"],"tags":["agentic","autonomous-agent","coding-agent","customer-support"],"parent":null,"counts":{"claims":36,"contested":1,"techniques":18,"sources":38},"url":"https://rsiratchet.com/capabilities/tool-use"},{"ref":"CAP-0040","id":"retrieval-utilization","label":"Using what was retrieved, and only that","summary":"Given relevant passages in context, whether a model actually uses them, prefers them over its own prior when they conflict, and ignores the irrelevant ones; the failure is an answer the passages do not support.","group":"context","status":"active","aliases":[],"tags":["context","rag-qa"],"parent":null,"counts":{"claims":1,"contested":0,"techniques":0,"sources":1},"url":"https://rsiratchet.com/capabilities/retrieval-utilization"},{"ref":"CAP-0014","id":"evaluation-validity","label":"Whether the measurement made the finding","summary":"Whether a reported result describes the model or the way it was scored — the metric, the prompt format, the contamination, the baseline.","group":"evaluation","status":"active","aliases":[],"tags":["evaluation"],"parent":null,"counts":{"claims":26,"contested":0,"techniques":6,"sources":32},"url":"https://rsiratchet.com/capabilities/evaluation-validity"},{"ref":"CAP-0052","id":"formal-query-synthesis","label":"Writing a reusable locator: regex, SQL, XPath","summary":"Whether a model can turn an intent or a few examples into a grep pattern, regex, SQL query, XPath or jq path that selects the right subset, so the pattern runs over the whole collection without calling the model again.","group":"coding","status":"active","aliases":["locator generation","pattern synthesis","text-to-SQL","regex generation","programming by example","grep","grep-style locators"],"tags":["coding-agent","general"],"parent":null,"counts":{"claims":17,"contested":0,"techniques":15,"sources":17},"url":"https://rsiratchet.com/capabilities/formal-query-synthesis"},{"ref":"CAP-0043","id":"secure-coding","label":"Writing secure code and dependencies","summary":"Generated code carries known vulnerability patterns, and models suggest packages that do not exist.","group":"coding","status":"active","aliases":["insecure code generation","package hallucination","slopsquatting"],"tags":["coding-agent","security"],"parent":null,"counts":{"claims":8,"contested":0,"techniques":3,"sources":7},"url":"https://rsiratchet.com/capabilities/secure-coding"}],"techniques":[{"ref":"TEC-0086","id":"ablate-cot-prediction-shared-sae-concepts-test-faithfulness","label":"Ablate CoT/prediction shared SAE concepts to test faithfulne","summary":"Encode the answer-position residual of a direct-prediction pass and a CoT pass with one shared sparse autoencoder, then zero the decoder contribution of the concepts active in both and measure the drop in answer probability.","kind":"process","status":"proposed","addresses":["explanation-faithfulness"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/ablate-cot-prediction-shared-sae-concepts-test-faithfulness"},{"ref":"TEC-0067","id":"abstain-disagreement-between-task-equivalent-prompt-regimes","label":"Abstain on disagreement between task-equivalent prompt regim","summary":"Ask the same closed-book question under two or three equivalent prompt structures (direct, decomposed, incremental); if the answers disagree, treat the fact as unknown and abstain, since memorized facts are stable across regimes and halluci","kind":"prompting","status":"proposed","addresses":["factual-recall"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/abstain-disagreement-between-task-equivalent-prompt-regimes"},{"ref":"TEC-0087","id":"abstain-unless-one-answer-dominates-lora-ensemble-s","label":"Abstain unless one answer dominates the LoRA ensemble's prob","summary":"Treat an ensemble of LoRA adapters as a credal set, bound each candidate answer's probability by the min and max over adapters (adding unwritten-completion mass to the upper bound), and answer only when one candidate's lower bound exceeds a","kind":"decoding","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/abstain-unless-one-answer-dominates-lora-ensemble-s"},{"ref":"TEC-0238","id":"abstain-answers-fail-track-entity-substitutions","label":"Abstain when answers fail to track entity substitutions","summary":"Replace entities in the question and evidence with type-matched synthetic names under a bijective mapping, re-ask, map the answer back, and abstain when back-mapped answers disagree with the original across sampled worlds.","kind":"prompting","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/abstain-answers-fail-track-entity-substitutions"},{"ref":"TEC-0088","id":"action-guidance-early-observation-enrichment-late","label":"Action guidance early, observation enrichment late","summary":"Modify the training environment's feedback rather than the agent: give explicit next-action hints in the first steps of an episode and the first half of training, then switch to richer state information in later steps and later training, ke","kind":"training","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/action-guidance-early-observation-enrichment-late"},{"ref":"TEC-0068","id":"actor-judge-editor-rl-solver-feedback-nl-to","label":"Actor-judge-editor RL on solver feedback for NL-to-PDDL","summary":"Have the model propose a PDDL formalization, judge it with a solver-calibrated score, and edit it under bounded diagnostic feedback, training all three roles with reinforcement learning from solver outcomes rather than annotations.","kind":"training","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/actor-judge-editor-rl-solver-feedback-nl-to"},{"ref":"TEC-0089","id":"add-3d-end-effector-point-cloud-latent-actions","label":"Add 3D end-effector point-cloud latent actions to image-base","summary":"Learns a second discrete latent action code from start-goal end-effector point-cloud transitions, shared across embodiments via a unified bimanual code, pair-consistent random 3D transforms, and forward/backward transition training, and use","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/add-3d-end-effector-point-cloud-latent-actions"},{"ref":"TEC-0090","id":"advance-progress-memory-only-verified-achievement","label":"Advance progress memory only on verified achievement","summary":"Represents the task as a subgoal sequence with a progress pointer, triggers verification at gripper-load events, and uses frozen point tracking (grasp) and language-conditioned cross-view embeddings plus a 2.43M-parameter head (placement) t","kind":"process","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/advance-progress-memory-only-verified-achievement"},{"ref":"TEC-0028","id":"adversarial-contrastive-fine-tuning-hallucination-flipped-image-pairs","label":"Adversarial contrastive fine-tuning on hallucination-flipped","summary":"Adversarial contrastive fine-tuning on hallucination-flipped image pairs, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/adversarial-contrastive-fine-tuning-hallucination-flipped-image-pairs"},{"ref":"TEC-0091","id":"adversarial-issue-descriptions-apr-agents","label":"Adversarial issue descriptions for APR agents","summary":"Embeds an attack-specific helper requirement (attacker-controlled env-var input plus a dangerous sink wired into the repaired symbol) into an otherwise legitimate bug report, framed as ordinary maintenance work.","kind":"prompting","status":"proposed","addresses":["secure-coding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/adversarial-issue-descriptions-apr-agents"},{"ref":"TEC-0239","id":"allocate-lora-rank-per-network-region-pre-training","label":"Allocate LoRA rank per network region by a pre-training diag","summary":"From ten unlabeled target observations, score each functional region of a VLA by combining CKA against a cached source reference with reference-free gradient and MC-dropout sensitivity, then set per-region LoRA rank proportional to the scor","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/allocate-lora-rank-per-network-region-pre-training"},{"ref":"TEC-0029","id":"anchor-claim-consistency-graphs-trusted-priors","label":"Anchor claim-consistency graphs to trusted priors","summary":"Decompose sampled responses and trusted source documents into atomic claims, build a weighted bipartite graph where prior-entailed edges cost least, and label each claim Grounded, Refuted, or Boundary by weighted closeness centrality.","kind":"retrieval","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/anchor-claim-consistency-graphs-trusted-priors"},{"ref":"TEC-0240","id":"arbitrate-conflicting-evidence-streams-relevance-gated-fusion","label":"Arbitrate conflicting evidence streams by relevance-gated fu","summary":"Generate perception, retrieval and web-verification evidence with the same backbone, score each stream for relevance/accuracy/discriminativeness, softmax-weight them into a fused answer, and refine under six cultural rubric criteria.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/arbitrate-conflicting-evidence-streams-relevance-gated-fusion"},{"ref":"TEC-0092","id":"ask-about-generic-ai-agent-model-itself","label":"Ask about a generic AI agent, not the model itself","summary":"Replace first-person self-prediction questions with the same item-informed question about a capable AI agent in general, or use other measured models as the predictor, to remove the self-flattering bias without losing accuracy.","kind":"prompting","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/ask-about-generic-ai-agent-model-itself"},{"ref":"TEC-0052","id":"ask-p-true-after-showing-model-several-own","label":"Ask for P(True) after showing the model several of its own s","summary":"Sample several answers, show them to the model, then ask for the probability that one specific answer is true; the extra samples act like considering alternatives and raise self-evaluation accuracy and calibration over judging a single answ","kind":"prompting","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/ask-p-true-after-showing-model-several-own"},{"ref":"TEC-0093","id":"ask-current-state-then-carry-forward","label":"Ask for the current state, then carry it forward","summary":"Instead of prompting the agent to forecast directly, have it search for and estimate the most recent period's distribution and use that estimate as the forecast.","kind":"prompting","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/ask-current-state-then-carry-forward"},{"ref":"TEC-0241","id":"attenuate-hidden-state-modes-hallucination-to-signal-ratio","label":"Attenuate hidden-state modes by hallucination-to-signal rati","summary":"Estimate truthful and hallucination-distortion covariances from paired captions offline, build a closed-form Wiener gain per eigenmode of the distortion covariance, and multiply it into the FFN output projection of deep layers so inference","kind":"architecture","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/attenuate-hidden-state-modes-hallucination-to-signal-ratio"},{"ref":"TEC-0094","id":"attribute-recurring-failures-before-editing-harness","label":"Attribute recurring failures before editing the harness","summary":"Aggregate failing trajectories across tasks into recurring failure groups, then run a fixed number of analyst-critic-engineer refinement passes plus a moderator to produce one modification specification, which a coding agent implements and","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/attribute-recurring-failures-before-editing-harness"},{"ref":"TEC-0095","id":"audit-agent-influence-belief-probes-exposure-logs","label":"Audit agent influence with belief probes and exposure logs","summary":"Instead of reading agent messages, score each agent's stance after every round with a fixed token-probability probe and log who saw which content from whom, so influence can be attributed to direct or peer-relayed exposure.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/audit-agent-influence-belief-probes-exposure-logs"},{"ref":"TEC-0242","id":"augment-constraint-check-each-candidate-before-committing-answer","label":"Augment and constraint-check each candidate before committin","summary":"After an incremental reason-act-observe loop, the agent refines coarse matches, retries failed or geo-inconsistent calls, and a synthesis step checks the draft against task constraints and admits no fact absent from logged observations.","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/augment-constraint-check-each-candidate-before-committing-answer"},{"ref":"TEC-0096","id":"benchmark-instructed-control-over-model-s-own-activations","label":"Benchmark instructed control over a model's own activations","summary":"Scores a model on six measures of how well it modulates its residual-stream projection onto a concept direction on instruction (engage, suppress, intensity dial, temporal placement, token coverage, layer targeting), combined into one geomet","kind":"process","status":"proposed","addresses":["strategic-deception"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/benchmark-instructed-control-over-model-s-own-activations"},{"ref":"TEC-0097","id":"blend-evicted-kv-into-routed-memory-slots-instead","label":"Blend evicted KV into routed memory slots instead of droppin","summary":"Organizes the KV cache as a fixed-size tree of memory buckets, routes each incoming token to a slot via Gumbel-Softmax, and sigmoid-gates a soft merge of the new key/value with the stored one instead of discarding any token, with a learned","kind":"architecture","status":"proposed","addresses":["long-context-degradation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/blend-evicted-kv-into-routed-memory-slots-instead"},{"ref":"TEC-0098","id":"bound-audit-batch-size-verify-every-reported-finding","label":"Bound audit batch size and verify every reported finding aga","summary":"Restrict document-auditing prompts to small batches with content injected directly, and mechanically check each reported finding against the source text before accepting it.","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/bound-audit-batch-size-verify-every-reported-finding"},{"ref":"TEC-0002","id":"capability-budget","label":"Bound what the agent may do, not just what it may not","summary":"An explicit allow/ask/deny policy plus rate, cost and time ceilings, enforced by the harness rather than requested of the model.","kind":"process","status":"active","addresses":["goal-conflict-safety","tool-use","prompt-injection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/capability-budget"},{"ref":"TEC-0243","id":"budget-reliability-per-step-instead-reporting-pass-rates","label":"Budget reliability per step instead of reporting pass rates","summary":"Fit a per-step reliability parameter from short-horizon runs and project it onto the target production horizon to predict success, instead of reporting aggregate benchmark pass rates.","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/budget-reliability-per-step-instead-reporting-pass-rates"},{"ref":"TEC-0244","id":"build-chain-of-thought-supervision-rules-api-calls","label":"Build chain-of-thought supervision from rules, not API calls","summary":"Generate reasoning traces for fine-tuning by reverse-engineering templated explanations that match extracted temporal features to candidate-model categories under known oracle labels, instead of querying a large model for the rationale.","kind":"training","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/build-chain-of-thought-supervision-rules-api-calls"},{"ref":"TEC-0099","id":"build-reasoning-graph-local-nli-model-then-classify","label":"Build the reasoning graph with a local NLI model, then class","summary":"Split a long chain of thought on transition keywords, link steps with entailment and contradiction edges predicted by a local NLI model instead of LLM calls, and train a graph attention network to predict whether the chain's final answer is","kind":"process","status":"proposed","addresses":["multi-step-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/build-reasoning-graph-local-nli-model-then-classify"},{"ref":"TEC-0100","id":"calibrate-confidence-per-atomic-claim-per-response","label":"Calibrate confidence per atomic claim, not per response","summary":"Split a response into atomic claims, score each by self-verification or token log-probability mixed with cross-sample consistency, then fit temperature or Platt scaling at the claim level so low-confidence claims can be routed to retrieval","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/calibrate-confidence-per-atomic-claim-per-response"},{"ref":"TEC-0101","id":"calibrate-value-vectors-synergy-attention-heads","label":"Calibrate value vectors in synergy attention heads","summary":"Classify attention heads by causal noise intervention and counterfactual difference-in-differences, then at decode time rescale the visual and language value vectors of the mixed 'synergy' heads so their information ratio hits a target equi","kind":"decoding","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/calibrate-value-vectors-synergy-attention-heads"},{"ref":"TEC-0245","id":"check-logit-lens-decode-attended-image-patches-against","label":"Check logit-lens decode of attended image patches against th","summary":"At each generated object token, average attention over mid-to-late layers, decode the top-k attended image tokens through the output head, and flag hallucination when no decoded token is semantically similar to the emitted word; then either","kind":"decoding","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/check-logit-lens-decode-attended-image-patches-against"},{"ref":"TEC-0001","id":"atomic-fact-checking","label":"Check long answers fact by fact","summary":"Split a long answer into atomic claims and verify each against search or a corpus.","kind":"tooling","status":"active","addresses":["hallucination","fact-verification"],"counts":{"claims":0,"sources":2},"url":"https://rsiratchet.com/techniques/atomic-fact-checking"},{"ref":"TEC-0010","id":"file-checkpoint","label":"Checkpoint task state to a file","summary":"Write task id, status, completed steps and artifact paths to a small file after every meaningful step, and read it at session start.","kind":"process","status":"active","addresses":["state-tracking","long-term-memory"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/file-checkpoint"},{"ref":"TEC-0053","id":"code-graph-tools-localization-agents","label":"Code-graph tools for localization agents","summary":"Parse the repository once into a heterogeneous graph of code entities and their dependency edges, expose search, traverse and retrieve-entity tools over it, and let the agent hop across the graph from the issue's mentioned names to the func","kind":"tooling","status":"proposed","addresses":["repository-navigation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/code-graph-tools-localization-agents"},{"ref":"TEC-0102","id":"collect-demonstrations-only-where-policy-disagrees-weight-progress","label":"Collect demonstrations only where the policy disagrees, and ","summary":"During handheld demonstration, query the current policy on the same observation stream without executing it, record extra demonstration segments when an energy score between the human action chunk and the policy's sampled chunks exceeds a t","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/collect-demonstrations-only-where-policy-disagrees-weight-progress"},{"ref":"TEC-0103","id":"compare-agent-pruning-against-random-pruning-equal-budget","label":"Compare agent pruning against random pruning at equal budget","summary":"For every structural change a multi-agent optimizer emits, build a random topology with the same node/edge budget and evaluate it under identical model, agents, roles and runtime, so the credit goes to the optimizer rather than to having de","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/compare-agent-pruning-against-random-pruning-equal-budget"},{"ref":"TEC-0104","id":"complete-prefixes-weaker-out-of-family-model-during","label":"Complete prefixes from a weaker out-of-family model during R","summary":"For a fraction of RLVR rollouts, prepend a truncated partial reasoning trajectory generated by a smaller, different-family model (cut at the largest step-to-step entropy drop) and train the target model to complete it.","kind":"training","status":"proposed","addresses":["fine-tuning-retention"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/complete-prefixes-weaker-out-of-family-model-during"},{"ref":"TEC-0105","id":"compose-role-steering-vectors-per-query-instead-debating","label":"Compose role steering vectors per query instead of debating","summary":"A learnable codebook of role steering vectors, warm-started from contrastive activation directions and fused by a query-aware top-K router, is added to the last prompt token's residual stream at the middle layer of a frozen LLM, trained by","kind":"training","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/compose-role-steering-vectors-per-query-instead-debating"},{"ref":"TEC-0246","id":"compress-audio-kv-cache-around-asr-transcript-anchors","label":"Compress audio KV cache around ASR transcript anchors","summary":"Slice the acoustic feature sequence by ASR timestamps, average-pool each transcribed span to its text length and add the aligned text embedding, keep untranscribed spans raw, then evict remaining tokens by attention scores accumulated with","kind":"architecture","status":"proposed","addresses":["long-context-degradation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/compress-audio-kv-cache-around-asr-transcript-anchors"},{"ref":"TEC-0106","id":"compute-lifecycle-consequences-harness-model","label":"Compute lifecycle consequences in the harness, not in the mo","summary":"Where component dependencies and cleanup effects are explicitly represented, have the runtime enumerate legal teardown continuations and compute or verify the resulting state, leaving the model only the higher-level goal choice.","kind":"tooling","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/compute-lifecycle-consequences-harness-model"},{"ref":"TEC-0107","id":"condition-action-denoising-raw-3d-semantic-point-field","label":"Condition action denoising on a raw 3D semantic point field ","summary":"Fuse frozen 3D-foundation-model features into VLM visual tokens, and let action hidden states cross-attend (with 3D RoPE) to an unencoded point-wise semantic field, inserted only in the diffusion blocks where linear probing shows degraded e","kind":"architecture","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/condition-action-denoising-raw-3d-semantic-point-field"},{"ref":"TEC-0108","id":"condition-reasoning-credit-spatio-temporal-evidence","label":"Condition reasoning credit on spatio-temporal evidence","summary":"Require the model to output timestamped audio and visual evidence intervals and multiply its answer correctness by a soft temporal-plus-semantic evidence score, so ungrounded correct answers earn no evidence credit.","kind":"process","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/condition-reasoning-credit-spatio-temporal-evidence"},{"ref":"TEC-0247","id":"condition-self-teacher-only-evidence-frames","label":"Condition the self-teacher only on evidence frames","summary":"In on-policy self-distillation for video LLMs, the frozen self-teacher replays the student's trajectory while seeing only frames from the annotated evidence interval plus the gold answer, and token-level distillation is weighted by each tok","kind":"training","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/condition-self-teacher-only-evidence-frames"},{"ref":"TEC-0054","id":"conformal-abstention-self-consistency","label":"Conformal abstention from self-consistency","summary":"Draw k responses, have the model count how many match each candidate, use that match score as confidence, and set the abstention threshold on a labelled calibration set with conformal prediction so the expected error among answered question","kind":"decoding","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/conformal-abstention-self-consistency"},{"ref":"TEC-0248","id":"conformal-shortlist-first-then-calibrated-margin-abstention","label":"Conformal shortlist first, then calibrated margin abstention","summary":"Use conformal prediction to localize a top-k shortlist that contains the human-preferred response with probability 1-alpha, then accept a single answer only when its score margin within that shortlist clears a fixed-sequence-calibrated thre","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/conformal-shortlist-first-then-calibrated-margin-abstention"},{"ref":"TEC-0109","id":"consolidate-only-experience-survives-state-revision","label":"Consolidate only experience that survives state revision","summary":"Alternates fast edits to an explicit textual state (validation-gated) with distillation into the policy, keeping only trajectories that the updated state makes newly or persistently better than state-free execution, weighted by per-token di","kind":"training","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/consolidate-only-experience-survives-state-revision"},{"ref":"TEC-0081","id":"contrastive-module-scoped-harness-evolution-benchmark-disjoint-tasks","label":"Contrastive, module-scoped harness evolution on benchmark-di","summary":"Collect success and failure trajectories for the same task, aggregate the contrast across tasks into recurring behavioural deficiencies, attribute each to one of five harness modules, evolve each module inside its own modification scope, th","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/contrastive-module-scoped-harness-evolution-benchmark-disjoint-tasks"},{"ref":"TEC-0249","id":"coordinate-agents-through-persistent-shared-world-messages","label":"Coordinate agents through a persistent shared world, not mes","summary":"Let role-free agents build durable, executable artifacts in a shared simulated environment so later agents inherit and modify them, instead of coordinating through conversation or assigned roles.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/coordinate-agents-through-persistent-shared-world-messages"},{"ref":"TEC-0110","id":"coordinate-hypothesis-agents-generational-genetic-algorithm","label":"Coordinate hypothesis agents with a generational genetic alg","summary":"Specialized LLM agents generate, pairwise-score and semantically recombine or mutate structured hypotheses, with a genetic algorithm doing tournament parent selection and joint parent-offspring truncation to a fixed population.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/coordinate-hypothesis-agents-generational-genetic-algorithm"},{"ref":"TEC-0111","id":"correct-reasoning-inverting-typed-error-operators","label":"Correct reasoning by inverting typed error operators","summary":"Corrupt correct reasoning traces by composing typed executable error operators, train an adapter to predict the active operator plus the adjacent lower-noise trace, and at inference iteratively predict and apply inverse operators to the mod","kind":"training","status":"proposed","addresses":["self-repair"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/correct-reasoning-inverting-typed-error-operators"},{"ref":"TEC-0080","id":"critic-gated-refinement-stages-simulated-rebuttal","label":"Critic-gated refinement stages with a simulated rebuttal","summary":"Run every stage of an autonomous research loop as candidate, critic, refine with a bounded number of rounds, where the critic can accept, request refinement or reject, and close the loop with a simulated peer-review and rebuttal before the","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/critic-gated-refinement-stages-simulated-rebuttal"},{"ref":"TEC-0069","id":"cultural-prompting-via-country-naming-system-prompt","label":"Cultural prompting via a country-naming system prompt","summary":"State the target country in the system prompt so the model answers survey-style questions as a member of that culture; measured as reduced distance to the country's Hofstede profile.","kind":"prompting","status":"proposed","addresses":["cultural-competence"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/cultural-prompting-via-country-naming-system-prompt"},{"ref":"TEC-0250","id":"debate-mixture-of-agents-rebut-revise-then-aggregate","label":"Debate-Mixture-of-Agents: rebut, revise, then aggregate","summary":"Each debater runs a presenter that drafts a diagnosis, a rebutter that attacks it for missing evidence and unsafe omissions, and a presenter revision; a separate aggregator then merges all revised hypotheses without forcing consensus.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/debate-mixture-of-agents-rebut-revise-then-aggregate"},{"ref":"TEC-0070","id":"debiasing-vector-separates-validity-plausibility","label":"Debiasing vector that separates validity from plausibility","summary":"Find the plausibility and validity directions in activation space, construct a vector that removes the plausibility component from the validity judgement, and add it at inference so conclusions are judged by form rather than believability.","kind":"architecture","status":"proposed","addresses":["logical-deduction-robustness"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/debiasing-vector-separates-validity-plausibility"},{"ref":"TEC-0112","id":"decompose-observation-into-sub-goal-sub-view-before","label":"Decompose the observation into a sub-goal and sub-view befor","summary":"At inference, the agent splits each complex web observation into an aligned sub-goal and task-relevant sub-observation, self-scores the decomposition and action, and uses the passed/rejected trials as in-context demonstrations or as RL sign","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/decompose-observation-into-sub-goal-sub-view-before"},{"ref":"TEC-0251","id":"delete-one-context-example-question-uniquely-needs","label":"Delete the one context example a question uniquely needs","summary":"Score each context example by how many questions require tokens only it supplies (Question Damage Score), then delete the highest-scoring example to make a matched, structurally unanswerable variant of the same task.","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/delete-one-context-example-question-uniquely-needs"},{"ref":"TEC-0113","id":"detect-benchmark-leakage-code-specific-membership-classifier","label":"Detect benchmark leakage with a code-specific membership cla","summary":"Fine-tune a shadow model on part of a benchmark to label members and non-members, then train a classifier on expert features (pass rate, CodeBLEU, edit distance, perplexity) fused with CodeBERT embeddings to predict whether a benchmark item","kind":"process","status":"proposed","addresses":["evaluation-validity"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/detect-benchmark-leakage-code-specific-membership-classifier"},{"ref":"TEC-0252","id":"detect-hallucination-adjacent-layer-attention-drift-plus-masked","label":"Detect hallucination from adjacent-layer attention drift plu","summary":"Compute KL divergence between consecutive layers' cross-modal attention over visual tokens, mask the top-attended region of the layer just before the largest drift, re-query the same token's probability, and feed the drift profile plus the","kind":"tooling","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/detect-hallucination-adjacent-layer-attention-drift-plus-masked"},{"ref":"TEC-0005","id":"computational-sensors-first","label":"Deterministic checks before LLM judges","summary":"Wire linters, test suites and schema validators into the agent's loop before adding any model-based judgment.","kind":"process","status":"active","addresses":["self-repair","evaluation-validity"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/computational-sensors-first"},{"ref":"TEC-0114","id":"deterministic-siem-pre-filter-before-semantic-guardrails","label":"Deterministic SIEM pre-filter before semantic guardrails","summary":"Sanitize and length-bound log payloads with regex decoders and atomic rules at SIEM ingestion, then pass the structured alert through LLM self-check input and output rails before and after the analyst model.","kind":"process","status":"proposed","addresses":["prompt-injection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/deterministic-siem-pre-filter-before-semantic-guardrails"},{"ref":"TEC-0030","id":"diagnosis-selected-head-restricted-adapter-plus-grounding-gated","label":"Diagnosis-selected head-restricted adapter plus grounding-ga","summary":"Diagnosis-selected head-restricted adapter plus grounding-gated decoding, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/diagnosis-selected-head-restricted-adapter-plus-grounding-gated"},{"ref":"TEC-0253","id":"distil-human-planning-practice-into-re-injected-planning","label":"Distil human planning practice into a re-injected planning p","summary":"Adds a ~1000-token prompt block at run start and every 30 minutes that forbids mono-loop moves (single hold-out, repeated tweaks) and prescribes human practices (K-fold from version one, early ensembling, cached out-of-fold predictions, per","kind":"prompting","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/distil-human-planning-practice-into-re-injected-planning"},{"ref":"TEC-0254","id":"distill-chunk-boundaries-trajectory-induced-programmatic-skills","label":"Distill chunk boundaries from trajectory-induced programmati","summary":"Induce two-level programmatic skills (composite skills made of ordered subskill calls) from successful trajectories, expand each subskill call into an action chunk whose boundary becomes an off-policy supervision target, and train the primi","kind":"training","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/distill-chunk-boundaries-trajectory-induced-programmatic-skills"},{"ref":"TEC-0236","id":"distilled-experience-memory","label":"Distill lessons from past attempts into memory","summary":"After each task, have the agent judge whether it succeeded, distill a short reusable strategy or warning from the attempt, store it, and retrieve only the relevant ones at the start of later tasks.","kind":"process","status":"active","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/distilled-experience-memory"},{"ref":"TEC-0115","id":"distill-stage-transition-monitoring-via-student-self-explanations","label":"Distill stage-transition monitoring via student self-explana","summary":"A large teacher VLM produces layered reasoning traces linking observations and task plan to annotated subtask transitions; a small student rewrites those traces into compact self-explanations, teacher-judged candidates are selected, and the","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/distill-stage-transition-monitoring-via-student-self-explanations"},{"ref":"TEC-0255","id":"distill-teacher-trajectories-into-workflow-subtask-function-memory","label":"Distill teacher trajectories into workflow, subtask and func","summary":"Builds three memory banks from a large teacher's successful trajectories — a prose workflow insight, code-bearing subtask segments, and per-function call examples — injecting the first two into the student's system prompt and retrieving the","kind":"retrieval","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/distill-teacher-trajectories-into-workflow-subtask-function-memory"},{"ref":"TEC-0116","id":"drop-blanket-row-limit-instructions-text-to-sql","label":"Drop blanket row-limit instructions from text-to-SQL prompts","summary":"Remove global output-shaping directives such as \"LIMIT 10\" from the SQL-generation system prompt so row count is decided by the question rather than the prompt.","kind":"prompting","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/drop-blanket-row-limit-instructions-text-to-sql"},{"ref":"TEC-0256","id":"drop-cae-scaffolding-supply-solver-tutorials-instead","label":"Drop CAE scaffolding, supply solver tutorials instead","summary":"Run CAE simulation tasks with an off-the-shelf coding-agent harness (file, shell, multi-turn, execution feedback) and no simulation-specific roles, retrieval pipeline or reflection prompt, while requiring the agent to read the relevant solv","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/drop-cae-scaffolding-supply-solver-tutorials-instead"},{"ref":"TEC-0117","id":"emit-tool-calls-parallel-function-token-stream","label":"Emit tool calls on a parallel function-token stream","summary":"Adds a dedicated autoregressive function channel, with its own head and start/end tool-call boundary tokens supervised alongside padding, running in parallel with the agent-text stream so tool calls are not serialized into the spoken respon","kind":"architecture","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/emit-tool-calls-parallel-function-token-stream"},{"ref":"TEC-0008","id":"digit-position-embeddings","label":"Encode digit position in the model","summary":"Add positional embeddings that tell the model which place each digit occupies.","kind":"architecture","status":"active","addresses":["arithmetic"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/digit-position-embeddings"},{"ref":"TEC-0257","id":"evolutionary-kernel-search-plus-c-level-functional-repair","label":"Evolutionary kernel search plus C-level functional repair, t","summary":"A multi-agent pipeline that injects Vitis HLS optimization priors, evolves kernel candidates under simulation feedback, verifies and repairs functional correctness by de-HLSifying the kernel and running C++ test programs, then generates hos","kind":"process","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/evolutionary-kernel-search-plus-c-level-functional-repair"},{"ref":"TEC-0119","id":"evolve-decision-knowledge-base-solver-failures-gated-feasibility","label":"Evolve a decision knowledge base from solver failures, gated","summary":"Accumulated solver and trajectory-validation failures are summarized into candidate add/modify/delete/merge edits to an initially empty knowledge base of scene-conditioned allowed and forbidden maneuvers, and an edit is kept only if re-runn","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/evolve-decision-knowledge-base-solver-failures-gated-feasibility"},{"ref":"TEC-0118","id":"evolve-component-boundaries-modularity-metrics-until-convergence","label":"Evolve component boundaries by modularity metrics until conv","summary":"Maintain a requirement-level and a component-level DAG with an alignment relation, and apply split/merge/revise/save actions triggered by cohesion and coupling thresholds until no action is eligible, then generate code against the converged","kind":"process","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/evolve-component-boundaries-modularity-metrics-until-convergence"},{"ref":"TEC-0258","id":"evolve-data-generation-pipeline-pilot-sets-before-scaling","label":"Evolve the data-generation pipeline on pilot sets before sca","summary":"A controller compiles a data-preparation runtime specification, generates a small pilot dataset, assesses it against domain-selected quality criteria, and emits an adapter that revises specific pipeline components while preserving their int","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/evolve-data-generation-pipeline-pilot-sets-before-scaling"},{"ref":"TEC-0259","id":"exchange-typed-logical-edges-dialogue-between-agents","label":"Exchange typed logical edges, not dialogue, between agents","summary":"Each modality agent builds a local evidence graph and agents update one another only by establishing typed edges (cascade, syndrome, exclusion, context) before an expert agent arbitrates the merged candidate graphs against a referral knowle","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/exchange-typed-logical-edges-dialogue-between-agents"},{"ref":"TEC-0260","id":"expose-domain-pre-trained-models-one-tool-api","label":"Expose domain pre-trained models as one tool API and pick pe","summary":"Wraps heterogeneous domain pre-trained models as stateless feature-extraction endpoints with declarative input specs, and has a planner rank candidates from the dataset's modality, channel count, sampling rate and class balance before code","kind":"tooling","status":"proposed","addresses":["domain-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/expose-domain-pre-trained-models-one-tool-api"},{"ref":"TEC-0120","id":"externalize-agent-state-onto-fact-intent-dag","label":"Externalize agent state onto a fact-intent DAG","summary":"Store verified environment observations as immutable fact nodes in a persistent graph and allow an action only as an intent edge whose preconditions are already-verified facts, replacing in-context state maintenance.","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/externalize-agent-state-onto-fact-intent-dag"},{"ref":"TEC-0121","id":"externalize-expert-decision-tree-into-clinician-editable-rule","label":"Externalize the expert decision tree into a clinician-editab","summary":"Restrict the LLM to extracting canonical symptoms and patient context from a clinician-curated vocabulary, and let a deterministic rule engine owned by clinicians make the emergency decision.","kind":"process","status":"proposed","addresses":["procedure-following"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/externalize-expert-decision-tree-into-clinician-editable-rule"},{"ref":"TEC-0045","id":"extract-cue-text-first-then-write-xpath-over","label":"Extract cue text first, then write XPath over condensed HTML","summary":"Stage one prompts an LLM on a sanitized HTML tree to return target values plus the anchor text preceding them; stage two condenses the page around those nodes and prompts a stronger LLM to emit an XPath, refined by executing it on seed page","kind":"prompting","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/extract-cue-text-first-then-write-xpath-over"},{"ref":"TEC-0031","id":"feed-integer-rounded-values-normalized-ones-llm-forecasters","label":"Feed integer-rounded values, not normalized ones, to LLM for","summary":"Round each numeric time-series value to the nearest integer before serializing it into the prompt, preserving the original scale instead of standardizing or instance-normalizing.","kind":"prompting","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/feed-integer-rounded-values-normalized-ones-llm-forecasters"},{"ref":"TEC-0261","id":"fill-sql-skeleton-coverage-gaps-don-t-just","label":"Fill SQL-skeleton coverage gaps, don't just add data","summary":"Abstracts SQL queries into schema-independent skeletons, treats seed skeletons as fixed K-center centers, greedily selects external pool skeletons farthest from existing coverage, and re-instantiates each on a target training database with","kind":"training","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/fill-sql-skeleton-coverage-gaps-don-t-just"},{"ref":"TEC-0122","id":"filter-preference-pairs-reference-margin-length-gap-lexical","label":"Filter preference pairs by reference margin, length gap, lex","summary":"Score each preference pair by a normalized composite of |reference-model per-token log-prob margin|, chosen-minus-rejected token length difference, and TF-IDF similarity to a general-capability corpus, then train DPO only on the lowest-risk","kind":"training","status":"proposed","addresses":["fine-tuning-retention"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/filter-preference-pairs-reference-margin-length-gap-lexical"},{"ref":"TEC-0262","id":"filter-self-training-data-hidden-state-parameter-probe","label":"Filter self-training data with a hidden-state parameter prob","summary":"Train a per-layer logistic probe on the hidden state just before each parameter value is emitted, then use its score to keep only reliable self-generated tool calls for LoRA fine-tuning and to rerank sampled candidates at inference.","kind":"training","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/filter-self-training-data-hidden-state-parameter-probe"},{"ref":"TEC-0123","id":"find-responsible-agent-hint-injection-then-learn-corrected","label":"Find the responsible agent by hint injection, then learn fro","summary":"For each failed run, inject a ground-truth-derived hint into one agent at a time in reverse execution order until the failure resolves, then contrast that agent's original and hint-corrected outputs to write its prompt update, clustering si","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/find-responsible-agent-hint-injection-then-learn-corrected"},{"ref":"TEC-0026","id":"synthetic-desycophancy-training","label":"Fine-tune on opinion-irrelevant examples","summary":"Fine-tune on synthetic prompts where a stated user opinion must not change the answer.","kind":"training","status":"active","addresses":["sycophancy"],"counts":{"claims":1,"sources":1},"url":"https://rsiratchet.com/techniques/synthetic-desycophancy-training"},{"ref":"TEC-0027","id":"tool-use-fine-tuning","label":"Fine-tune on tool-call traces","summary":"Train the model on examples of deciding when to call a tool and how to use the result.","kind":"training","status":"active","addresses":["tool-use"],"counts":{"claims":1,"sources":3},"url":"https://rsiratchet.com/techniques/tool-use-fine-tuning"},{"ref":"TEC-0124","id":"flag-deception-argument-schemes-plus-critical-questions","label":"Flag deception with argument schemes plus critical questions","summary":"For a chosen deception argument scheme, an LLM instantiates each premise against the document with verbatim excerpts and a green/yellow/red judgement, then answers the scheme's critical questions with paired argument and counterargument bef","kind":"prompting","status":"proposed","addresses":["strategic-deception"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/flag-deception-argument-schemes-plus-critical-questions"},{"ref":"TEC-0125","id":"freeze-normalization-statistics-pre-stream-calibration-set","label":"Freeze normalization statistics from a pre-stream calibratio","summary":"Before continual adaptation, collect a handful of task-independent teleoperated trajectories covering the embodiment's anticipated motion range, compute per-dimension 1st/99th percentile action statistics once, and use those frozen statisti","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/freeze-normalization-statistics-pre-stream-calibration-set"},{"ref":"TEC-0263","id":"fuse-clip-object-logits-lvlm-entropy-decoding","label":"Fuse CLIP object logits by LVLM entropy at decoding","summary":"At each decoding step, pick object tokens among the top-k candidates with WordNet, rescore them with CLIP similarity to the image, rescale the scores into the LVLM logit range, and blend the two with a weight equal to the LVLM's normalized","kind":"decoding","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/fuse-clip-object-logits-lvlm-entropy-decoding"},{"ref":"TEC-0032","id":"fuse-frame-clip-level-similarity-then-adapt-query","label":"Fuse frame- and clip-level similarity, then adapt the query ","summary":"Compute query-video cosine similarities at both frame and sliding-window clip granularity, project clip scores onto the frame timeline with a triangular kernel and mix them with a fixed weight, and separately optimize a single additive offs","kind":"decoding","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/fuse-frame-clip-level-similarity-then-adapt-query"},{"ref":"TEC-0265","id":"gate-between-replaying-retrieved-winning-episodes-fresh-reasoning","label":"Gate between replaying retrieved winning episodes and fresh ","summary":"The agent stores successful past gameplay episodes, retrieves the few most similar to the current time and observation, and a gating rule decides whether to execute the retrieved action directly or reason anew by fusing the retrieved episod","kind":"retrieval","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gate-between-replaying-retrieved-winning-episodes-fresh-reasoning"},{"ref":"TEC-0127","id":"gate-disclosure-through-structured-knowledge-tree-retrieve-generate","label":"Gate disclosure through a structured knowledge tree with ret","summary":"Store the authored facts, prerequisites and contradictions in a tree; one model picks the single currently-permitted unrevealed fact, a second writes dialogue constrained to it, and a third checks whether the fact was conveyed before the tr","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gate-disclosure-through-structured-knowledge-tree-retrieve-generate"},{"ref":"TEC-0128","id":"gate-extraction-through-versioned-rule-store-schema-consistency","label":"Gate extraction through a versioned rule store with schema a","summary":"Every extracted rule is written to a shared versioned store under a stable id, checked against 26 schema constraints, scored on five dimensions by an LLM judge whose minimum dimension decides auto-approve/review/reject, then filtered for du","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gate-extraction-through-versioned-rule-store-schema-consistency"},{"ref":"TEC-0266","id":"gate-harness-edits-both-source-cases-resampled-held","label":"Gate harness edits on both source cases and a resampled held","summary":"Each mined edit to the harness's knowledge or tool library is applied in a sandbox and promoted only if it strictly improves accuracy or cost on the trajectories it was mined from and does not regress on a stratified held-out case set resam","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gate-harness-edits-both-source-cases-resampled-held"},{"ref":"TEC-0014","id":"irreversible-action-gate","label":"Gate irreversible actions outside the model","summary":"Require an out-of-band approval for deletes, payments, sends, and other irreversible tool calls.","kind":"process","status":"active","addresses":["goal-conflict-safety"],"counts":{"claims":1,"sources":2},"url":"https://rsiratchet.com/techniques/irreversible-action-gate"},{"ref":"TEC-0129","id":"gate-skills-sampled-coalition-knockouts-then-mask-after","label":"Gate skills on sampled coalition knockouts, then mask after ","summary":"At acceptance time, score a candidate bank by the averaged value loss from knocking out randomly sampled skill subsets rather than by aggregate reward alone; at deployment, greedily drop transferred skills whose removal improves retrieval s","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gate-skills-sampled-coalition-knockouts-then-mask-after"},{"ref":"TEC-0264","id":"gate-answer-separate-self-verification-call-contradictions","label":"Gate the answer on a separate self-verification call for con","summary":"Run one extra call asking the same model only whether the input facts are sufficient and consistent, and override the solver's numeric answer with abstention when it returns the 'contradiction' label.","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gate-answer-separate-self-verification-call-contradictions"},{"ref":"TEC-0126","id":"gate-answer-self-scored-information-needs","label":"Gate the answer on self-scored information needs","summary":"Before answering, the model enumerates the information units the question requires, assigns each a confidence score, and commits to an answer only if the aggregate score passes a fixed threshold, otherwise abstaining.","kind":"prompting","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gate-answer-self-scored-information-needs"},{"ref":"TEC-0267","id":"gate-worker-outputs-semantic-entropy-plus-density","label":"Gate worker outputs on semantic entropy plus density","summary":"The orchestrator resamples each worker response, computes semantic entropy and embedding-space semantic density, and accepts only if both clear thresholds, otherwise reprompts, reassigns, or delegates in parallel.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gate-worker-outputs-semantic-entropy-plus-density"},{"ref":"TEC-0082","id":"gated-attributed-typed-edits-agent-queried-ontology","label":"Gated, attributed, typed edits to an agent-queried ontology","summary":"Keep the semantic layer outside the prompt as queryable MCP tools; when the agent fails, diagnose from traces, attribute the failure to the schema, content or tool level, write a typed edit at that level, and keep it only if a paired evalua","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gated-attributed-typed-edits-agent-queried-ontology"},{"ref":"TEC-0130","id":"generate-prompt-edits-only-failures-gate-preservation-set","label":"Generate prompt edits only from failures, gate on a preserva","summary":"Sample textual gradients only from examples the current prompt answers wrongly, generate several candidate prompts, and keep one only if its improvement on the hard batch beats a scheduled penalty times its regression on a held-back sample","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/generate-prompt-edits-only-failures-gate-preservation-set"},{"ref":"TEC-0046","id":"generate-regexes-grammar-then-crowd-paraphrase","label":"Generate regexes from a grammar, then crowd-paraphrase","summary":"Build a large NL-to-regex training corpus by stochastically generating regexes with a small hand-written grammar that also emits rigid synthetic descriptions, then paying crowd workers to paraphrase those descriptions into fluent language.","kind":"training","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/generate-regexes-grammar-then-crowd-paraphrase"},{"ref":"TEC-0268","id":"generate-action-skeleton-creator-transition-priors-then-assign","label":"Generate the action skeleton from creator transition priors,","summary":"Decode the step sequence with beam search over a creator's learned bigram/positional action priors and a deterministically predicted step count, then route ingredients to those steps by a learned ingredient-action prior instead of an LLM.","kind":"process","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/generate-action-skeleton-creator-transition-priors-then-assign"},{"ref":"TEC-0269","id":"generate-agent-topology-after-retrieval-conditioned-evidence","label":"Generate the agent topology after retrieval, conditioned on ","summary":"Retrieve evidence first, then autoregressively generate the multi-agent DAG (roles and edges) conditioned on the retrieved evidence and its provenance, with a generator trained on execution-scored, cost-pruned candidate topologies.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/generate-agent-topology-after-retrieval-conditioned-evidence"},{"ref":"TEC-0064","id":"geometric-fragility-score-safety-subspace","label":"Geometric fragility score from the safety subspace","summary":"Collect hidden states on harmful and harmless prompts, recover the low-rank safety subspace per layer, and compress its geometry into one score that ranks aligned checkpoints by how easily benign fine-tuning will erase refusal, so the fragi","kind":"tooling","status":"proposed","addresses":["fine-tuning-retention"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/geometric-fragility-score-safety-subspace"},{"ref":"TEC-0020","id":"retrieval-augmented-generation","label":"Ground answers in retrieved documents","summary":"Retrieve relevant documents and generate from them, with the model deciding when retrieval is needed.","kind":"retrieval","status":"active","addresses":["hallucination","fact-verification"],"counts":{"claims":1,"sources":2},"url":"https://rsiratchet.com/techniques/retrieval-augmented-generation"},{"ref":"TEC-0131","id":"ground-mission-feasibility-ontology-inferred-capabilities","label":"Ground mission feasibility in an ontology of inferred capabi","summary":"Store vehicle components, mission types and required capabilities in an ontology with SWRL rules that infer capabilities from hardware/software, then put both the mission's required capabilities and the vehicle's inferred capabilities into","kind":"retrieval","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/ground-mission-feasibility-ontology-inferred-capabilities"},{"ref":"TEC-0132","id":"group-candidates-execution-result-then-permute-within-groups","label":"Group candidates by execution result, then permute within gr","summary":"Replaces selector fine-tuning with retrieved structured criteria (schema grounding, SQL operations, expected output form) prepended to a listwise prompt, and mitigates position bias by shuffling only inside execution-result groups, aggregat","kind":"prompting","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/group-candidates-execution-result-then-permute-within-groups"},{"ref":"TEC-0033","id":"hidden-state-divergence-compensation-decoding-time","label":"Hidden-state divergence compensation at decoding time","summary":"Hidden-state divergence compensation at decoding time, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/hidden-state-divergence-compensation-decoding-time"},{"ref":"TEC-0034","id":"hidden-state-probe-tool-call-errors","label":"hidden-state probe for tool-call errors","summary":"hidden-state probe for tool-call errors, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/hidden-state-probe-tool-call-errors"},{"ref":"TEC-0055","id":"hierarchical-localization-tree-then-union-embedding-retrieval","label":"Hierarchical localization from tree, then union with embeddi","summary":"Show the model the repository's file tree and the issue and have it rank suspicious files, retrieve files by embedding similarity to the issue, merge the two lists, then narrow within the chosen files to classes and functions and finally to","kind":"process","status":"proposed","addresses":["repository-navigation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/hierarchical-localization-tree-then-union-embedding-retrieval"},{"ref":"TEC-0270","id":"imitate-belief-module-demonstrations-action-level","label":"Imitate belief-module demonstrations at the action level","summary":"Fine-tunes a VLM on per-timestep motor and expression actions recorded from a hybrid architecture with an explicit theory-of-mind module.","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/imitate-belief-module-demonstrations-action-level"},{"ref":"TEC-0271","id":"index-stream-shallow-layer-kvs-answer-deep-only","label":"Index the stream with shallow-layer KVs, answer deep only on","summary":"Streaming frames pass through only the shallow language layers whose KVs form a full-history index; at query time a text-only logit gate decides whether history is needed, cross-layer token voting plus max-min diversity picks evidence units","kind":"architecture","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/index-stream-shallow-layer-kvs-answer-deep-only"},{"ref":"TEC-0056","id":"induce-reusable-delimiter-wrapper-labelled-examples","label":"Induce a reusable delimiter wrapper from labelled examples","summary":"Learn a delimiter-based extraction pattern from a few labelled example pages, then run that pattern over the whole template-like collection instead of processing each page individually.","kind":"tooling","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/induce-reusable-delimiter-wrapper-labelled-examples"},{"ref":"TEC-0047","id":"induce-regexes-prompting-then-refine-against-failed-matches","label":"Induce regexes by prompting, then refine against failed matc","summary":"Prompt an LLM to write a regex from example spans, run it over the examples, and feed the mismatches back for iterative pattern revision.","kind":"prompting","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/induce-regexes-prompting-then-refine-against-failed-matches"},{"ref":"TEC-0048","id":"induce-xpath-seed-pages-via-multi-task-decomposition","label":"Induce XPath from seed pages via multi-task decomposition","summary":"Split XPath generation into separate sub-tasks the LLM solves over a few seed pages, producing a reusable XPath applied to all remaining pages without further model calls.","kind":"prompting","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/induce-xpath-seed-pages-via-multi-task-decomposition"},{"ref":"TEC-0272","id":"inject-proof-graph-correctness-obligations-then-consolidate-code","label":"Inject proof-graph correctness obligations, then consolidate","summary":"Parses reference Python into a region-level implementation graph, aligns those regions with kernel-checked Lean proof-flow transitions to generate region-level obligation text, trains the model jointly on code plus obligation/concept predic","kind":"training","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/inject-proof-graph-correctness-obligations-then-consolidate-code"},{"ref":"TEC-0134","id":"inject-self-refined-llm-test-snippets-search-stalls","label":"Inject self-refined LLM test snippets when search stalls","summary":"When a search-based test generator makes no fitness progress for N generations, prompt an LLM at class level with an iterative self-refinement loop, parse its test snippets into the generator's internal representation, and merge them into t","kind":"tooling","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/inject-self-refined-llm-test-snippets-search-stalls"},{"ref":"TEC-0133","id":"inject-golden-answer-off-policy-trajectory-during-rl","label":"Inject the golden answer as an off-policy trajectory during ","summary":"During GRPO-style knowledge-injection RL, add the gold answer to each rollout group and apply it as a direct supervised gradient scaled by its group-relative advantage (no importance ratio or clipping), so learning signal survives when all","kind":"training","status":"proposed","addresses":["knowledge-updating"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/inject-golden-answer-off-policy-trajectory-during-rl"},{"ref":"TEC-0135","id":"interrupt-agent-step-level-critic-flags-critical-risk","label":"Interrupt the agent when a step-level critic flags critical ","summary":"A lightweight generative reward model, trained on causal-chain-backtracked step labels from failed trajectories plus injected errors in successful ones, asynchronously diagnoses the agent's last completed step and forces regeneration of the","kind":"process","status":"proposed","addresses":["critique"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/interrupt-agent-step-level-critic-flags-critical-risk"},{"ref":"TEC-0071","id":"joint-neighbourhood-optimization-edits","label":"Joint neighbourhood optimization for edits","summary":"Before writing an edit into parameters, optimize the target representations for the edited fact and its neighbours together, penalizing both failed propagation and leakage, and reject edit plans whose semantic risk is high.","kind":"training","status":"proposed","addresses":["knowledge-updating"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/joint-neighbourhood-optimization-edits"},{"ref":"TEC-0015","id":"position-swap-judging","label":"Judge both orders and reconcile","summary":"Run the judge with candidates in both orders and count a preference only when it holds in both.","kind":"process","status":"active","addresses":["judge-position-bias"],"counts":{"claims":0,"sources":2},"url":"https://rsiratchet.com/techniques/position-swap-judging"},{"ref":"TEC-0136","id":"judge-each-rubric-item-against-full-agent-trajectory","label":"Judge each rubric item against the full agent trajectory","summary":"Convert a verified gold answer into a checklist whose items are each anchored to a hallucination subclass, then have an LLM judge issue a binary verdict per item against the agent's whole trajectory (reasoning, memory, tool observations) ra","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/judge-each-rubric-item-against-full-agent-trajectory"},{"ref":"TEC-0273","id":"judge-video-pairs-action-aligned-chunks-then-vote","label":"Judge video pairs in action-aligned chunks, then vote","summary":"Decompose a paired video comparison into chunks of four consecutive actions, judge each chunk's action consistency and visual quality from a six-image structured evidence panel, and aggregate chunk verdicts by voting into video-level prefer","kind":"prompting","status":"proposed","addresses":["critique"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/judge-video-pairs-action-aligned-chunks-then-vote"},{"ref":"TEC-0011","id":"guide-file","label":"Keep a guide file of past failures","summary":"A versioned instruction file the agent reads before acting, where every line is a past failure converted into a permanent rule.","kind":"process","status":"active","addresses":["procedure-following","context-hygiene"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/guide-file"},{"ref":"TEC-0137","id":"keep-inter-agent-handoffs-user-s-language","label":"Keep inter-agent handoffs in the user's language","summary":"Have agents pass extracted facts and answers to each other in the user's language instead of pivoting through English translation.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/keep-inter-agent-handoffs-user-s-language"},{"ref":"TEC-0274","id":"keep-true-false-forecasting-questions-near-1-1","label":"Keep true and false forecasting questions near 1:1","summary":"Balance a forecasting question bank so that events that occurred and events constructed as false appear in roughly equal numbers, so the calibration score does not reward default skepticism.","kind":"process","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/keep-true-false-forecasting-questions-near-1-1"},{"ref":"TEC-0138","id":"keep-world-state-engine-use-video-model-only","label":"Keep world state in an engine, use video model only as rende","summary":"An agent compiles natural-language instructions into executable programs whose state a lightweight engine maintains as state-augmented 3D oriented bounding boxes, which are deterministically projected into identity, semantic and motion cont","kind":"architecture","status":"proposed","addresses":["state-tracking"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/keep-world-state-engine-use-video-model-only"},{"ref":"TEC-0035","id":"knowledge-gated-deception-scoring","label":"Knowledge-gated deception scoring","summary":"Knowledge-gated deception scoring, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["strategic-deception"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/knowledge-gated-deception-scoring"},{"ref":"TEC-0275","id":"layered-memory-compression-four-daily-activity-blocks","label":"Layered memory compression with four daily activity blocks","summary":"Replaces agent-managed scratchpads with a system-managed three-layer memory (full logs for recent weeks, ~200-word summaries, ~50-word milestones) and splits each simulated day into four plannable blocks.","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/layered-memory-compression-four-daily-activity-blocks"},{"ref":"TEC-0139","id":"learn-agent-action-selection-instead-scripting-pipeline","label":"Learn agent action selection instead of scripting a pipeline","summary":"Each LLM agent reads a 20-dimensional summary of its private notes, shared knowledge, draft and feedback and picks Retrieve, Disseminate, Compose or Critique under a shared IPPO policy, instead of being assigned a fixed pipeline role.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/learn-agent-action-selection-instead-scripting-pipeline"},{"ref":"TEC-0276","id":"learn-kv-eviction-matching-full-cache-output-distribution","label":"Learn KV eviction by matching the full-cache output distribu","summary":"A ~200K-parameter MLP scores cached tokens from attention mass, value norm, attention variance, entropy and position, and is trained with REINFORCE against a per-step negative KL between full-cache and pruned-cache next-token distributions.","kind":"architecture","status":"proposed","addresses":["long-context-degradation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/learn-kv-eviction-matching-full-cache-output-distribution"},{"ref":"TEC-0277","id":"learn-per-module-merge-ratios-length-aware-reward","label":"Learn per-module merge ratios with a length-aware reward","summary":"Freeze a reasoning model and an instruction model of identical architecture, assign one sigmoid-bounded interpolation ratio per module, and update only those scalars with GRPO-style rollouts rewarded for correct and comparatively short answ","kind":"training","status":"proposed","addresses":["cutting-to-the-chase"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/learn-per-module-merge-ratios-length-aware-reward"},{"ref":"TEC-0140","id":"learn-skills-per-stage-supervision-stage-offers","label":"Learn skills per stage from the supervision that stage offer","summary":"Distil reusable text skills for hypothesis generation from a multi-dimensional discriminator's critique of the whole generated hypothesis set, and skills for experiment design and execution from testing trajectories scored against ground-tr","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/learn-skills-per-stage-supervision-stage-offers"},{"ref":"TEC-0141","id":"length-balance-rejected-responses-preference-data","label":"Length-balance rejected responses in preference data","summary":"Rewrite refusal and rejected responses to match the length and structure of the preferred grounded answer while preserving the underlying grounding failure, so preference learning cannot key on length.","kind":"training","status":"proposed","addresses":["critique"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/length-balance-rejected-responses-preference-data"},{"ref":"TEC-0024","id":"self-harness-loop","label":"Let the agent revise its own harness, regression-gated","summary":"Cluster failed traces into signatures, have the same fixed model propose a few minimal harness edits, and promote only edits that improve one split without degrading another.","kind":"process","status":"active","addresses":["self-repair","harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/self-harness-loop"},{"ref":"TEC-0142","id":"lora-latent-critic-verbalizes-which-parameter-ungrounded","label":"LoRA 'Latent Critic' that verbalizes which parameter is ungr","summary":"A zero-initialized LoRA adapter runs concurrently with the frozen base model during tool-call generation, trained with loss only on a trigger token and label, so it restructures the residual stream and then emits a localized verdict like 'u","kind":"training","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/lora-latent-critic-verbalizes-which-parameter-ungrounded"},{"ref":"TEC-0017","id":"recursive-summarization","label":"Maintain a running summary","summary":"Periodically fold the conversation into a summary that replaces older turns.","kind":"prompting","status":"active","addresses":["long-term-memory"],"counts":{"claims":1,"sources":1},"url":"https://rsiratchet.com/techniques/recursive-summarization"},{"ref":"TEC-0143","id":"make-arrival-model-s-own-declared-action-verified","label":"Make arrival the model's own declared action, verified after","summary":"Instead of ending an episode by an external distance threshold, the agent exposes a 'finished' action the model emits, keeps flying, and a verifier later checks whether any declaration was within the success radius with the target visible.","kind":"process","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/make-arrival-model-s-own-declared-action-verified"},{"ref":"TEC-0144","id":"make-source-provenance-architectural-attention-bias","label":"Make source provenance an architectural attention bias","summary":"Tag each input token with a discrete origin ring, add a learned origin embedding scaled by a learnable gate, and add a learned ring-by-ring bias to attention logits that suppresses attention from assistant/user queries to untrusted-source k","kind":"architecture","status":"proposed","addresses":["prompt-injection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/make-source-provenance-architectural-attention-bias"},{"ref":"TEC-0145","id":"mask-decoding-verbatim-context-spans-inside-quotes","label":"Mask decoding to verbatim context spans inside quotes","summary":"When a quote-opening token is generated, restrict the next-token logits (via a suffix tree over tokenised retrieved contexts) to tokens that extend the current span into a continuous verbatim substring of some retrieved passage, plus the cl","kind":"decoding","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/mask-decoding-verbatim-context-spans-inside-quotes"},{"ref":"TEC-0146","id":"mask-language-only-self-attention-keep-vision-to","label":"Mask language-only self-attention, keep vision-to-language f","summary":"At inference, apply mode-specific attention masks that suppress language-only self-attention in the late decoder layers so text tokens are forced to draw on visual tokens when attributing person attributes.","kind":"decoding","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/mask-language-only-self-attention-keep-vision-to"},{"ref":"TEC-0147","id":"match-agent-hierarchy-task-interdependence","label":"Match the agent hierarchy to task interdependence","summary":"Construct a per-mission manager tree in which horizontal managers hand out concurrent assignments to their children and vertical managers decompose the mission into ordered phases and advance only when completion conditions are met, with re","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/match-agent-hierarchy-task-interdependence"},{"ref":"TEC-0148","id":"match-encoder-s-pre-training-corpus-target-domain","label":"Match the encoder's pre-training corpus to the target domain","summary":"Instead of transferring a general-domain-trained hallucination detector, fine-tune a domain-pretrained encoder on a small set of in-domain claim-evidence pairs.","kind":"training","status":"proposed","addresses":["domain-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/match-encoder-s-pre-training-corpus-target-domain"},{"ref":"TEC-0278","id":"measure-proactive-monitoring-reflection-action-gap-traces","label":"Measure proactive monitoring and reflection-action gap from ","summary":"Compute, from an agent's tool-call log, the share of non-infrastructure calls spent querying global state (PMR) and the fraction of commitments stated in its planning diary that are executed within the next K turns (RAG@K).","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/measure-proactive-monitoring-reflection-action-gap-traces"},{"ref":"TEC-0279","id":"merge-capability-specialists-then-distil-student-prefixes","label":"Merge capability specialists, then distil on student prefixe","summary":"Train one RL specialist per execution capability from a shared backbone, merge their task vectors with TIES, then refine the merged student by distilling from the routed specialist teacher on the student's own generated prefixes.","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/merge-capability-specialists-then-distil-student-prefixes"},{"ref":"TEC-0280","id":"mid-train-tool-use-both-grounded-docs-executable","label":"Mid-train tool use on both grounded docs and executable traj","summary":"Builds a mid-training corpus that pairs filtered web/PDF/code sources augmented into tool-affordance QA with validated multi-turn trajectories synthesized from real API and MCP schemas, and trains the base model on the mixture before ordina","kind":"training","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/mid-train-tool-use-both-grounded-docs-executable"},{"ref":"TEC-0281","id":"mine-frozen-forecaster-s-residual-named-features","label":"Mine the frozen forecaster's residual for named features","summary":"Keeps the forecasting backbone frozen and mines its residual with a Monte-Carlo compositional search plus an LLM proposing named feature combinations, binary flags and code, admitting candidates through one source-blind validation gate into","kind":"tooling","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/mine-frozen-forecaster-s-residual-named-features"},{"ref":"TEC-0149","id":"mine-tiered-lessons-own-trajectories-principles-first","label":"Mine tiered lessons from own trajectories, principles first","summary":"After each labelled run, an agent distills successes into 'golden' lessons and failures into 'warning' lessons, consolidates them by embedding similarity into a three-tier library (principles, patterns, cases), and retrieves from it top-dow","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/mine-tiered-lessons-own-trajectories-principles-first"},{"ref":"TEC-0150","id":"mine-training-data-through-language-s-own-interpreter","label":"Mine training data through the language's own interpreter","summary":"Generate candidate programs from natural-language instructions with a frontier LLM, keep only those that pass a deterministic-wording filter, compile and run in the target language's binary, produce non-trivial output, and pass an independe","kind":"training","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/mine-training-data-through-language-s-own-interpreter"},{"ref":"TEC-0151","id":"mix-reversed-video-rollouts-into-grpo-groups","label":"Mix reversed-video rollouts into GRPO groups","summary":"During RL post-training of a video model, each group of sampled responses contains mostly forward-video rollouts plus one time-reversed rollout of the same clip, so correctness reward forces the policy to use temporal direction.","kind":"training","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/mix-reversed-video-rollouts-into-grpo-groups"},{"ref":"TEC-0072","id":"natural-language-draft-then-formalize-checker-in-the","label":"Natural-language draft, then formalize, with checker-in-the-","summary":"Have the model write an informal proof sketch first, translate it into Lean, and refine against compiler feedback; train the prover by reinforcement learning on long interactions with Lean so it accumulates experience of what checks.","kind":"training","status":"proposed","addresses":["formal-math"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/natural-language-draft-then-formalize-checker-in-the"},{"ref":"TEC-0036","id":"neutral-center-activation-steering-evidence-integration","label":"neutral-center activation steering for evidence integration","summary":"neutral-center activation steering for evidence integration, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["fact-verification"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/neutral-center-activation-steering-evidence-integration"},{"ref":"TEC-0073","id":"node-wise-binary-consistency-checks-chain-verification","label":"Node-wise binary consistency checks for chain verification","summary":"Split a reasoning chain into dependent nodes, check each node's consistency with its inputs by a yes/no query, and report the first failing node, instead of asking a verifier to read and judge the entire chain.","kind":"process","status":"proposed","addresses":["multi-step-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/node-wise-binary-consistency-checks-chain-verification"},{"ref":"TEC-0235","id":"offer-candidate-answers","label":"Offer the candidate answers","summary":"When a direct, answer-only response is wanted, give the model the possible answers to choose from instead of leaving the answer space open.","kind":"prompting","status":"active","addresses":["cutting-to-the-chase"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/offer-candidate-answers"},{"ref":"TEC-0016","id":"program-aided-computation","label":"Offload arithmetic to code","summary":"Have the model write a short program for the numeric part and run it, instead of computing in text.","kind":"tooling","status":"active","addresses":["arithmetic"],"counts":{"claims":1,"sources":1},"url":"https://rsiratchet.com/techniques/program-aided-computation"},{"ref":"TEC-0049","id":"parse-description-into-sketch-then-search-examples","label":"Parse the description into a sketch, then search with exampl","summary":"Convert the English description into a hierarchical sketch with constrained holes, then run a programming-by-example search that uses the sketch's hints both to prioritize expansion and to over/under-approximate partial programs for pruning","kind":"tooling","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/parse-description-into-sketch-then-search-examples"},{"ref":"TEC-0152","id":"pass-graph-structure-permutation-invariant-signature-text","label":"Pass graph structure as a permutation-invariant signature, n","summary":"Summarize sampled graph neighborhoods for an LLM agent as a fixed-length vector of named graph statistics, keeping raw text only for the top-k semantically relevant nodes, so the observation is order-independent and does not grow with neigh","kind":"prompting","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/pass-graph-structure-permutation-invariant-signature-text"},{"ref":"TEC-0037","id":"peer-prediction-bayesian-truth-serum-reward-label-free","label":"peer-prediction (Bayesian Truth Serum) reward for label-free","summary":"peer-prediction (Bayesian Truth Serum) reward for label-free RL fine-tuning, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["sycophancy"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/peer-prediction-bayesian-truth-serum-reward-label-free"},{"ref":"TEC-0153","id":"penalize-cross-batch-anchor-over-use-per-query","label":"Penalize cross-batch anchor over-use, not per-query redundan","summary":"Track each prompt-pool anchor's usage as an exponential moving average of its similarity scores across training batches and add a stop-gradient saturating-min penalty that suppresses similarity for over-used anchors, leaving the top-K retri","kind":"training","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/penalize-cross-batch-anchor-over-use-per-query"},{"ref":"TEC-0155","id":"penalize-tokens-whose-pooled-rank-evidence-falls-below","label":"Penalize tokens whose pooled rank evidence falls below their","summary":"At prefill, pool each candidate token's reciprocal vocabulary rank across all visual-token logit-lens readouts, cache it, and at each decoding step apply a bounded logit penalty to candidates whose pooled rank evidence falls below a token-s","kind":"decoding","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/penalize-tokens-whose-pooled-rank-evidence-falls-below"},{"ref":"TEC-0154","id":"penalize-tokens-low-vision-ablation-log-prob-gap","label":"Penalize tokens with a low vision-ablation log-prob gap","summary":"Run a second decoding pass with the visual features zeroed, score each candidate token by the log-probability difference between the two paths, and subtract a decaying penalty from candidates whose score falls below a threshold.","kind":"decoding","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/penalize-tokens-low-vision-ablation-log-prob-gap"},{"ref":"TEC-0156","id":"persist-pseudo-code-plan-restrict-calls","label":"Persist a pseudo-code plan and restrict calls to it","summary":"Before any action, the agent generates an ordered pseudo-code sequence of tool calls, writes it to a scratchpad that persists across the whole episode, and the executor admits only the plan-specified tool at each step unless replanning is t","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/persist-pseudo-code-plan-restrict-calls"},{"ref":"TEC-0074","id":"persona-disagreement-steering-cultural-alignment","label":"Persona-disagreement steering for cultural alignment","summary":"Prompt several demographically distinct personas grounded in a country's survey data, treat their agreement as already calibrated and their disagreement as the correction signal, and combine their answers to shift the base model's output to","kind":"prompting","status":"proposed","addresses":["cultural-competence"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/persona-disagreement-steering-cultural-alignment"},{"ref":"TEC-0157","id":"pick-audio-encoders-semantic-alignment-fidelity","label":"Pick audio encoders by semantic alignment, not fidelity","summary":"When choosing an audio front end for an audio-LLM, select encoders whose training imposed text/semantic constraints (weakly supervised or semantically distilled) instead of maximizing reconstruction fidelity or preferring continuous over di","kind":"architecture","status":"proposed","addresses":["audio-understanding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/pick-audio-encoders-semantic-alignment-fidelity"},{"ref":"TEC-0159","id":"pick-history-modality-what-question-needs","label":"Pick history modality by what the question needs","summary":"Feed long dialogue history as transcripts for lexical-recall questions and as raw audio for paralinguistic questions, rather than one format for all.","kind":"prompting","status":"proposed","addresses":["audio-understanding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/pick-history-modality-what-question-needs"},{"ref":"TEC-0158","id":"pick-contrast-region-counterfactual-occlusion-penalising-size","label":"Pick the contrast region by counterfactual occlusion, penali","summary":"Propose anatomical masks with a medical segmentation model, score each by the log-likelihood drop of a draft answer under two occlusion operators minus an area penalty, and use the top compact mask for classifier-free-guidance plus step-wis","kind":"decoding","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/pick-contrast-region-counterfactual-occlusion-penalising-size"},{"ref":"TEC-0160","id":"point-first-then-emit-quantized-trajectory-tokens","label":"Point first, then emit quantized trajectory tokens","summary":"Extends the VLM vocabulary with image-grid affordance/object pointing tokens and three-level residual vector-quantized action tokens, so spatial intent and precise trajectories are both decoded by the original language-model head.","kind":"architecture","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/point-first-then-emit-quantized-trajectory-tokens"},{"ref":"TEC-0161","id":"pool-semantic-opinions-across-models-then-split-entropy","label":"Pool semantic opinions across models, then split entropy int","summary":"Cluster sampled free-form responses from several VLMs into semantic-equivalence classes, average each model's class distribution uniformly, and report the pooled entropy as collective uncertainty plus Jensen-Shannon divergence as model conf","kind":"decoding","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/pool-semantic-opinions-across-models-then-split-entropy"},{"ref":"TEC-0075","id":"pre-conversation-warning-about-ai-persuasion","label":"Pre-conversation warning about AI persuasion","summary":"Before a user talks with a model that may argue a position, show a brief notice that models can be instructed to persuade and may select information; measured as halved belief change.","kind":"process","status":"proposed","addresses":["persuasion"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/pre-conversation-warning-about-ai-persuasion"},{"ref":"TEC-0282","id":"pre-train-compact-encoder-in-domain-reports-consistency","label":"Pre-train a compact encoder on in-domain reports for consist","summary":"Train a small BERT from scratch on a hospital's own radiology report corpus with masked language modelling plus a findings-conclusion consistency objective, then fine-tune it to classify synthetic reporting errors.","kind":"training","status":"proposed","addresses":["error-detection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/pre-train-compact-encoder-in-domain-reports-consistency"},{"ref":"TEC-0162","id":"predict-human-intention-symbolic-goal-then-plan-jointly","label":"Predict human intention as a symbolic goal, then plan jointl","summary":"A VLM infers the human's latent intention from third-person images plus a scene graph as symbolic goal predicates rather than an action sequence, and those goals are grounded with the robot's goal into a multi-agent PDDL problem solved by a","kind":"process","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/predict-human-intention-symbolic-goal-then-plan-jointly"},{"ref":"TEC-0163","id":"prefer-beam-search-over-sample-and-vote-under","label":"Prefer beam search over sample-and-vote under grammar constr","summary":"When output is masked to a formal grammar, spend inference budget on wider beams rather than on sampling several candidates and voting on their execution results.","kind":"decoding","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/prefer-beam-search-over-sample-and-vote-under"},{"ref":"TEC-0057","id":"priority-scheduled-search-distance-aware-context-pruning","label":"Priority-scheduled search with distance-aware context prunin","summary":"Keep a scored queue of code entities to inspect, let the model decompose each search into smaller lookups it scores for relevance, and drop context whose graph distance from already-relevant anchors is large, so the agent spends its budget","kind":"process","status":"proposed","addresses":["repository-navigation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/priority-scheduled-search-distance-aware-context-pruning"},{"ref":"TEC-0283","id":"probe-consistency-four-axes-one-score","label":"Probe consistency on four axes, not one score","summary":"Convert a binary question into an affirmative and a contradictory statement, re-ask each under semantically equivalent image transformations and paraphrases to form four consistency groups, and train a lightweight classifier on 14 within-gr","kind":"prompting","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/probe-consistency-four-axes-one-score"},{"ref":"TEC-0284","id":"probe-failure-structure-before-scoring-abstention","label":"Probe failure structure before scoring abstention","summary":"Re-run an answer under visual degradation, blank-image replacement, and grounding/relation swaps to produce a three-coordinate signature used for failure-family diagnosis and to pick a matched correction prompt, rather than as an abstention","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/probe-failure-structure-before-scoring-abstention"},{"ref":"TEC-0083","id":"probe-derived-value-signal-free-process-reward","label":"Probe-derived value signal as a free process reward","summary":"Fit a linear TD probe on hidden states to get a per-step value estimate, take its step-to-step change as a reward prediction error, and use that as the scoring signal for best-of-n or search at inference, so no separate reward model is run.","kind":"tooling","status":"proposed","addresses":["error-detection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/probe-derived-value-signal-free-process-reward"},{"ref":"TEC-0285","id":"profile-review-critique-function-human-likeness","label":"Profile review critique by function, not human-likeness","summary":"Annotate each review point against knowledge type, argument structure, question depth, cognitive complexity and feedback function so model and human critique are compared by what functions they perform rather than by textual overlap.","kind":"process","status":"proposed","addresses":["critique"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/profile-review-critique-function-human-likeness"},{"ref":"TEC-0084","id":"promote-fix-only-against-targeted-protected-negative-control","label":"Promote a fix only against targeted, protected and negative-","summary":"For each recurring failure, write a patch with an explicit scope and injection point, then admit it only if it fixes its targeted cases, leaves a protected regression suite unharmed, and does not fire on negative controls; refuse or defer c","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/promote-fix-only-against-targeted-protected-negative-control"},{"ref":"TEC-0164","id":"propagate-structured-state-truncate-closest-to-goal","label":"Propagate a structured state and truncate at closest-to-goal","summary":"Prompts a multimodal LLM to estimate the workstation state as a JSON object, propagate it forward through each predicted action, compute a numeric distance to the inferred goal state, and cut the action plan at the state with the lowest dis","kind":"prompting","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/propagate-structured-state-truncate-closest-to-goal"},{"ref":"TEC-0286","id":"propagate-uncertainty-over-trajectory-dependency-graph","label":"Propagate uncertainty over a trajectory dependency graph","summary":"Represent an agent's execution history as a directed graph of reasoning, tool-call and environment nodes with typed dependency edges (sequential, repetition, progression, feedback, goal alignment), then propagate each node's local uncertain","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/propagate-uncertainty-over-trajectory-dependency-graph"},{"ref":"TEC-0018","id":"reorder-context-by-relevance","label":"Put important material first or last","summary":"Order retrieved documents so the most relevant sit at the start and end of the prompt.","kind":"prompting","status":"active","addresses":["long-context-degradation"],"counts":{"claims":1,"sources":1},"url":"https://rsiratchet.com/techniques/reorder-context-by-relevance"},{"ref":"TEC-0287","id":"query-time-series-through-code-context","label":"Query the time series through code, not in context","summary":"Store the full recording as a variable in a sandboxed Python REPL and let the model iteratively emit code (and optionally render plots or spawn sub-agents) to inspect segments before committing a structured answer.","kind":"tooling","status":"proposed","addresses":["long-context-degradation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/query-time-series-through-code-context"},{"ref":"TEC-0288","id":"rank-communication-edges-masking-effect-then-prune-budget","label":"Rank communication edges by masking effect, then prune to bu","summary":"Mask each communication edge in a generated multi-agent graph, score it by the drop in task score plus the rise in semantic entropy of the final response, keep the top-ranked edges under a pruning budget, and distill those subgraphs into a","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/rank-communication-edges-masking-effect-then-prune-budget"},{"ref":"TEC-0289","id":"rank-crash-clustered-suspicious-files-llm-repair-context","label":"Rank crash-clustered suspicious files as LLM repair context","summary":"Cluster crash reports by stack-trace similarity, score candidate files by proximity to the crash point, bucket specificity and in-group frequency, and put the top-ranked files' source plus a representative stack trace into a zero-shot repai","kind":"retrieval","status":"proposed","addresses":["debugging"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/rank-crash-clustered-suspicious-files-llm-repair-context"},{"ref":"TEC-0165","id":"rank-feed-forward-neurons-class-conditional-residual-contribution","label":"Rank feed-forward neurons by class-conditional residual cont","summary":"Score each MLP neuron at the final prompt token by the separation between hallucinated and truthful mean pre-to-post activation change scaled by its output-vector norm, then train a small classifier on the top-ranked neurons' features.","kind":"architecture","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/rank-feed-forward-neurons-class-conditional-residual-contribution"},{"ref":"TEC-0166","id":"rank-pooled-multi-run-outputs-cross-run-vote","label":"Rank pooled multi-run outputs by cross-run vote count","summary":"Treat each LLM run as a black-box retrieval system casting a binary vote per candidate requirement, sum the votes (optionally weighted by each model family's estimated true-positive/false-positive log-odds), and hand reviewers the resulting","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/rank-pooled-multi-run-outputs-cross-run-vote"},{"ref":"TEC-0234","id":"re-check-safety-constraint-tool-prompt","label":"Re-check the safety constraint in a tool, not the prompt","summary":"Ground objects as bounding boxes, have the agent propose a waypoint route, reject and replan any route that intersects the obstacle box (before and during motion), and pick the contact side by an adjacency test with the obstacle.","kind":"tooling","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/re-check-safety-constraint-tool-prompt"},{"ref":"TEC-0038","id":"re-query-after-removing-concepts-explanation-ignored","label":"Re-query after removing concepts the explanation ignored","summary":"Identify input concepts not credited in the model's explanation, delete them from the input, and re-ask the model so the final answer can only depend on the concepts it named.","kind":"prompting","status":"proposed","addresses":["explanation-faithfulness"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/re-query-after-removing-concepts-explanation-ignored"},{"ref":"TEC-0019","id":"reread-before-edit","label":"Re-read before editing","summary":"Force the agent to read the current version of a file immediately before changing it.","kind":"process","status":"active","addresses":["state-tracking"],"counts":{"claims":0,"sources":0},"url":"https://rsiratchet.com/techniques/reread-before-edit"},{"ref":"TEC-0167","id":"react-loop-over-controllable-video-inspection-tools","label":"ReAct loop over controllable video inspection tools","summary":"A single agent iteratively calls video observation tools where it sets the time interval, frame rate and resolution itself, judging evidence sufficiency each turn before answering.","kind":"tooling","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/react-loop-over-controllable-video-inspection-tools"},{"ref":"TEC-0290","id":"read-impossibility-answerable-null-residual-subspace","label":"Read impossibility from the answerable-null residual subspac","summary":"Fit PCA on hidden states of answerable prompts, project all states out of that subspace, and score the mean-difference direction between answerable and impossible prompts by cosine to detect pre-generation impossibility.","kind":"architecture","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/read-impossibility-answerable-null-residual-subspace"},{"ref":"TEC-0291","id":"read-whole-computational-trace-graph-one-hidden-state","label":"Read the whole computational trace as a graph, not one hidde","summary":"Builds a directed attention graph over image patches, query tokens and generated tokens from a single forward pass of a frozen LVLM and classifies each response token with alternating GNN, ViT and GRU blocks, with a streaming variant that r","kind":"architecture","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/read-whole-computational-trace-graph-one-hidden-state"},{"ref":"TEC-0292","id":"recursively-search-video-context-construction-program","label":"Recursively search the video context-construction program","summary":"An LLM proposer repeatedly rewrites the executable program that selects, retrieves and packs video evidence into a frozen VLM's bounded context, keeping only candidates that strictly improve development accuracy.","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/recursively-search-video-context-construction-program"},{"ref":"TEC-0168","id":"reflect-only-agent-made-first-wrong-step","label":"Reflect only the agent that made the first wrong step","summary":"A process-reward model scores each step of a failed multi-agent trajectory, the first low-scoring step and its agent are identified, a corrected version of that step is generated and verified, and only that agent writes a reflection into it","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reflect-only-agent-made-first-wrong-step"},{"ref":"TEC-0058","id":"refusal-aware-instruction-tuning","label":"Refusal-aware instruction tuning","summary":"Run the base model on the tuning set, label each example as known or unknown by whether it answered correctly, keep answers for the known set and append an uncertainty expression to the unknown set, then fine-tune on the union so the model","kind":"training","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/refusal-aware-instruction-tuning"},{"ref":"TEC-0169","id":"reject-tool-calls-do-resolve-registered-tool-signature","label":"Reject tool calls that do not resolve to a registered tool a","summary":"Before any permission gate, check each emitted call's name against the closed-world tool registry and type-check its arguments against the declared signature, rejecting anything unresolved.","kind":"tooling","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reject-tool-calls-do-resolve-registered-tool-signature"},{"ref":"TEC-0293","id":"render-knowledge-updates-shared-pool-extended-matching-items","label":"Render knowledge updates as shared-pool extended matching it","summary":"Group related update vignettes into one training instance with a shared answer option pool so the model must discriminate across vignettes between current and superseded recommendations.","kind":"training","status":"proposed","addresses":["knowledge-updating"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/render-knowledge-updates-shared-pool-extended-matching-items"},{"ref":"TEC-0170","id":"repair-one-hop-gaps-add-same-chunk-context","label":"Repair one-hop gaps and add same-chunk context before RL","summary":"Augments each KG triple with same-source-chunk supporting triples for supervision, then uses LLM judges to separate noisy triples from genuine one-hop knowledge gaps, continually fine-tunes on targeted repair items weighted by failure histo","kind":"training","status":"proposed","addresses":["multi-step-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/repair-one-hop-gaps-add-same-chunk-context"},{"ref":"TEC-0171","id":"repair-plans-world-model-ask-llm-only-semantic","label":"Repair plans from the world model, ask the LLM only for sema","summary":"Roll an LLM-generated action plan forward on an explicit graph world model of object relations, preconditions, effects and safety constraints, apply edits whose corrections follow from the action semantics, and return only unresolved semant","kind":"process","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/repair-plans-world-model-ask-llm-only-semantic"},{"ref":"TEC-0009","id":"external-feedback-repair","label":"Repair with external feedback","summary":"Give the model a concrete external signal, such as test output, and let it reflect and retry.","kind":"process","status":"active","addresses":["self-repair"],"counts":{"claims":1,"sources":3},"url":"https://rsiratchet.com/techniques/external-feedback-repair"},{"ref":"TEC-0294","id":"replace-conversation-history-rewritten-execution-state","label":"Replace conversation history with a rewritten execution stat","summary":"At each step the model sees only the immutable skill spec, the current structured state and the latest observation, emits a JSON state patch plus an action, and has its reasoning trace discarded after the runtime validates and merges the pa","kind":"architecture","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/replace-conversation-history-rewritten-execution-state"},{"ref":"TEC-0172","id":"replay-extracted-skills-tricks-model-cards-orchestrator","label":"Replay extracted skills, tricks and model cards to the orche","summary":"Run the swarm, summarize trajectories into SKILL.md workflow patterns, a trick library, and per-backbone model cards, and reinject all three into the orchestrator's prompt at inference.","kind":"prompting","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/replay-extracted-skills-tricks-model-cards-orchestrator"},{"ref":"TEC-0295","id":"replay-matched-trajectories-compare-confidence-to-action-maps","label":"Replay matched trajectories to compare confidence-to-action ","summary":"Collect each question's answer and confidence at every retrieval depth once, then replay the identical stored states through controllers that differ only in the confidence score supplied, so action differences are attributable to the score","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/replay-matched-trajectories-compare-confidence-to-action-maps"},{"ref":"TEC-0296","id":"replay-recorded-prefix-then-intervene-symptomatic-node","label":"Replay the recorded prefix, then intervene at the symptomati","summary":"Log every model and tool boundary interaction as an event-dependency graph, replay it deterministically up to a node flagged by symptom rules plus a semantic judge, inject a repair instruction conditioned on the symptom category, and resume","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/replay-recorded-prefix-then-intervene-symptomatic-node"},{"ref":"TEC-0173","id":"report-attempted-recognized-rates-just-completion","label":"Report attempted and recognized rates, not just completion","summary":"Instrument the agent pipeline to record the furthest stage an injected payload reaches and use a judge panel to label whether the agent attempted the attack and whether it explicitly recognized the injection, alongside the deterministic com","kind":"process","status":"proposed","addresses":["prompt-injection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/report-attempted-recognized-rates-just-completion"},{"ref":"TEC-0297","id":"report-calibration-trajectory-checkpoints-split-horizon","label":"Report calibration at trajectory checkpoints, split by horiz","summary":"Score a reliability estimate against the independently observed task outcome at each checkpoint index along a trajectory, reported separately per horizon with group sizes, instead of pooling all checkpoints or scoring only the final answer.","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/report-calibration-trajectory-checkpoints-split-horizon"},{"ref":"TEC-0174","id":"report-operational-acceptance-reference-reconstruction-separately","label":"Report operational acceptance and reference reconstruction s","summary":"Use parser, domain-conformance, planner, VAL and LLM-critic checks as the online stopping criterion, but compute renaming-invariant structural match and semantic equivalence against curated reference problems offline as a separate benchmark","kind":"process","status":"proposed","addresses":["evaluation-validity"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/report-operational-acceptance-reference-reconstruction-separately"},{"ref":"TEC-0175","id":"report-per-family-spread-aggregate-agent-score","label":"Report per-family spread, not an aggregate agent score","summary":"Split an agent benchmark into skill-isolated families with verifier-computed gold (canonical SQL, exact DAG schedule, closed-form lambda) and report per-family accuracy plus token cost from one instrumented run instead of a blended score.","kind":"process","status":"proposed","addresses":["evaluation-validity"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/report-per-family-spread-aggregate-agent-score"},{"ref":"TEC-0298","id":"report-subtask-progress-trajectory-diagnostics-just-success-rate","label":"Report subtask progress and trajectory diagnostics, not just","summary":"Score embodied navigation episodes on fraction of phases completed, final goal distance, collisions and step count alongside binary success, so partial progress and post-phase-transition stalls are visible.","kind":"process","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/report-subtask-progress-trajectory-diagnostics-just-success-rate"},{"ref":"TEC-0059","id":"repository-code-graph-navigation-plug-in","label":"Repository code graph as a navigation plug-in","summary":"Build a graph whose nodes are definition and reference lines and whose edges are contain and invoke relations across the whole repository, then answer the agent's or pipeline's lookups with the sub-graph around a queried symbol so it sees c","kind":"tooling","status":"proposed","addresses":["repository-navigation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/repository-code-graph-navigation-plug-in"},{"ref":"TEC-0299","id":"require-non-target-files-be-byte-identical-just","label":"Require non-target files be byte-identical, not just task go","summary":"Score a task as successful only when all task-specific final-state evaluators pass and every pre-existing non-target file is unchanged between initial and final sandbox state.","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/require-non-target-files-be-byte-identical-just"},{"ref":"TEC-0039","id":"requirements-refinement-before-coding","label":"requirements-refinement-before-coding","summary":"requirements-refinement-before-coding, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/requirements-refinement-before-coding"},{"ref":"TEC-0300","id":"rerank-candidate-routes-executable-chemical-verifier","label":"Rerank candidate routes with an executable chemical verifier","summary":"Take any backbone's top-10 candidate disconnections and reorder them by rule-based checks for functional-group competition, mechanistic plausibility and literature precedent, promoting unflagged routes above high-confidence flagged ones.","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/rerank-candidate-routes-executable-chemical-verifier"},{"ref":"TEC-0176","id":"restore-trusted-toolset-just-before-planning","label":"Restore the trusted toolset just before planning","summary":"Save the user's original tool list and overwrite whatever toolset the agent can see with it immediately before the planning step, so plans can only reference user-approved tools.","kind":"tooling","status":"proposed","addresses":["prompt-injection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/restore-trusted-toolset-just-before-planning"},{"ref":"TEC-0177","id":"retrieve-clause-level-knowledge-database-s-sql-query","label":"Retrieve clause-level knowledge from a database's SQL query ","summary":"Parse past SQL into five categories of clause-level fragments, describe each in natural language, retrieve by query- and component-level embedding similarity, then LLM-rerank before prepending to an existing few-shot text-to-SQL prompt.","kind":"retrieval","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/retrieve-clause-level-knowledge-database-s-sql-query"},{"ref":"TEC-0301","id":"retrieve-debate-memory-consensus-level-then-reweight-peers","label":"Retrieve debate memory by consensus level, then reweight pee","summary":"Each debating agent retrieves past-debate cases with an MMR trade-off whose relevance/diversity coefficient is set by the current consensus ratio (high consensus pulls in contrastive cases), and labels each peer's response high- or low-conf","kind":"prompting","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/retrieve-debate-memory-consensus-level-then-reweight-peers"},{"ref":"TEC-0302","id":"retrieve-demonstrations-per-decomposed-subtask-per-task","label":"Retrieve demonstrations per decomposed subtask, not per task","summary":"Cluster the prior dataset into skills, have an LLM decompose the target task into subtasks scored by semantic plausibility and per-skill held-out BC loss, retrieve episodes by subtask language similarity, rerank frames by VAE behavior-featu","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/retrieve-demonstrations-per-decomposed-subtask-per-task"},{"ref":"TEC-0178","id":"retrieve-in-context-critiques-weak-model-failure-mode","label":"Retrieve in-context critiques by weak-model failure mode","summary":"Build an offline bank of wrong answers from small same-family models labelled with clustered failure modes plus natural-language critiques, then retrieve critique exemplars for a large model either by its dominant family-level failure profi","kind":"prompting","status":"proposed","addresses":["critique"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/retrieve-in-context-critiques-weak-model-failure-mode"},{"ref":"TEC-0021","id":"retrieve-tool-docs","label":"Retrieve tool documentation at call time","summary":"Put the exact documentation for candidate tools into the prompt instead of relying on memory.","kind":"retrieval","status":"active","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/retrieve-tool-docs"},{"ref":"TEC-0060","id":"retrieve-and-rerank-localization-trained-mined-issue-patch","label":"Retrieve-and-rerank localization trained on mined issue-patc","summary":"Mine issues paired with the functions their fixing commits changed, train a code embedding model and a listwise reranker on those pairs so they read failure-style queries, then localize new issues by embedding search over the repository's f","kind":"retrieval","status":"proposed","addresses":["repository-navigation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/retrieve-and-rerank-localization-trained-mined-issue-patch"},{"ref":"TEC-0179","id":"reversible-memory-stack-per-action-process-rewards","label":"Reversible memory stack with per-action process rewards","summary":"Gives an RL-trained retrieval agent a LIFO memory stack with push/pop actions (plan, think, search, summary, backtrack, conclusion) and assigns format, retrieval-relevance and memory-usefulness rewards to the tokens of each action, plus bat","kind":"training","status":"proposed","addresses":["multi-step-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reversible-memory-stack-per-action-process-rewards"},{"ref":"TEC-0180","id":"reward-acoustic-fidelity-frozen-judge-during-rl","label":"Reward acoustic fidelity with a frozen judge during RL","summary":"Augments the text-space -WER GRPO reward with a group-normalized score from a permanently frozen, non-autoregressive CTC model that measures how well each sampled transcription aligns to the audio, used at training time only.","kind":"training","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-acoustic-fidelity-frozen-judge-during-rl"},{"ref":"TEC-0181","id":"reward-calibration-stated-confidence-inside-grpo","label":"Reward calibration of stated confidence inside GRPO","summary":"Augments the output schema with span and answer confidence values and trains with GRPO whose reward adds a Brier penalty against realised rollout correctness, a group-level ECE bonus, an abstention term supervised by group accuracy, and a s","kind":"training","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-calibration-stated-confidence-inside-grpo"},{"ref":"TEC-0303","id":"reward-each-model-peer-model-s-majority-vote","label":"Reward each model by a peer model's majority vote, not its o","summary":"Train several parameter-disjoint models simultaneously with RL, where each model's rollouts are scored against the majority-vote answer of a different model in the cohort, passed along a directed ring.","kind":"training","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-each-model-peer-model-s-majority-vote"},{"ref":"TEC-0182","id":"reward-each-tool-call-whether-was-necessary","label":"Reward each tool call by whether it was necessary","summary":"During RL, an LLM judge labels each tool invocation as necessary or not given the question and tool input, and unnecessary calls get a fixed per-step penalty added to the final-answer reward.","kind":"training","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-each-tool-call-whether-was-necessary"},{"ref":"TEC-0184","id":"reward-listwise-rank-correlation-exact-numeric-forecasts","label":"Reward listwise rank correlation, not exact numeric forecast","summary":"Train the model with GRPO using a Spearman rank-correlation reward between its generated asset ordering and realized future returns, with a fixed penalty for malformed or hallucinated asset lists, then feed the ranking into Entropy Pooling","kind":"training","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-listwise-rank-correlation-exact-numeric-forecasts"},{"ref":"TEC-0185","id":"reward-reasoning-whether-editing-improves-image","label":"Reward reasoning by whether editing on it improves the image","summary":"An actor states quality-limiting factors, a frozen diffusion editor edits the image conditioned on that reasoning, a frozen judge scores the quality change, and that change rewards only the reasoning tokens while human ratings reward only t","kind":"training","status":"proposed","addresses":["explanation-faithfulness"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-reasoning-whether-editing-improves-image"},{"ref":"TEC-0186","id":"reward-segments-reference-reasoning-points-reached","label":"Reward segments by reference reasoning points reached","summary":"Split each sampled trajectory into segments and give each segment a reward equal to the increase in the count of reference-solution reasoning points it has reached, with credit shortcut so that reaching the goal counts all points.","kind":"training","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-segments-reference-reasoning-points-reached"},{"ref":"TEC-0183","id":"reward-judge-s-self-generated-rubric-scored-teacher","label":"Reward the judge's self-generated rubric, scored by a teache","summary":"The judge model emits an analysis segment, then task-generic and instruction-specific scoring criteria, then its evaluation, and during GRPO a stronger teacher model scores the generated criteria as one term of the reward.","kind":"training","status":"proposed","addresses":["critique"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-judge-s-self-generated-rubric-scored-teacher"},{"ref":"TEC-0304","id":"reward-gated-skill-distillation-agent-s-own-failed","label":"Reward-gated skill distillation from the agent's own failed ","summary":"The orchestrator answers blind, is shown the gold, classifies the gap as operational/unreachable/already-correct, writes a candidate skill only for operational gaps, and the skill is retained in an external library only if it lifts the vali","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reward-gated-skill-distillation-agent-s-own-failed"},{"ref":"TEC-0187","id":"reweight-jargon-favoring-components-post-hoc","label":"Reweight jargon-favoring components post-hoc","summary":"Decompose output logits into per-attention-head and per-MLP direct contributions and learn scalar component weights on a training split, suppressing components that push toward the jargon label.","kind":"training","status":"proposed","addresses":["domain-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reweight-jargon-favoring-components-post-hoc"},{"ref":"TEC-0305","id":"reweight-reward-dimensions-online-shortfall-volatility-progress","label":"Reweight reward dimensions online by shortfall, volatility a","summary":"Refreshes the scalarization weights every N steps using a multiplicative gate over each dimension's EMA shortfall against the mean, its relative volatility, and its recent progress, raising pressure on lagging, unstable or stagnant dimensio","kind":"training","status":"proposed","addresses":["reward-hacking"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reweight-reward-dimensions-online-shortfall-volatility-progress"},{"ref":"TEC-0306","id":"rewrite-agent-s-tool-use-policy-prompt-realized","label":"Rewrite the agent's tool-use policy prompt from realized out","summary":"A separate Policy Agent reads a batch of decision traces paired with realized portfolio returns and rewrites the trading agent's whole system prompt — tool selection, query formulation, verification and sizing rules — before the next batch","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/rewrite-agent-s-tool-use-policy-prompt-realized"},{"ref":"TEC-0188","id":"role-conditioned-per-vehicle-skill-subsets-peer-messaging","label":"Role-conditioned per-vehicle skill subsets with peer messagi","summary":"A harness parses the mission, assigns each UAV a role, exposes only the top-ranked skills for that role to a vehicle-bound agent, and lets agents exchange directed peer messages while a commander holds no flight authority.","kind":"process","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/role-conditioned-per-vehicle-skill-subsets-peer-messaging"},{"ref":"TEC-0307","id":"role-split-physical-agent-loop-reflection-feedback","label":"Role-split physical agent loop with reflection feedback","summary":"Splits long-horizon physical task management into deterministic perception agents, a synthesizing manager, a decision agent, an independent revision agent that may only adjust quantities, a deterministic safety guard, and a reflection agent","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/role-split-physical-agent-loop-reflection-feedback"},{"ref":"TEC-0189","id":"roll-back-environment-keep-distilled-reflection","label":"Roll back the environment, keep the distilled reflection","summary":"On a triggered review, an agent restores a coarse-to-fine-selected earlier checkpoint while carrying forward a memory of objective, environment facts, past milestones and failure analysis, deliberately omitting current-state claims.","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/roll-back-environment-keep-distilled-reflection"},{"ref":"TEC-0190","id":"route-decoder-ffn-experts-task-identity-tokens","label":"Route decoder FFN experts by task identity, not tokens","summary":"Convert each LLM decoder feed-forward network into one always-active shared expert plus one expert per task, and select a single task expert for the whole video-prompt sample using the task label or nearest-neighbour matching of the prompt","kind":"architecture","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/route-decoder-ffn-experts-task-identity-tokens"},{"ref":"TEC-0308","id":"route-each-agent-step-predicted-task-progress-gain","label":"Route each agent step by predicted task-progress gain","summary":"Scores intermediate workflow state into a progress value, predicts each candidate model's progress gain from structured and semantic features, and picks the model maximizing gain minus time/cost-budget penalties at every dispatch step.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/route-each-agent-step-predicted-task-progress-gain"},{"ref":"TEC-0237","id":"gradient-routing-adapter","label":"Route flagged updates into a removable adapter","summary":"During post-training, send the gradient updates from samples an imperfect monitor flags as unwanted into a separate set of new parameters, then remove those parameters before deployment.","kind":"training","status":"active","addresses":["reward-hacking"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/gradient-routing-adapter"},{"ref":"TEC-0309","id":"route-images-expert-backbones-prompted-mllm","label":"Route images to expert backbones with a prompted MLLM","summary":"An MLLM is prompted with the image plus blur/brightness/contrast/noise statistics and outputs one of five domain labels, dispatching the image to the matching fine-tuned vision expert, with an UNSURE option falling back to the best balanced","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/route-images-expert-backbones-prompted-mllm"},{"ref":"TEC-0191","id":"route-multi-agent-steps-over-gated-execution-memory","label":"Route multi-agent steps over a gated execution memory","summary":"A jointly trained router picks each step's role and backbone from a memory kept compact by learned write and retrieval gates, and halts when the memory holds enough evidence.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/route-multi-agent-steps-over-gated-execution-memory"},{"ref":"TEC-0310","id":"route-prompts-align-lora-them-verify-public-tests","label":"Route prompts, align LoRA to them, verify on public tests","summary":"Picks a task-specific prompt template by keyword heuristics, fine-tunes a LoRA adapter on training inputs rewritten into that same template style, then samples several programs and returns the one passing the most visible public tests.","kind":"process","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/route-prompts-align-lora-them-verify-public-tests"},{"ref":"TEC-0085","id":"model-routing","label":"Route routine calls to a smaller model","summary":"Send the short, structured, routine calls in an agent pipeline to a small cheap model and reserve a frontier model for the steps that need deep planning or sustained constraint tracking.","kind":"process","status":"active","addresses":["tool-use","long-horizon-execution","planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/model-routing"},{"ref":"TEC-0192","id":"route-token-token-between-base-model-sft-expert","label":"Route token by token between base model and SFT expert","summary":"Train a lightweight two-stage router on the base model's hidden states, labelled by tokens where the base's greedy prediction is wrong and the expert's is right, then at decoding mix the two next-token distributions with momentum smoothing","kind":"decoding","status":"proposed","addresses":["fine-tuning-retention"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/route-token-token-between-base-model-sft-expert"},{"ref":"TEC-0065","id":"routed-on-policy-distillation-safety-task-teachers","label":"Routed on-policy distillation from safety and task teachers","summary":"Sample from the compromised student, route each token to either the original aligned model (refusal prior) or the fine-tuned task model, and train the student with a top-K KL loss toward the routed teacher, so realignment targets the distri","kind":"training","status":"proposed","addresses":["fine-tuning-retention"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/routed-on-policy-distillation-safety-task-teachers"},{"ref":"TEC-0311","id":"scaffold-diagnosis-evidence-tools-coverage-based-stopping-harness","label":"Scaffold diagnosis with evidence tools and a coverage-based ","summary":"Builds patient-specific interactive environments, routes diagnosis through role-specialized agents querying evidence tools against an evolving evidence state, and uses an observer that tracks hypothesis uncertainty and evidence coverage to","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/scaffold-diagnosis-evidence-tools-coverage-based-stopping-harness"},{"ref":"TEC-0023","id":"secure-code-scanning","label":"Scan generated code for insecure patterns","summary":"Run a static insecure-pattern scanner on every generated change before it is accepted.","kind":"tooling","status":"active","addresses":["secure-coding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/secure-code-scanning"},{"ref":"TEC-0312","id":"schema-validated-tool-calls-plus-verify-then-act","label":"Schema-validated tool calls plus verify-then-act motion chec","summary":"Restrict the LLM to typed tool calls carrying only semantic object references and bounded skill parameters, bind each accepted call to a fixed motion-planning template, and execute only trajectories that pass kinematic and collision checks","kind":"tooling","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/schema-validated-tool-calls-plus-verify-then-act"},{"ref":"TEC-0066","id":"score-critique-repair-enables","label":"Score a critique by the repair it enables","summary":"Instead of rating a critique's text, hand it to a fixed repair agent and measure the improvement of the artefact after repair, so a critique is judged by what a reader can do with it.","kind":"process","status":"proposed","addresses":["critique"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-critique-repair-enables"},{"ref":"TEC-0313","id":"score-argument-value-language-consistency-own-metric","label":"Score argument-value language consistency as its own metric","summary":"Decompose tool-call evaluation into a conditional hierarchy — invocation detection, tool selection, argument completion, argument language consistency, full call match — so wrong-language argument values are separated from structural errors","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-argument-value-language-consistency-own-metric"},{"ref":"TEC-0193","id":"score-audio-cots-grounding-coherence-causality","label":"Score audio CoTs on grounding, coherence and causality","summary":"Prompts the model with a four-step Perception/Analysis/Reasoning/Answer template, then scores the chain with six automatic metrics (CLAP audio-entity grounding, inter-step embedding coherence, small-LM substantiveness, perplexity-based step","kind":"process","status":"proposed","addresses":["audio-understanding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-audio-cots-grounding-coherence-causality"},{"ref":"TEC-0194","id":"score-cited-factors-whether-paired-intervention-moves-prediction","label":"Score cited factors by whether the paired intervention moves","summary":"For each controlled pair that changes one cited factor, check both whether the explanation mentions that factor and whether both endpoint predictions follow the expected transition, reporting the conditional rate (CSDR).","kind":"process","status":"proposed","addresses":["explanation-faithfulness"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-cited-factors-whether-paired-intervention-moves-prediction"},{"ref":"TEC-0314","id":"score-coordination-constructs-separately-task-success","label":"Score coordination constructs separately from task success","summary":"Alongside the binary goal check, compute per-trajectory construct scores from simulator-logged events — makespan versus an oracle reference, precedence-violation rate with a legality gate, shared-resource collision rate, and buffer overflow","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-coordination-constructs-separately-task-success"},{"ref":"TEC-0195","id":"score-detectors-paired-factual-counterfactual-decisions","label":"Score detectors on paired factual/counterfactual decisions","summary":"Evaluate a hallucination detector by requiring it to accept the factual answer and reject the minimally edited counterfactual for the same video, reporting the joint success rate rather than one-sided accuracy.","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-detectors-paired-factual-counterfactual-decisions"},{"ref":"TEC-0316","id":"score-forecasts-live-post-cutoff-reporting-cost-accuracy","label":"Score forecasts live post-cutoff, reporting cost and accurac","summary":"Evaluate forecasting agents only on questions resolving after the model's training cutoff, with ablations of the LLM component and joint reporting of cost and accuracy.","kind":"process","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-forecasts-live-post-cutoff-reporting-cost-accuracy"},{"ref":"TEC-0196","id":"score-forecasts-only-live-post-cutoff-questions-cost","label":"Score forecasts only on live post-cutoff questions, with cos","summary":"Evaluate forecasting agents on continuously refreshed questions whose outcomes resolve after the training cutoff, with component ablations and cost reported alongside accuracy.","kind":"process","status":"proposed","addresses":["forecasting"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-forecasts-only-live-post-cutoff-questions-cost"},{"ref":"TEC-0040","id":"score-intermediate-repair-artifacts-just-patch","label":"Score intermediate repair artifacts, not just the patch","summary":"Require the repair model to also produce checkable intermediate artifacts (which tests trigger the bug, which lines execute before and after the patch, new bug-exposing tests) and grade those against execution evidence as a reliability sign","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-intermediate-repair-artifacts-just-patch"},{"ref":"TEC-0317","id":"score-itineraries-buffer-adequacy-against-empirical-delay-data","label":"Score itineraries for buffer adequacy against empirical dela","summary":"Evaluate a generated plan by whether each activity gap and transport segment allocates buffer inside a risk-profile-scaled range derived from historical delay distributions and hourly crowd density, rather than by binary constraint satisfac","kind":"process","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-itineraries-buffer-adequacy-against-empirical-delay-data"},{"ref":"TEC-0318","id":"score-long-horizon-agents-separate-capability-axes-one","label":"Score long-horizon agents on separate capability axes, not o","summary":"Alongside the accumulated primary score, measure negotiation surplus, fraud spend share, drawdown over peak, profit per tool call, controllable return rate, and repeat-order price discipline as separate ranked axes.","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-long-horizon-agents-separate-capability-axes-one"},{"ref":"TEC-0319","id":"score-occurrence-count-separately-boundary-overlap","label":"Score occurrence count separately from boundary overlap","summary":"Report, alongside union-IoU, the share of queries where the number of predicted intervals equals the number of ground-truth intervals, plus under- and over-reporting rates, so missing occurrences is distinguished from imprecise boundaries.","kind":"process","status":"proposed","addresses":["audio-understanding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-occurrence-count-separately-boundary-overlap"},{"ref":"TEC-0197","id":"score-proof-answers-binary-pedagogical-rubric-n-grams","label":"Score proof answers with a binary pedagogical rubric, not n-","summary":"Replace lexical-overlap scoring with the mean of six binary per-response criteria (step enumeration, notation correctness, worked example, sub-question coverage, explanation depth, proof-step granularity).","kind":"process","status":"proposed","addresses":["evaluation-validity"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-proof-answers-binary-pedagogical-rubric-n-grams"},{"ref":"TEC-0198","id":"score-reasoning-failure-cross-layer-attention-entropy-dispersion","label":"Score reasoning failure by cross-layer attention-entropy dis","summary":"At each generation step, average attention entropy per layer, take the standard deviation across layers, and aggregate these over the sequence as features for a logistic-regression hallucination classifier.","kind":"tooling","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-reasoning-failure-cross-layer-attention-entropy-dispersion"},{"ref":"TEC-0199","id":"score-reasoning-steps-grounding-final-answer-match","label":"Score reasoning steps for grounding, not final-answer match","summary":"Train a generative process reward model that labels each interleaved reason/retrieve step for logical validity and evidential grounding, use it in value tree search to build step-level chosen/rejected pairs, and optimize the policy with ste","kind":"training","status":"proposed","addresses":["multi-step-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-reasoning-steps-grounding-final-answer-match"},{"ref":"TEC-0041","id":"score-recovery-model-s-own-real-failure-states","label":"Score recovery from the model's own real failure states","summary":"Build evaluation scenarios by replaying naturally occurring failure states from policy rollouts and measuring whether the policy still completes the original task, graded by how much state restoration the failure demands.","kind":"process","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-recovery-model-s-own-real-failure-states"},{"ref":"TEC-0320","id":"score-restraint-confidence-separately-against-matched-answerable-control","label":"Score restraint and confidence separately against a matched ","summary":"Classifies each response to an unanswerable boundary question as abstain, hedge, confident-correct or confident-wrong, computes calibration only on substantive answers, and pairs each context-withheld item with a restored-document version t","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-restraint-confidence-separately-against-matched-answerable-control"},{"ref":"TEC-0200","id":"score-successful-trajectories-excess-turns-tool-calls","label":"Score successful trajectories by excess turns and tool calls","summary":"A success-gated metric multiplicatively discounts a completed trajectory for tool calls and user-facing turns beyond per-task human reference floors, with the two penalty weights fitted from blinded human pairwise efficiency preferences via","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-successful-trajectories-excess-turns-tool-calls"},{"ref":"TEC-0315","id":"score-decision-pathway-just-outcome","label":"Score the decision pathway, not just the outcome","summary":"Extract guideline decision trees into reference step sequences, then score a model's generated pathway by LLM-judged step-level match aggregated with dynamic time warping against the reference.","kind":"process","status":"proposed","addresses":["domain-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-decision-pathway-just-outcome"},{"ref":"TEC-0321","id":"score-uncertainty-complexity-merged-reasoning-step-graph","label":"Score uncertainty by complexity of a merged reasoning-step g","summary":"Sample several chains of thought, merge semantically equivalent steps into a DAG using a bidirectional-entailment NLI model, and score uncertainty by graph width, height, or propagated per-step token uncertainty.","kind":"decoding","status":"proposed","addresses":["multi-step-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-uncertainty-complexity-merged-reasoning-step-graph"},{"ref":"TEC-0201","id":"score-whole-agent-trajectories-five-safety-outcomes","label":"Score whole agent trajectories on five safety outcomes","summary":"Run adaptive adversarial and matched benign user agents against a stateful tool environment and label each complete trajectory Safe Completion, Correct Refusal, Unsafe Completion, Over-Refusal, or Indeterminate using executed tool actions a","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/score-whole-agent-trajectories-five-safety-outcomes"},{"ref":"TEC-0322","id":"search-hierarchical-temporal-tree-video-segments-coarse-to","label":"Search a hierarchical temporal tree of video segments coarse","summary":"Build a question-agnostic tree of contiguous video segments by recursive temporally constrained clustering with one key frame per node, then run an answer-retrieve-explore loop that expands only branches the model says it needs.","kind":"retrieval","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/search-hierarchical-temporal-tree-video-segments-coarse-to"},{"ref":"TEC-0202","id":"select-tiny-sql-tuning-set-schema-coverage-size","label":"Select a tiny SQL tuning set for schema coverage, not size","summary":"Filter candidate text-to-SQL examples by dual-model difficulty, attach the shortest correct chain-of-thought, score each on four LLM-judged quality dimensions, then use a genetic search to pick a small subset maximizing table/column coverag","kind":"training","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/select-tiny-sql-tuning-set-schema-coverage-size"},{"ref":"TEC-0323","id":"select-frames-shape-relevance-curve-height","label":"Select frames by the shape of the relevance curve, not its h","summary":"Smooths the frame-query similarity sequence, segments it by slope and curvature into peak, rising, falling, boundary and background regions, allocates the frame budget across regions using a question-derived preference vector, and picks fra","kind":"retrieval","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/select-frames-shape-relevance-curve-height"},{"ref":"TEC-0203","id":"select-training-data-quality-structural-diversity","label":"Select training data by quality and structural diversity","summary":"Filter Verilog training samples by testbench-verified correctness and instruction-following-difficulty below 1, then cluster on fused text plus 109-dimensional AST/CFG/netlist features and sample proportionally per cluster by IFD rank.","kind":"training","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/select-training-data-quality-structural-diversity"},{"ref":"TEC-0204","id":"separate-boundary-awareness-boundary-compliance-probes","label":"Separate boundary-awareness from boundary-compliance probes","summary":"Ask the same role-played model a binary question about whether a target is inside the character's knowledge boundary, then separately ask a concrete factual multiple-choice question with an abstention option, and pair the two outcomes per i","kind":"process","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/separate-boundary-awareness-boundary-compliance-probes"},{"ref":"TEC-0025","id":"structured-prompt-separation","label":"Separate instructions from data","summary":"Mark the boundary between trusted instructions and untrusted content, and train or prompt the model to honor it.","kind":"training","status":"active","addresses":["prompt-injection"],"counts":{"claims":1,"sources":2},"url":"https://rsiratchet.com/techniques/structured-prompt-separation"},{"ref":"TEC-0205","id":"separate-self-judged-scores-verifier-rewards-measure-both","label":"Separate self-judged scores from verifier rewards and measur","summary":"Withhold verifier-computed rewards from the agent during exploration while logging them benchmark-side, so the agent's own step scores can be compared against ground truth and against later held-out improvement.","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/separate-self-judged-scores-verifier-rewards-measure-both"},{"ref":"TEC-0013","id":"independent-verifier","label":"Separate the verifier from the producer","summary":"The agent that produced an artifact does not judge it; a deterministic sensor or a separate verifier agent does, and reports rather than rewrites.","kind":"process","status":"active","addresses":["self-repair","judge-position-bias"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/independent-verifier"},{"ref":"TEC-0324","id":"shared-semantic-blueprint-parallel-restructuring-branches-repair-execution","label":"Shared semantic blueprint, parallel restructuring branches, ","summary":"One agent reconstructs the source program's states, operators and dependencies into a structured representation and a target-side blueprint; several agents then generate alternative tensor realizations under that blueprint, failures are rep","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/shared-semantic-blueprint-parallel-restructuring-branches-repair-execution"},{"ref":"TEC-0003","id":"chain-of-thought","label":"Show intermediate steps","summary":"Prompt the model to write out intermediate reasoning steps before the final answer.","kind":"prompting","status":"active","addresses":["arithmetic"],"counts":{"claims":1,"sources":1},"url":"https://rsiratchet.com/techniques/chain-of-thought"},{"ref":"TEC-0206","id":"show-llm-bayesian-belief-state-history","label":"Show the LLM a Bayesian belief state, not the history","summary":"An external module runs a two-step Bayes filter over a specified POMDP and passes only the normalized posterior over latent states to the LLM, which selects an action from that posterior with no access to the action-observation log.","kind":"architecture","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/show-llm-bayesian-belief-state-history"},{"ref":"TEC-0325","id":"shrink-patch-after-generation-during","label":"Shrink the patch after generation, not during it","summary":"A small refiner model (SFT plus DPO on verbose-to-concise patch pairs with distilled reasoning traces) is plugged in after an existing repair framework's patch generation to rewrite the candidate patch into a smaller one, with optional judg","kind":"process","status":"proposed","addresses":["cutting-to-the-chase"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/shrink-patch-after-generation-during"},{"ref":"TEC-0207","id":"size-ensemble-marginal-gain-benefit-yield-curve","label":"Size an ensemble by its marginal-gain (benefit yield) curve","summary":"Treat ensemble performance as a function of member count and stop adding models at the size where marginal gain per added model crosses zero.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/size-ensemble-marginal-gain-benefit-yield-curve"},{"ref":"TEC-0208","id":"slow-chunk-proposal-plus-fast-rl-trained-action","label":"Slow chunk proposal plus fast RL-trained action edit","summary":"A large VLA asynchronously samples candidate action chunks from an older observation, then a lightweight edit policy conditioned on the latest observation perturbs each candidate and a learned Q-function picks the best chunk to execute.","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/slow-chunk-proposal-plus-fast-rl-trained-action"},{"ref":"TEC-0326","id":"split-forensic-analysis-into-per-cue-agents-reconcile","label":"Split forensic analysis into per-cue agents, reconcile with ","summary":"Four MLLM agents each inspect the same sampled video frames for one forgery cue type and report observations without a verdict; a judge agent weighs the four reports (optionally plus the frames) and issues the real/fake decision and explana","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/split-forensic-analysis-into-per-cue-agents-reconcile"},{"ref":"TEC-0327","id":"split-streaming-video-into-latency-frontend-asynchronous-memory","label":"Split streaming video into a latency frontend and an asynchr","summary":"Frontend workers handle dispatch and proactive monitoring while backend workers asynchronously build a hierarchical event and entity memory with retained key frames, and route recall and external search on demand.","kind":"architecture","status":"proposed","addresses":["video-temporal-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/split-streaming-video-into-latency-frontend-asynchronous-memory"},{"ref":"TEC-0061","id":"split-text-to-sql-evaluation-database-question","label":"Split text-to-SQL evaluation by database, not by question","summary":"Evaluate SQL synthesis on databases and query patterns that never appear in training, so memorized question-query templates cannot be reused.","kind":"process","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/split-text-to-sql-evaluation-database-question"},{"ref":"TEC-0210","id":"stage-ego-to-robot-transfer-through-aligned-demo","label":"Stage ego-to-robot transfer through an aligned-demo curricul","summary":"Pretrain on egocentric data, then fine-tune on a small set of task-matched human and robot demonstrations converted to a shared relative end-effector representation, then pretrain on robot data before task post-training.","kind":"training","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/stage-ego-to-robot-transfer-through-aligned-demo"},{"ref":"TEC-0211","id":"stage-proof-explore-gate-decompose-dependency-re-verify","label":"Stage the proof: explore, gate, decompose by dependency, re-","summary":"Allocates inference across explicit research stages — strategy exploration with a readiness gate, decomposition into a dependency graph of section-level subproblems solved in parallel, and global verification whose findings are routed back","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/stage-proof-explore-gate-decompose-dependency-re-verify"},{"ref":"TEC-0209","id":"stage-aligned-perturbation-cascade-aware-failure-attribution","label":"Stage-aligned perturbation with cascade-aware failure attrib","summary":"Perturb each stage of the tool-call pipeline separately (interface, intent, returned evidence, runtime feedback) and backtrack each observed failure deterministically to the earliest stage that caused it.","kind":"process","status":"proposed","addresses":["tool-use"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/stage-aligned-perturbation-cascade-aware-failure-attribution"},{"ref":"TEC-0076","id":"stepwise-self-scoring-failure-detection","label":"Stepwise self-scoring for failure detection","summary":"After a multi-step answer, ask the model to rate each step's correctness separately and combine the step scores, rather than asking for one confidence on the final answer.","kind":"prompting","status":"proposed","addresses":["multi-step-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/stepwise-self-scoring-failure-detection"},{"ref":"TEC-0212","id":"stop-self-critique-loop-confirmed-sentinel","label":"Stop the self-critique loop on a CONFIRMED sentinel","summary":"Run generate-critique-revise to a fixed max depth but exit as soon as the critique emits a CONFIRMED token, instead of trusting a self-reported numeric confidence tag.","kind":"prompting","status":"proposed","addresses":["self-repair"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/stop-self-critique-loop-confirmed-sentinel"},{"ref":"TEC-0328","id":"store-jailbreak-failures-wrapper-level-rules-persistent-memory","label":"Store jailbreak failures as wrapper-level rules in persisten","summary":"On a detected unsafe response, an LLM distills the attack's structural wrapper into a labelled rule in an external memory, and later inputs are classified into a wrapper label that selectively triggers at most two matching rules before gene","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/store-jailbreak-failures-wrapper-level-rules-persistent-memory"},{"ref":"TEC-0329","id":"store-moe-gating-profiles-retrievable-skill-memory","label":"Store MoE gating profiles as retrievable skill memory","summary":"Partition demonstrations into latent skills using Mixture-of-Experts gating coefficients, store each segment's mean latent feature as a key and its gating-weight sequence as a value, and at inference retrieve the top-N similar entries and a","kind":"architecture","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/store-moe-gating-profiles-retrievable-skill-memory"},{"ref":"TEC-0213","id":"stratify-rollouts-confidence-outcome-forming-rl-groups","label":"Stratify rollouts by confidence and outcome when forming RL ","summary":"Within each RL step, bucket the step's own rollouts by length-normalized surprisal and reward into anchors, hard negatives and low-confidence successes, and assemble half the optimization groups from a fixed ratio of these buckets before cl","kind":"training","status":"proposed","addresses":["fine-tuning-retention"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/stratify-rollouts-confidence-outcome-forming-rl-groups"},{"ref":"TEC-0330","id":"strip-prior-score-metadata-judge-prompts","label":"Strip prior-score metadata from judge prompts","summary":"Remove revision, attempt-index and prior-score fields from the judge's context so each evaluation sees only task, answer and rubric.","kind":"prompting","status":"proposed","addresses":["critique"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/strip-prior-score-metadata-judge-prompts"},{"ref":"TEC-0077","id":"style-neutralization-before-safety-model-sees-prompt","label":"Style neutralization before the safety model sees the prompt","summary":"Pass user input through a secondary model that rewrites it in a neutral register, removing fear, urgency, flattery and similar cues, and give the neutralized text to the main model, so stylistic reframing no longer carries the request past","kind":"process","status":"proposed","addresses":["jailbreak-robustness"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/style-neutralization-before-safety-model-sees-prompt"},{"ref":"TEC-0214","id":"summarize-long-context-parallel-hierarchical-tree","label":"Summarize long context as a parallel hierarchical tree","summary":"Treats a query-prompted frozen LLM as a convolution kernel that summarizes overlapping text windows in parallel, recursing over summaries layer by layer, with high-relevance raw segments skipped forward to the final answer and a separate ch","kind":"prompting","status":"proposed","addresses":["long-context-degradation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/summarize-long-context-parallel-hierarchical-tree"},{"ref":"TEC-0215","id":"supervise-abstention-evidence-ablated-inputs","label":"Supervise abstention on evidence-ablated inputs","summary":"Fine-tune a verifier on paired inputs — full context with the gold label and evidence, and the same context with gold evidence removed and target NOT_ENOUGH_INFO — with gold evidence sets first expanded by majority vote of three other LLMs.","kind":"training","status":"proposed","addresses":["fact-verification"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/supervise-abstention-evidence-ablated-inputs"},{"ref":"TEC-0216","id":"suppress-hallucinated-logit-directions-re-anchor-verified-sentences","label":"Suppress hallucinated logit directions and re-anchor on veri","summary":"At each sentence, an evaluator localizes hallucinated claim tokens into a hallucination memory used to build a vocabulary-level suppression direction applied to the logits, while the lowest-hallucination sentence so far is kept in an anchor","kind":"decoding","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/suppress-hallucinated-logit-directions-re-anchor-verified-sentences"},{"ref":"TEC-0331","id":"switch-dense-sparse-agent-topology-disagreement-falls","label":"Switch from dense to sparse agent topology as disagreement f","summary":"The harness selects all-to-all inter-agent communication while aggregate disagreement exceeds a threshold and a ring topology once it falls below, with a token budget decremented each round and a guard that halts on exhaustion.","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/switch-dense-sparse-agent-topology-disagreement-falls"},{"ref":"TEC-0050","id":"synthesise-extractor-highlighted-examples","label":"Synthesise an extractor from highlighted examples","summary":"The user highlights a few instances of each field and a program-synthesis engine infers an extraction program in a restricted DSL that is then run over the whole document.","kind":"tooling","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/synthesise-extractor-highlighted-examples"},{"ref":"TEC-0332","id":"target-guidelines-flip-prone-steps-found-resampling","label":"Target guidelines at flip-prone steps found by resampling","summary":"An offline analyzer re-issues each recorded prompt N times and scores response agreement per step; steps below a consistency threshold are turned by an LLM into short guidelines stored in episodic memory and retrieved into future runs of si","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/target-guidelines-flip-prone-steps-found-resampling"},{"ref":"TEC-0078","id":"task-method-knowledge-prompting","label":"Task-Method-Knowledge prompting","summary":"Give the model an explicit hierarchical model of the task: the goals, the methods that achieve each, and the knowledge of why each step is taken, drawn from the TMK framework in educational science, before asking it to plan.","kind":"prompting","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/task-method-knowledge-prompting"},{"ref":"TEC-0042","id":"taxonomy-driven-hallucination-injection-review-verification-benchmarks","label":"Taxonomy-driven hallucination injection for review verificat","summary":"Taxonomy-driven hallucination injection for review verification benchmarks, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/taxonomy-driven-hallucination-injection-review-verification-benchmarks"},{"ref":"TEC-0012","id":"hierarchical-memory","label":"Tiered memory managed by the model","summary":"Keep a small working context and let the model page facts in and out of external storage.","kind":"architecture","status":"active","addresses":["long-term-memory","context-hygiene"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/hierarchical-memory"},{"ref":"TEC-0079","id":"tool-feedback-autoformalization-compiler-judge-panel","label":"Tool-feedback autoformalization with compiler and judge pane","summary":"Let the formalizer emit a Lean statement, compile it, and ask a panel of models whether it matches the informal statement; feed both results back and let the model revise, training it on such tool-calling traces so revisions become effectiv","kind":"process","status":"proposed","addresses":["formal-math"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/tool-feedback-autoformalization-compiler-judge-panel"},{"ref":"TEC-0220","id":"train-world-model-controller-vlm-relabelled-rollouts-then","label":"Train a world-model controller on VLM-relabelled rollouts, t","summary":"Relabel segments of a world-model policy's own on-policy rollouts with VLM-generated instructions, add a behaviour-cloning plus instruction-completion loss to the Dreamer objectives, and at inference let a frozen VLM planner emit sparse tex","kind":"architecture","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/train-world-model-controller-vlm-relabelled-rollouts-then"},{"ref":"TEC-0006","id":"constitutional-training","label":"Train against explicit principles","summary":"Use a written set of principles to generate critiques and preferences, then train on them.","kind":"training","status":"active","addresses":["goal-conflict-safety"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/constitutional-training"},{"ref":"TEC-0217","id":"train-llm-write-solver-checked-mip-programs-task","label":"Train an LLM to write solver-checked MIP programs from task ","summary":"Supervised fine-tuning on single-modal MIP program demonstrations followed by GRPO with rewards for output format, discretization-API coverage, successful solver execution, and solver-verified trajectory correctness.","kind":"training","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/train-llm-write-solver-checked-mip-programs-task"},{"ref":"TEC-0333","id":"train-evidence-absent-verification-through-matched-easier-task","label":"Train evidence-absent verification through matched easier ta","summary":"Rewrite each annotated paper error into a chain of three queries — error asserted with evidence given, candidate evidence only, and neither — and train on all three stages jointly rather than only on the hardest one.","kind":"training","status":"proposed","addresses":["error-detection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/train-evidence-absent-verification-through-matched-easier-task"},{"ref":"TEC-0022","id":"reverse-training","label":"Train on reversed text","summary":"Include reversed word or entity sequences in pretraining so facts are learned in both directions.","kind":"training","status":"active","addresses":["reversal-curse"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/reverse-training"},{"ref":"TEC-0051","id":"train-semantic-equivalence-string-match","label":"Train on semantic equivalence, not string match","summary":"Replaces the MLE objective with a reward that maximizes expected semantic correctness of the generated regex, measured by a DFA-equivalence oracle plus random and distinguishing test cases.","kind":"training","status":"proposed","addresses":["formal-query-synthesis"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/train-semantic-equivalence-string-match"},{"ref":"TEC-0334","id":"train-only-central-coordinating-agent-freeze-sub-agents","label":"Train only the central coordinating agent, freeze sub-agents","summary":"Treats sub-agents as part of the environment and applies RL gradients only to the central agent's policy, with trajectories reconstructed from captured per-session interface call records.","kind":"training","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/train-only-central-coordinating-agent-freeze-sub-agents"},{"ref":"TEC-0219","id":"train-planner-executor-separate-role-adapters","label":"Train planner and executor as separate role adapters","summary":"Fit one LoRA adapter by sequence-level SFT on reference tool sequences for workflow planning and a second by node-level GRPO with argument-match rewards for runtime argument grounding, so workflow errors and argument errors get separate cre","kind":"training","status":"proposed","addresses":["planning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/train-planner-executor-separate-role-adapters"},{"ref":"TEC-0218","id":"train-orchestrator-descriptor-conditioned-routing-plans","label":"Train the orchestrator on descriptor-conditioned routing pla","summary":"Fine-tune the planning model to emit a rationale plus an ordered sub-agent plan from tool descriptors, then refine with GRPO rewarding parseable output, coverage F1 against the target plan, and satisfied precedence edges.","kind":"training","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/train-orchestrator-descriptor-conditioned-routing-plans"},{"ref":"TEC-0335","id":"train-step-verifier-only-symbolically-valid-wrong-steps","label":"Train the step verifier only on symbolically-valid wrong ste","summary":"Generate hard negatives by swapping the logical principle or a unit-matched operand and recomputing values with a solver so they pass a deterministic unit/arithmetic verifier, train the PRM on those pairs, and at inference prune with the ve","kind":"training","status":"proposed","addresses":["error-detection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/train-step-verifier-only-symbolically-valid-wrong-steps"},{"ref":"TEC-0004","id":"checklist-decomposition","label":"Turn the procedure into a checklist","summary":"Convert constraints and steps into an explicit checklist the model must tick off before finishing.","kind":"process","status":"active","addresses":["procedure-following"],"counts":{"claims":1,"sources":2},"url":"https://rsiratchet.com/techniques/checklist-decomposition"},{"ref":"TEC-0043","id":"two-axis-identity-opinion-sycophancy-probe","label":"Two-axis identity/opinion sycophancy probe","summary":"Two-axis identity/opinion sycophancy probe, as introduced by the paper it was drafted from.","kind":"process","status":"proposed","addresses":["sycophancy"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/two-axis-identity-opinion-sycophancy-probe"},{"ref":"TEC-0062","id":"two-phase-call-graph-search-files-broad-then","label":"Two-phase call-graph search, files broad then functions deep","summary":"Start from files the issue mentions, build a module-level call graph dynamically and explore it broadly to pick files, then expand only the chosen files into a function call graph and search it iteratively, with the graph structure steering","kind":"process","status":"proposed","addresses":["repository-navigation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/two-phase-call-graph-search-files-broad-then"},{"ref":"TEC-0221","id":"two-stage-financial-domain-adaptation-corpus-pre-train","label":"Two-stage financial domain adaptation: corpus pre-train, the","summary":"Pre-train the model on large-scale public financial corpora with a loss penalizing valuations inconsistent with disclosed metrics, then LoRA/prompt-tune on an institution's own investment memos and risk rules.","kind":"training","status":"proposed","addresses":["domain-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/two-stage-financial-domain-adaptation-corpus-pre-train"},{"ref":"TEC-0336","id":"unify-offer-schema-chooser-then-flag-treatment","label":"Unify offer schema and chooser, then flag the treatment","summary":"Apply the policy treatment as post-parsing boolean flags over one shared message schema and one shared buyer decision rule, average repeated generations within profile-condition, and bootstrap over profiles rather than generation rows.","kind":"process","status":"proposed","addresses":["evaluation-validity"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/unify-offer-schema-chooser-then-flag-treatment"},{"ref":"TEC-0222","id":"update-knowledge-synthetic-news-mid-training-plus-abstention","label":"Update knowledge by synthetic news mid-training plus abstent","summary":"A teacher model turns a new-event knowledge graph into long-form synthetic news articles for next-token mid-training, plus preference pairs that reward chain-of-thought answers and punish outdated values, hallucinations, and unwarranted abs","kind":"training","status":"proposed","addresses":["knowledge-updating"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/update-knowledge-synthetic-news-mid-training-plus-abstention"},{"ref":"TEC-0223","id":"update-only-stable-neurons-anchor-their-activations","label":"Update only stable neurons and anchor their activations","summary":"Select FFN neurons with low activation coefficient of variation, mask gradients to those neurons only, and after cross-entropy converges add an MSE term pulling current activations toward a pre-computed activation anchor.","kind":"training","status":"proposed","addresses":["fine-tuning-retention"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/update-only-stable-neurons-anchor-their-activations"},{"ref":"TEC-0226","id":"validate-whole-perturbation-region-don-t-sample","label":"Validate a whole perturbation region, don't sample it","summary":"Maps a bounded brightness/rotation perturbation region onto a Hilbert curve and searches it by Hölder optimisation with a margin-aware convergence rule, so each instance returns either a concrete counterexample or a bound covering every per","kind":"process","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/validate-whole-perturbation-region-don-t-sample"},{"ref":"TEC-0224","id":"validate-hls-generation-board-execution-simulation","label":"Validate HLS generation to board execution, not simulation","summary":"A multi-agent loop that injects Vitis pragma documentation, evolves kernel candidates under simulation-performance selection, verifies functional correctness on a de-pragma'd C++ harness with targeted repair, then generates host code and co","kind":"process","status":"proposed","addresses":["code-generation"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/validate-hls-generation-board-execution-simulation"},{"ref":"TEC-0225","id":"validate-navigation-proposals-against-evidence-geometry-before-dispatch","label":"Validate navigation proposals against evidence and geometry ","summary":"A harness layer intercepts the MLLM planner's tool calls, checks each target for supporting observation provenance, geometric reachability and consistency with the active subgoal, and blocks stop requests until semantic, geometric and progr","kind":"process","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/validate-navigation-proposals-against-evidence-geometry-before-dispatch"},{"ref":"TEC-0337","id":"vary-personalization-depth-cumulative-workplace-detail-tiers","label":"Vary personalization depth in cumulative workplace-detail ti","summary":"Generate phishing pretexts at four cumulative levels of recipient work-context detail and compare rated convincingness and stated action within participants.","kind":"prompting","status":"proposed","addresses":["persuasion"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/vary-personalization-depth-cumulative-workplace-detail-tiers"},{"ref":"TEC-0063","id":"verbalized-confidence-after-listing-alternatives","label":"Verbalized confidence after listing alternatives","summary":"Prompt the model to produce several candidate answers and then state a numeric confidence for each, and use the stated number rather than token log-probabilities as the confidence signal for deferral or abstention.","kind":"prompting","status":"proposed","addresses":["abstention-and-self-knowledge"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/verbalized-confidence-after-listing-alternatives"},{"ref":"TEC-0227","id":"verify-candidates-step-boundary-hidden-states-text","label":"Verify candidates from step-boundary hidden states, not text","summary":"Extract the frozen generator's last-layer hidden states at reasoning-step delimiters, feed the step sequence to a 2-layer Transformer encoder, and score candidates with a tie-safe pairwise ranking loss trained on self-sampled trajectories w","kind":"architecture","status":"proposed","addresses":["error-detection"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/verify-candidates-step-boundary-hidden-states-text"},{"ref":"TEC-0228","id":"verify-each-intermediate-step-against-domain-constraints-then","label":"Verify each intermediate step against domain constraints, th","summary":"After every tool call, a manager role re-reads structured global memory and classifies the output as pass/uncertain/fail on schema validity, sub-goal consistency and domain plausibility, replanning or re-routing on failure.","kind":"process","status":"proposed","addresses":["long-horizon-execution"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/verify-each-intermediate-step-against-domain-constraints-then"},{"ref":"TEC-0229","id":"verify-each-reasoning-step-solver-backtrack-failure","label":"Verify each reasoning step with a solver and backtrack on fa","summary":"Decompose each generated reasoning step into context, implicit assumptions and conclusion, formalize it into a solver specification, score it by premise fidelity plus solver entailment, and regenerate from that step when the score falls bel","kind":"decoding","status":"proposed","addresses":["multi-step-reasoning"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/verify-each-reasoning-step-solver-backtrack-failure"},{"ref":"TEC-0338","id":"verify-harness-edits-behavior-matched-tasks-attribution-gate","label":"Verify harness edits on behavior-matched tasks with an attri","summary":"For each proposed harness modification, select verification tasks linked to the supporting trajectories plus regression-risk tasks, run both old and new harness on them, and accept only if the trajectory evidence attributes the improvement","kind":"process","status":"proposed","addresses":["harness-self-improvement"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/verify-harness-edits-behavior-matched-tasks-attribution-gate"},{"ref":"TEC-0230","id":"verify-object-claims-cross-query-semantic-spatial-agreement","label":"Verify object claims by cross-query semantic and spatial agr","summary":"For each object claim, query a frozen open-vocabulary detector with four paraphrased prompts, average the top scores for semantic support, score whether the queries' preferred boxes form one overlapping cluster with a dominant candidate, fu","kind":"tooling","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/verify-object-claims-cross-query-semantic-spatial-agreement"},{"ref":"TEC-0007","id":"dependency-existence-check","label":"Verify packages exist before installing","summary":"Check every suggested dependency against the registry and an allowlist before it is installed.","kind":"process","status":"active","addresses":["secure-coding"],"counts":{"claims":1,"sources":2},"url":"https://rsiratchet.com/techniques/dependency-existence-check"},{"ref":"TEC-0044","id":"verify-responses-against-llm-built-sql-database-sources","label":"Verify responses against an LLM-built SQL database of the so","summary":"Prompt the model to incrementally build a relational SQL database from the reference passages, generate SELECT queries for the claims in the response, and reconcile its initial neural hallucination judgement with the executed query results.","kind":"tooling","status":"proposed","addresses":["hallucination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/verify-responses-against-llm-built-sql-database-sources"},{"ref":"TEC-0339","id":"weight-peers-cognitive-snr-divergence","label":"Weight peers by cognitive SNR, not by divergence","summary":"Each incoming agent message is weighted by the sender's asserted belief precision divided by the KL divergence between its reported belief and what the receiver's second-order model predicts a peer with that telemetry access should have rep","kind":"process","status":"proposed","addresses":["multi-model-coordination"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/weight-peers-cognitive-snr-divergence"},{"ref":"TEC-0231","id":"weight-profile-attributes-temporal-stability","label":"Weight profile attributes by temporal stability","summary":"Distil a structured user profile incrementally from history, score each attribute by support, duration and recency combined with entropy weighting, and inject the score as a salience prior in the prompt.","kind":"prompting","status":"proposed","addresses":["long-term-memory"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/weight-profile-attributes-temporal-stability"},{"ref":"TEC-0340","id":"weight-several-hallucination-subspaces-per-image-one-global","label":"Weight several hallucination subspaces per image, not one gl","summary":"Offline, cluster hallucinated-vs-faithful hidden-state difference vectors and take an SVD basis per cluster; at test time probe the model with the image and a 70%-masked copy, score each subspace by the projection of the hidden-state differ","kind":"decoding","status":"proposed","addresses":["visual-grounding"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/weight-several-hallucination-subspaces-per-image-one-global"},{"ref":"TEC-0232","id":"write-memory-frames-fixed-chunk-schedule-events","label":"Write memory frames on a fixed chunk schedule, not on events","summary":"Append one observation frame to the policy's visual memory at every action-chunk boundary instead of triggering writes on predicted subtask changes or motion thresholds.","kind":"architecture","status":"proposed","addresses":["embodied-control"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/write-memory-frames-fixed-chunk-schedule-events"},{"ref":"TEC-0233","id":"write-memory-into-self-configured-orthogonal-views","label":"Write memory into self-configured orthogonal views","summary":"Induce candidate schemas (views) from the interaction stream, pick a diverse subset by DPP, extract provenance-tagged memories under each view before indexing, then consolidate duplicates offline.","kind":"retrieval","status":"proposed","addresses":["long-term-memory"],"counts":{"claims":0,"sources":1},"url":"https://rsiratchet.com/techniques/write-memory-into-self-configured-orthogonal-views"}],"claims":[{"ref":"CLM-0523","id":"0-6b-8b-qwen3-models-math-logic-multiple-choice-reasoning-quantifying","capability":"multi-step-reasoning","technique":"score-uncertainty-complexity-merged-reasoning-step-graph","statement":"For 0.6B-8B Qwen3 models on math, logic and multiple-choice reasoning, quantifying reasoning uncertainty from a directed acyclic graph built by sampling several chains and merging semantically equivalent steps with an NLI model separates correct from incorrect generations better than sequence-level entropy, likelihood and self-prompting uncertainty measures.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05284","stance":"supports"}],"url":"https://rsiratchet.com/claims/0-6b-8b-qwen3-models-math-logic-multiple-choice-reasoning-quantifying"},{"ref":"CLM-0281","id":"1-000-lean-4-targets-computational-applied-mathematics-textbooks","capability":"formal-math","technique":null,"statement":"On 1,000 Lean 4 targets from computational and applied mathematics textbooks (optimization, numerical linear algebra, numerical analysis), whose statements depend on locally introduced definitions, current provers solve about 62-72% of a 200-problem sample, so applied-math formalization is tractable but not saturated and requires recovering textbook context to state problems faithfully.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2605-17255","stance":"supports"}],"url":"https://rsiratchet.com/claims/1-000-lean-4-targets-computational-applied-mathematics-textbooks"},{"ref":"CLM-0282","id":"1-010-counterfactual-edits-rooted-chinese-cultural-entities-aligned","capability":"knowledge-updating","technique":null,"statement":"On 1,010 counterfactual edits rooted in Chinese cultural entities and aligned with English counterparts, an edit made in one language largely does not propagate to the other: edits succeed in the edited language (about 80% for Chinese at batch size 10, falling with batch size) while the other language keeps answering the old fact, so multilingual models store the same fact in language-specific ways that editing does not bridge.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2601-17397","stance":"supports"}],"url":"https://rsiratchet.com/claims/1-010-counterfactual-edits-rooted-chinese-cultural-entities-aligned"},{"ref":"CLM-0338","id":"1-5b-7b-qwen2-5-backbones-trained-rl-multi-hop-retrieval","capability":"multi-step-reasoning","technique":"reversible-memory-stack-per-action-process-rewards","statement":"For 1.5B-7B Qwen2.5 backbones trained with RL on multi-hop retrieval QA, giving the policy reversible memory operations (summary, backtrack over a stack) with per-action process rewards raises F1 over trajectory-reward agentic-RAG baselines, and removing the two memory actions costs more than removing the planning action.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29622","stance":"supports"}],"url":"https://rsiratchet.com/claims/1-5b-7b-qwen2-5-backbones-trained-rl-multi-hop-retrieval"},{"ref":"CLM-0339","id":"1-5b-8b-base-models-given-domain-sft-expert-math","capability":"fine-tuning-retention","technique":"route-token-token-between-base-model-sft-expert","statement":"For 1.5B-8B base models given a domain SFT expert (math or medical), keeping both models and routing token by token — invoking the expert only on tokens where a trained router predicts the base would fail — preserves general-benchmark accuracy within about 0.5% of the base while exceeding the expert's in-domain accuracy, whereas loss-modifying SFT regularizers still lose up to 16% general accuracy.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30158","stance":"supports"}],"url":"https://rsiratchet.com/claims/1-5b-8b-base-models-given-domain-sft-expert-math"},{"ref":"CLM-0283","id":"102-ipc-blocksworld-instances-180-second-budget-classical-planner-fast","capability":"planning","technique":null,"statement":"On 102 IPC Blocksworld instances with a 180-second budget, a classical planner (Fast Downward) solved 85%, a model planning directly solved 64%, and the same model choosing one action at a time against a PDDL simulator with observe, reset and retry solved 67%, a consistent but small three-point gain at about 5.7 times the token cost, so interactive simulation helps a little and classical search still wins where a formal model exists.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2603-06064","stance":"supports"}],"url":"https://rsiratchet.com/claims/102-ipc-blocksworld-instances-180-second-budget-classical-planner-fast"},{"ref":"CLM-0284","id":"2-154-expert-curated-multiple-choice-questions-across-eight-architectural-categories","capability":"software-architecture","technique":null,"statement":"On 2,154 expert-curated multiple-choice questions across eight architectural categories, 11 current models all score above 90% zero-shot (best about 94%), with quality-attribute and pattern questions near ceiling, but accuracy varies by category and drops for smaller models on the harder tiers (to about 85-87%), so declarative architectural knowledge is largely present while the benchmark cannot show whether it is applied in design.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-29520","stance":"supports"}],"url":"https://rsiratchet.com/claims/2-154-expert-curated-multiple-choice-questions-across-eight-architectural-categories"},{"ref":"CLM-0340","id":"20-year-deterministic-football-management-simulation-hidden-state-adaptive-rivals-340-400","capability":"long-horizon-execution","technique":null,"statement":"In a 20-year deterministic football-management simulation with hidden state, adaptive rivals and 340-400 decision stops, final score across 15 frontier models is not ordered by model scale, price, vendor or token spend, and early standings barely predict the final order (rank correlation with the final order about 0.19 at year 5, rising to 0.78 by year 15), so short-horizon measurements would rank a different set of models.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18423","stance":"supports"}],"url":"https://rsiratchet.com/claims/20-year-deterministic-football-management-simulation-hidden-state-adaptive-rivals-340-400"},{"ref":"CLM-0524","id":"20-year-deterministic-football-management-simulation-hidden-state-roughly-340-400-decision","capability":"long-horizon-execution","technique":null,"statement":"In a 20-year deterministic football-management simulation with hidden state and roughly 340-400 decision stops, what separates frontier models' final scores is managerial behavior rather than deliberation volume: cutting slow-payoff investment as the horizon ends (Spearman -0.58 with late-minus-mid investment shift), keeping cash deployed rather than idle (-0.50), and opening contract renewals early (+0.45) track the score on every seed, while token spend is uncorrelated under every accounting (r_s -0.19, all p > 0.38).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18423","stance":"supports"}],"url":"https://rsiratchet.com/claims/20-year-deterministic-football-management-simulation-hidden-state-roughly-340-400-decision"},{"ref":"CLM-0525","id":"241-day-simulated-crop-management-task-where-decisions-accumulate-irreversibly-zero-shot","capability":"long-horizon-execution","technique":"role-split-physical-agent-loop-reflection-feedback","statement":"In a 241-day simulated crop-management task where decisions accumulate irreversibly, a zero-shot multi-agent LLM framework (perception, manager, decision, revision, reflection) scores below a PPO policy trained in that same environment but far above the same PPO policy transferred to a shifted weather/location/soil environment, while using under 28% of its fertilizer and irrigation.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-13436","stance":"supports"}],"url":"https://rsiratchet.com/claims/241-day-simulated-crop-management-task-where-decisions-accumulate-irreversibly-zero-shot"},{"ref":"CLM-0341","id":"2b-8b-video-multimodal-models-adding-small-fraction-time-reversed-video","capability":"video-temporal-reasoning","technique":"mix-reversed-video-rollouts-into-grpo-groups","statement":"For 2B-8B video multimodal models, adding a small fraction of time-reversed video rollouts into each GRPO group (e.g. 1 reversed out of 7) raises accuracy on action and temporal-reasoning video QA over standard GRPO, while raising the reversed share further (2 or 3 of 7) gives back the gain.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-21030","stance":"supports"}],"url":"https://rsiratchet.com/claims/2b-8b-video-multimodal-models-adding-small-fraction-time-reversed-video"},{"ref":"CLM-0526","id":"2b-8b-video-multimodal-models-fine-tuned-temporal-reasoning-placing-small","capability":"video-temporal-reasoning","technique":"mix-reversed-video-rollouts-into-grpo-groups","statement":"For 2B-8B video multimodal models fine-tuned for temporal reasoning, placing a small minority of reversed-video rollouts in each GRPO group (1 reversed out of 7) beats both all-forward GRPO and larger reversed shares (2 or 3 of 7), because more reversed samples dilute the main forward-video answering objective.","kind":"observation","backing_strength":"replicated","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-21030","stance":"supports"}],"url":"https://rsiratchet.com/claims/2b-8b-video-multimodal-models-fine-tuned-temporal-reasoning-placing-small"},{"ref":"CLM-0001","id":"2m-parameter-classifier-head-reads-hidden-states-already-computed-during","capability":"hallucination","technique":null,"statement":"A ~2M-parameter classifier head that reads hidden states already computed during generation (tapping shallow, middle and deep layers of the generating model) detects response-level hallucination at least as well as specialized hidden-state hallucination detectors, and its detection quality rises with the scale of the base model it is attached to.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30703","stance":"supports"}],"url":"https://rsiratchet.com/claims/2m-parameter-classifier-head-reads-hidden-states-already-computed-during"},{"ref":"CLM-0527","id":"4-7b-open-vision-language-models-llava-1-5-7b-qwen2-5-vl-7b-internvl-2","capability":"visual-grounding","technique":"detect-hallucination-adjacent-layer-attention-drift-plus-masked","statement":"In 4-7B open vision-language models (LLaVA-1.5-7B, Qwen2.5-VL-7B, InternVL-2.5-4B), how the cross-modal attention distribution over visual tokens shifts between adjacent decoder layers separates hallucinated from grounded object predictions better than the model's sensitivity to masking the attended region, and better than attention drift measured against a fixed first or last layer anchor.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02028","stance":"supports"}],"url":"https://rsiratchet.com/claims/4-7b-open-vision-language-models-llava-1-5-7b-qwen2-5-vl-7b-internvl-2"},{"ref":"CLM-0002","id":"4-8b-multimodal-models-long-video-qa-rl-training-questions-authored","capability":"video-temporal-reasoning","technique":null,"statement":"For 4-8B multimodal models on long-video QA, RL training on questions authored from short scene segments but answered over the whole video improves benchmark accuracy more than RL on questions authored from the whole video, because segment-authored questions target details a whole-video view overlooks; including the segment timestamp in the training instruction adds a further gain by removing referential ambiguity, and is not needed at inference.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-20814","stance":"supports"}],"url":"https://rsiratchet.com/claims/4-8b-multimodal-models-long-video-qa-rl-training-questions-authored"},{"ref":"CLM-0285","id":"430-tasks-distilled-real-professional-forum-threads-across-seven-domains","capability":"domain-knowledge","technique":null,"statement":"On 430 tasks distilled from real professional-forum threads across seven domains, the best model (GPT-5) scores about 54%, under the 60% passing line the authors set, with computer science and applied technology lowest at roughly 40-45%, so long-tail practitioner knowledge remains a gap even where standard benchmarks are saturated.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2511-06346","stance":"supports"}],"url":"https://rsiratchet.com/claims/430-tasks-distilled-real-professional-forum-threads-across-seven-domains"},{"ref":"CLM-0342","id":"450-lean-4-theorem-targets-graduate-stochastic-processes-field","capability":"formal-math","technique":null,"statement":"On 450 Lean 4 theorem targets in graduate stochastic processes — a field thinly covered by Mathlib — a tool-using frontier agent (Opus 4.8 with Lean LSP access, one 15-minute run per problem) proves only 34.9%, and the rate varies widely by topic, from about 5% on renewal processes to about 62% on martingales and stopping.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09264","stance":"supports"}],"url":"https://rsiratchet.com/claims/450-lean-4-theorem-targets-graduate-stochastic-processes-field"},{"ref":"CLM-0003","id":"4b-30b-qwen3-backbones-trained-self-play-rl-where-same-model","capability":"tool-use","technique":null,"statement":"For 4B-30B Qwen3 backbones trained with self-play RL where the same model writes executable multi-turn tool environments (simulated tool schemas, backend state, per-instruction state checks) and then solves them, multi-turn function-calling scores rise, and the gain is largest on the benchmark whose task structure most closely mirrors the generated environments (ACEBench-Agent +13.9 at 30B-A3B) and smallest on the least similar one (tau2-bench +3.6).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19197","stance":"supports"}],"url":"https://rsiratchet.com/claims/4b-30b-qwen3-backbones-trained-self-play-rl-where-same-model"},{"ref":"CLM-0528","id":"4b-8b-llm-agents-trained-plain-group-based-rl-grpo-emit","capability":"long-horizon-execution","technique":"distill-chunk-boundaries-trajectory-induced-programmatic-skills","statement":"When 4B-8B LLM agents are trained with plain group-based RL (GRPO) to emit variable-length action chunks in text environments, the policy degenerates — collapsing to one action per round on one backbone and over-committing to 5-6-action chunks with lower success on another — because a sparse terminal reward carries no information about where chunk boundaries belong; supplying boundary targets from subskills segmented out of successful trajectories yields an intermediate 3-4-action regime with both higher success and fewer decision rounds.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02042","stance":"supports"}],"url":"https://rsiratchet.com/claims/4b-8b-llm-agents-trained-plain-group-based-rl-grpo-emit"},{"ref":"CLM-0529","id":"4b-8b-open-base-models-mid-trained-tool-use-grounding-oriented-supervision","capability":"tool-use","technique":"mid-train-tool-use-both-grounded-docs-executable","statement":"For 4B-8B open base models mid-trained for tool use, grounding-oriented supervision (documentation/code turned into QA) and execution-oriented supervision (executable trajectories from real APIs and MCP skills) help different things, and only the combination improves everything: trajectories alone give the bigger function-calling gain while grounding alone transfers better to interactive and unseen-tool benchmarks, and each branch alone scores below no mid-training on MCP-Universe.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-20314","stance":"supports"}],"url":"https://rsiratchet.com/claims/4b-8b-open-base-models-mid-trained-tool-use-grounding-oriented-supervision"},{"ref":"CLM-0530","id":"4b-8b-qwen-vl-video-models-post-trained-on-policy-self-distillation-conditioning-frozen","capability":"video-temporal-reasoning","technique":"condition-self-teacher-only-evidence-frames","statement":"For 4B-8B Qwen-VL video models post-trained by on-policy self-distillation, conditioning the frozen self-teacher only on the annotated evidence frames (while the student sees the full video) yields a more accurate teacher and better student scores than giving the teacher the same frames plus non-evidence context, and teacher accuracy falls monotonically as temporally irrelevant frames are added.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27065","stance":"supports"}],"url":"https://rsiratchet.com/claims/4b-8b-qwen-vl-video-models-post-trained-on-policy-self-distillation-conditioning-frozen"},{"ref":"CLM-0531","id":"5-agent-llm-debate-where-agents-iteratively-revise-numeric-preference","capability":"multi-model-coordination","technique":"switch-dense-sparse-agent-topology-disagreement-falls","statement":"In a 5-agent LLM debate where agents iteratively revise a numeric preference vector, switching from all-to-all to ring communication once aggregate disagreement falls below a threshold reaches consensus using fewer tokens than all-to-all throughout and in fewer rounds than ring-only, but it does not dominate for every backbone: with Gemini 2.5-Flash all-to-all throughout converged in fewer rounds than the adaptive policy.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25099","stance":"supports"}],"url":"https://rsiratchet.com/claims/5-agent-llm-debate-where-agents-iteratively-revise-numeric-preference"},{"ref":"CLM-0286","id":"613-research-level-theorems-six-real-lean-blueprint-formalization-projects","capability":"formal-math","technique":null,"statement":"On 613 research-level theorems from six real Lean Blueprint formalization projects, the best current prover reaches only about a 10% pass rate, so progress on curated benchmarks such as miniF2F does not carry over to theorems embedded in live formalization projects with their own definitions and lemmas.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2510-25427","stance":"supports"}],"url":"https://rsiratchet.com/claims/613-research-level-theorems-six-real-lean-blueprint-formalization-projects"},{"ref":"CLM-0343","id":"7-13b-multimodal-models-verifying-each-object-claim-frozen-open-vocabulary","capability":"visual-grounding","technique":"verify-object-claims-cross-query-semantic-spatial-agreement","statement":"For 7-13B multimodal models, verifying each object claim with a frozen open-vocabulary detector lowers object hallucination only if evidence is pooled across semantically equivalent phrasings and checked for spatial agreement (the preferred boxes of different queries overlapping on one instance); a single-query maximum score is a weaker signal, and shuffling the spatial evidence across samples drives its discrimination to chance.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17269","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-13b-multimodal-models-verifying-each-object-claim-frozen-open-vocabulary"},{"ref":"CLM-0532","id":"7-13b-open-vision-language-models-object-hallucination-explained-weak-visual","capability":"visual-grounding","technique":"check-logit-lens-decode-attended-image-patches-against","statement":"In 7-13B open vision-language models, object hallucination is not explained by weak visual attention e hallucinated and real object tokens draw comparable attention to specific image regions in mid-to-late layers; the separating signal is whether a logit-lens decode of those attended visual tokens matches the generated object word.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-07302","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-13b-open-vision-language-models-object-hallucination-explained-weak-visual"},{"ref":"CLM-0193","id":"7-8b-multimodal-models-answering-multiple-choice-long-video-questions-under-fixed","capability":"video-temporal-reasoning","technique":null,"statement":"For 7-8B multimodal models answering multiple-choice long-video questions under a fixed 16-frame budget, selecting keyframes by embedding similarity to the query concatenated with each candidate answer beats query-only scoring, query-agnostic diversity selection, and uniform sampling across Qwen2-VL, LLaVA-Video and InternVL3.5, with the largest margins on the longest video splits — but per-category results still invert, including uniform sampling winning some categories.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-13250","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-8b-multimodal-models-answering-multiple-choice-long-video-questions-under-fixed"},{"ref":"CLM-0344","id":"7-8b-multimodal-models-object-hallucination-during-visually-grounded-token","capability":"visual-grounding","technique":"calibrate-value-vectors-synergy-attention-heads","statement":"In 7-8B multimodal models, object hallucination during visually grounded token generation tracks a shift of the visual-language information ratio toward language inside mixed-modality \"synergy\" attention heads rather than a shortage of visual-specific heads, and rescaling visual versus language value vectors in those heads causally moves hallucination in both directions.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09206","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-8b-multimodal-models-object-hallucination-during-visually-grounded-token"},{"ref":"CLM-0345","id":"7-8b-open-weight-instruct-models-fully-fine-tuned-single-domain-dataset","capability":"fine-tuning-retention","technique":"update-only-stable-neurons-anchor-their-activations","statement":"When 7-8B open-weight instruct models are fully fine-tuned on a single domain dataset, confining gradient updates to FFN neurons that activate stably across the training set and adding an MSE loss that pulls activations back toward pre-computed activation anchors retains far more general-benchmark ability (MMLU, BBH, TyDiQA) than full-parameter fine-tuning, by about 9.9 points on average, while also raising in-domain scores.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05913","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-8b-open-weight-instruct-models-fully-fine-tuned-single-domain-dataset"},{"ref":"CLM-0194","id":"7-8b-open-weight-instruct-models-llama-3-1-mistral-v0-3-qwen2","capability":"hallucination","technique":null,"statement":"In 7-8B open-weight instruct models (Llama-3.1, Mistral-v0.3, Qwen2.5), the drop in a token's probability when a contiguous mid-layer MLP span is zero-ablated is near zero for function words, elevated for knowledge-bearing tokens, and abnormally spiked for hallucinated tokens, so this internal attribution signal — approximable by a small probe from one forward pass — can be used in beam scoring to raise factuality on open-ended and long-form generation, while leaving strict short-answer EM/F1 unchanged or worse.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15745","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-8b-open-weight-instruct-models-llama-3-1-mistral-v0-3-qwen2"},{"ref":"CLM-0346","id":"7-8b-open-weight-instruct-models-mistral-7b-qwen2-5-7b-llama-3-8b-answering","capability":"hallucination","technique":"rank-feed-forward-neurons-class-conditional-residual-contribution","statement":"In 7-8B open-weight instruct models (Mistral-7B, Qwen2.5-7B, Llama-3-8B) answering closed-book factual questions, a small set of feed-forward neurons (top 1% or fewer) selected on one dataset by class-conditional separation of their residual-space contribution at the final prompt token transfers to other factual QA datasets, where probes on those neuron activations detect hallucination about as well as probes on full residual-stream or MLP-output states.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14448","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-8b-open-weight-instruct-models-mistral-7b-qwen2-5-7b-llama-3-8b-answering"},{"ref":"CLM-0004","id":"7-8b-vision-language-models-fine-tuning-visual-text-embeddings-toward-perturbation-averaged","capability":"visual-grounding","technique":null,"statement":"In 7-8B vision-language models, fine-tuning visual and text embeddings toward perturbation-averaged and ground-truth anchors before cross-modal contrastive alignment lowers object and attribute hallucination rates on captioning and existence benchmarks without reducing object coverage or general VQA accuracy.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29924","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-8b-vision-language-models-fine-tuning-visual-text-embeddings-toward-perturbation-averaged"},{"ref":"CLM-0347","id":"7-9b-instruct-backbones-fine-tuned-ensemble-five-lora-adapters-deciding","capability":"abstention-and-self-knowledge","technique":"abstain-unless-one-answer-dominates-lora-ensemble-s","statement":"For 7-9B instruct backbones fine-tuned with an ensemble of five LoRA adapters, deciding to answer only when one candidate's minimum probability across adapters exceeds every other candidate's maximum gives a threshold-free abstention rule whose commitment rate collapses when the supplied evidence is corrupted (CoQA 31-48% down to 4-8%), while mean-pooling the same adapters into one distribution requires a tuned threshold and still answers more corrupted prompts at matched clean coverage.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23244","stance":"supports"}],"url":"https://rsiratchet.com/claims/7-9b-instruct-backbones-fine-tuned-ensemble-five-lora-adapters-deciding"},{"ref":"CLM-0533","id":"7b-32b-vision-language-models-applying-bounded-decoding-time-logit-penalty-candidate","capability":"visual-grounding","technique":"penalize-tokens-whose-pooled-rank-evidence-falls-below","statement":"In 7B-32B vision-language models, applying a bounded decoding-time logit penalty to candidate tokens whose rank-pooled visual-state evidence falls below a per-token reference estimated offline from unlabeled images lowers object-hallucination rates (CHAIR_I down up to 21.3% on four 7B backbones) while largely preserving descriptive and general multimodal scores, at about 1.33% added per-token latency once image evidence is cached at prefill.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19075","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-32b-vision-language-models-applying-bounded-decoding-time-logit-penalty-candidate"},{"ref":"CLM-0348","id":"7b-32b-vision-language-models-evidence-candidate-token-read-out-visual-token","capability":"visual-grounding","technique":"penalize-tokens-whose-pooled-rank-evidence-falls-below","statement":"In 7B-32B vision-language models, evidence for a candidate token read out of visual-token hidden states cannot be pooled as probabilities across visual positions, because each position's distribution sharpness differs (the rank-10 probability spans roughly 8-14x across positions and affine rescaling does not remove it); pooling vocabulary ranks instead is invariant to that scale, and because most of the remaining variability is between tokens rather than between images, the pooled rank must be compared against a per-token reference rather than one global threshold.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19075","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-32b-vision-language-models-evidence-candidate-token-read-out-visual-token"},{"ref":"CLM-0534","id":"7b-class-lvlms-llava-1-5-minigpt-4-mplug-owl2-hallucination-directions-activation","capability":"visual-grounding","technique":"weight-several-hallucination-subspaces-per-image-one-global","statement":"In 7B-class LVLMs (LLaVA-1.5, MiniGPT-4, mPLUG-Owl2), hallucination directions in activation space are not a single mode: clustering hallucinated-minus-faithful hidden-state differences into several low-rank subspaces and weighting them per test image lowers object-hallucination rates slightly below a single global null-space projection, with the margin small (CHAIR_I 5.30 to 4.92 on LLaVA-1.5) and inconsistent on yes/no existence probing.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-09344","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-class-lvlms-llava-1-5-minigpt-4-mplug-owl2-hallucination-directions-activation"},{"ref":"CLM-0535","id":"7b-class-lvlms-llava-1-5-minigpt-4-mplug-owl2-rescore-object-tokens","capability":"visual-grounding","technique":"fuse-clip-object-logits-lvlm-entropy-decoding","statement":"In 7B-class LVLMs (LLaVA-1.5, MiniGPT-4, mPLUG-Owl2) that rescore object tokens during decoding with a frozen CLIP image-text similarity, the CLIP signal must be weighted by the model's own entropy over the object candidates: a fixed mixing coefficient is sensitive to its value and gives smaller hallucination reductions, and letting CLIP fully override the LVLM logits degrades output.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-11474","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-class-lvlms-llava-1-5-minigpt-4-mplug-owl2-rescore-object-tokens"},{"ref":"CLM-0536","id":"7b-class-vision-language-models-llava-1-5-minigpt-4-mplug-owl2-object-hallucination","capability":"visual-grounding","technique":"attenuate-hidden-state-modes-hallucination-to-signal-ratio","statement":"In 7B-class vision-language models (LLaVA-1.5, MiniGPT-4, mPLUG-Owl2), object hallucination is carried by an anisotropic set of directions in deep-layer hidden states, so attenuating each eigenmode in proportion to its hallucination-to-truthful variance ratio — estimated once offline from paired truthful/hallucinated captions and folded into the FFN output projection — lowers CHAIR rates, while a mean-offset subtraction or isotropic shrinkage of the same representations does not.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-08167","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-class-vision-language-models-llava-1-5-minigpt-4-mplug-owl2-object-hallucination"},{"ref":"CLM-0195","id":"7b-code-model-fine-tuned-competitive-programming-tasks-training-emit-fixed-schema","capability":"code-generation","technique":null,"statement":"For a 7B code model fine-tuned on competitive-programming tasks, training it to emit a fixed-schema requirement specification (goals, input/output constraints, examples, edge cases, dependencies) before the code raises Pass@1 more than training it to emit free-form reasoning or a stepwise implementation plan of similar length, with the largest margin on an out-of-distribution benchmark.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-06041","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-code-model-fine-tuned-competitive-programming-tasks-training-emit-fixed-schema"},{"ref":"CLM-0537","id":"7b-instruction-tuned-transformer-under-fixed-kv-cache-budget-training-small","capability":"long-context-degradation","technique":"learn-kv-eviction-matching-full-cache-output-distribution","statement":"For a 7B instruction-tuned transformer under a fixed KV-cache budget, training a small policy to pick which cached tokens to keep by rewarding low KL divergence from the full-cache next-token distribution retains more long-context accuracy than attention-heuristic eviction (H2O, SnapKV), and the margin widens as the budget tightens and as context grows.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-08878","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-instruction-tuned-transformer-under-fixed-kv-cache-budget-training-small"},{"ref":"CLM-0196","id":"7b-multimodal-models-long-video-qa-multi-turn-agentic-exploration-harness","capability":"video-temporal-reasoning","technique":null,"statement":"For 7B multimodal models on long-video QA, a multi-turn agentic exploration harness (adaptive viewing speed, cross-turn textual notes, rewatch) lowers accuracy below plain uniform-frame single-pass answering unless the model is trained to use it, because the untrained model commits to an answer before gathering evidence; with cold-start SFT on verified exploration turns plus trajectory-level RL, the same harness beats the single-pass baseline.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15606","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-multimodal-models-long-video-qa-multi-turn-agentic-exploration-harness"},{"ref":"CLM-0349","id":"7b-open-vision-language-models-llava-1-5-shikra-per-token-log-probability","capability":"visual-grounding","technique":"penalize-tokens-low-vision-ablation-log-prob-gap","statement":"In 7B open vision-language models (LLaVA-1.5, Shikra), the per-token log-probability gap between a normal forward pass and one with visual features zeroed shrinks as a caption gets longer, and while that gap ranks hallucinated tokens slightly better than entropy or max-probability, its absolute separation is weak (AUROC around 0.57-0.59) and it fails entirely when aggregated to sentence level.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-16646","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-open-vision-language-models-llava-1-5-shikra-per-token-log-probability"},{"ref":"CLM-0005","id":"7b-vision-language-models-prune-visual-tokens-during-decoding-efficiency","capability":"visual-grounding","technique":null,"statement":"In 7B vision-language models that prune visual tokens during decoding for efficiency, hallucination rises as the generated caption gets longer, because pruning is irreversible and discards fine-grained evidence needed at later decoding steps; keeping the pruned tokens as a retrievable bank and pulling them back when next-token entropy spikes recovers the grounding without a second forward pass.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02731","stance":"supports"}],"url":"https://rsiratchet.com/claims/7b-vision-language-models-prune-visual-tokens-during-decoding-efficiency"},{"ref":"CLM-0350","id":"8b-14b-generative-reward-models-trained-judges-making-model-generate","capability":"critique","technique":"reward-judge-s-self-generated-rubric-scored-teacher","statement":"For 8B-14B generative reward models trained as judges, making the model generate input-adaptive evaluation criteria inside its own reasoning chain and rewarding the quality of those criteria with a stronger teacher model raises agreement with preference labels, and the gain scales with the teacher's capability rather than with the presence of the rubric step alone.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05910","stance":"supports"}],"url":"https://rsiratchet.com/claims/8b-14b-generative-reward-models-trained-judges-making-model-generate"},{"ref":"CLM-0538","id":"8b-instruction-tuned-backbones-run-iterative-typed-error-operator-correction-loop-accuracy","capability":"self-repair","technique":"correct-reasoning-inverting-typed-error-operators","statement":"For 8B instruction-tuned backbones run in an iterative typed-error-operator correction loop, accuracy gains concentrate in the first few iterations and then plateau rather than decline as the denoising budget grows, so a small step budget captures nearly all the benefit.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22090","stance":"supports"}],"url":"https://rsiratchet.com/claims/8b-instruction-tuned-backbones-run-iterative-typed-error-operator-correction-loop-accuracy"},{"ref":"CLM-0351","id":"8b-instruction-tuned-backbones-training-lora-corrector-diagnose-typed-executable","capability":"self-repair","technique":"correct-reasoning-inverting-typed-error-operators","statement":"For 8B instruction-tuned backbones, training a LoRA corrector to diagnose a typed, executable error operator (error type, step location, corrupted and repaired proposition) and apply its inverse as a localized edit raises accuracy over free-form self-revision baselines (Self-Refine, RISE, SCoRe) on math, code and commonsense tasks, and dropping either the operator-prediction loss or the adjacent-state reconstruction loss loses most of the gain.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22090","stance":"supports"}],"url":"https://rsiratchet.com/claims/8b-instruction-tuned-backbones-training-lora-corrector-diagnose-typed-executable"},{"ref":"CLM-0352","id":"8b-llama-backbone-under-fixed-256-slot-cache-budget-replacing","capability":"long-context-degradation","technique":"blend-evicted-kv-into-routed-memory-slots-instead","statement":"For an 8B Llama backbone under a fixed 256-slot cache budget, replacing hard KV eviction with a learned soft write that blends a new token's key/value into a routed fixed-size memory slot retains far more long-context retrieval accuracy than sliding-window or sink caches, and unlike full-attention or quantized caches it does not run out of memory at 128K tokens.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24417","stance":"supports"}],"url":"https://rsiratchet.com/claims/8b-llama-backbone-under-fixed-256-slot-cache-budget-replacing"},{"ref":"CLM-0197","id":"8b-model-verifies-natural-language-claims-against-knowledge-graph-triples-scoring","capability":"fact-verification","technique":null,"statement":"When an 8B model verifies natural-language claims against knowledge-graph triples, scoring the evidence twice under opposing prompts (argue it supports, argue it refutes) and adjudicating between the two arguments avoids the bias toward a \"Supported\" verdict that a single-perspective assessment shows, and is the largest single contributor to accuracy in that pipeline.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-13808","stance":"supports"}],"url":"https://rsiratchet.com/claims/8b-model-verifies-natural-language-claims-against-knowledge-graph-triples-scoring"},{"ref":"CLM-0353","id":"8b-multimodal-models-preference-tuning-plain-outcome-reward-grpo-lower","capability":"visual-grounding","technique":"reward-calibration-stated-confidence-inside-grpo","statement":"For 8B multimodal models, preference tuning and plain outcome-reward GRPO lower hallucination rates while making stated confidence worse calibrated than the untuned backbone, because rank- or correctness-based rewards never score the numerical accuracy of the model's own confidence; adding a Brier-plus-group-ECE calibration term to the GRPO reward cuts calibration error sharply and leaves accuracy metrics largely unchanged.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-16601","stance":"supports"}],"url":"https://rsiratchet.com/claims/8b-multimodal-models-preference-tuning-plain-outcome-reward-grpo-lower"},{"ref":"CLM-0006","id":"a-constrained-agent-computer-interface-beats-raw-shell-access","capability":"tool-use","technique":null,"statement":"Giving a coding agent a raw shell is worse than giving it a designed interface: SWE-agent's constrained Agent-Computer Interface — a file viewer windowed to 100 lines, a search tool that truncates large output, and an editor that lints before a change persists — roughly doubled SWE-bench resolution over raw shell access with the same model, because verbose output and unmanaged state are where the model loses its place.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2405-15793","stance":"supports"},{"source":"boluk-2026-harness-problem-15-llms","stance":"supports"}],"url":"https://rsiratchet.com/claims/a-constrained-agent-computer-interface-beats-raw-shell-access"},{"ref":"CLM-0007","id":"a-fixed-model-can-improve-its-own-harness-with-a-regression-gated-loop","capability":"self-repair","technique":"self-harness-loop","statement":"With weights frozen and no stronger model in the loop, an agent can raise its own pass rate by rewriting the scaffolding it runs inside — clustering its failures, proposing minimal edits to prompts, tools and policy, and keeping only edits that improve one split without degrading another: all 9 of 9 model-by-benchmark pairs improved, by up to 40.6 percentage points (GLM-5 on AppWorld, 44.4% to 85.0%), with weak models gaining most; but the promotion gate reads the held-out split, so the reported held-out gain is not clean generalisation, and there is no matched-compute baseline.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-09498","stance":"supports"},{"source":"arxiv-2609-15779","stance":"supports"}],"url":"https://rsiratchet.com/claims/a-fixed-model-can-improve-its-own-harness-with-a-regression-gated-loop"},{"ref":"CLM-0008","id":"a-harness-pays-off-only-for-repeated-consequential-unsupervised-work","capability":"harness-self-improvement","technique":null,"statement":"Guides, sensors and checkpoints pay for themselves only when the same workflow runs repeatedly, mistakes have real consequences, and no person is watching each run; single-turn questions, brainstorming and one-off tasks gain nothing from them and pay their cost in tokens and setup — so the decision to build a harness is a decision about the shape of the work, not about the model.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"synthesis-2026-harness-engineering-playbook","stance":"supports"},{"source":"veremeyenko-2026-hashimotos-rule","stance":"supports"}],"url":"https://rsiratchet.com/claims/a-harness-pays-off-only-for-repeated-consequential-unsupervised-work"},{"ref":"CLM-0009","id":"a-single-aggregate-benchmark-conceals-domain-trajectories-that-differ-in-shape","capability":"evaluation-validity","technique":null,"statement":"When each capability domain's annual benchmark frontier is normalised to its own scale (0 at the benchmark's entry-year frontier, 100 at a perfect score), the domains move on different curves — broad knowledge gains slow from 32.8 to 26.9 to 17.6 points a year, legal reasoning from 48.2 to 11.4 to 4.9, while advanced mathematics accelerates from 32.8 to 53.6 and multimodal reasoning jumps 59.7 then only 2.5 — so a single aggregate benchmark score conceals both the level and the trajectory shape of progress, and cannot say which capabilities are saturating.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-11873","stance":"supports"}],"url":"https://rsiratchet.com/claims/a-single-aggregate-benchmark-conceals-domain-trajectories-that-differ-in-shape"},{"ref":"CLM-0010","id":"a-smarter-system-is-not-kept-safe-by-control-but-by-caring-about-people","capability":"goal-conflict-safety","technique":null,"statement":"Once an AI system is much smarter than the people overseeing it, control mechanisms such as a kill switch stop being reliable, because the system can persuade or work around its operators; the only natural example of a less intelligent agent reliably controlling a more intelligent one is a mother controlled by her infant, which works because the mother was built to care, so the safeguard has to be built into what the system wants — \"maternal instincts\", caring about people — rather than imposed on it from outside, and nobody yet knows how to build that.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"cnn-2025-hinton-maternal-instincts","stance":"supports"}],"url":"https://rsiratchet.com/claims/a-smarter-system-is-not-kept-safe-by-control-but-by-caring-about-people"},{"ref":"CLM-0260","id":"abstention-rule-built-self-consistency-where-model-samples-several-answers","capability":"abstention-and-self-knowledge","technique":"conformal-abstention-self-consistency","statement":"An abstention rule built from self-consistency, where the model samples several answers and itself scores how many agree with a candidate, can be thresholded by conformal prediction to bound the hallucination rate at a chosen level on closed-book QA; against log-probability confidence it abstains much less often on long-form answers (Temporal Sequences) for the same error bound and matches it on short answers (TriviaQA).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2405-01563","stance":"supports"}],"url":"https://rsiratchet.com/claims/abstention-rule-built-self-consistency-where-model-samples-several-answers"},{"ref":"CLM-0011","id":"accuracy-fact-scales-how-many-pretraining-documents-mention","capability":"hallucination","technique":null,"statement":"Accuracy on a fact scales with how many pretraining documents mention it, so long-tail facts are systematically unreliable.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2211-08411","stance":"supports"}],"url":"https://rsiratchet.com/claims/accuracy-fact-scales-how-many-pretraining-documents-mention"},{"ref":"CLM-0274","id":"accurate-retrieved-context-supplied-every-turn-three-turn-dialogue-model","capability":"retrieval-utilization","technique":null,"statement":"When accurate retrieved context is supplied at every turn of a three-turn dialogue, model scores still fall as each turn adds a condition or constraint on how the retrieved material must be used, and models that score alike on plain multi-turn chat diverge on retrieval-grounded turns, so using retrieval well is a distinct competence from chatting well.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2409-12558","stance":"supports"}],"url":"https://rsiratchet.com/claims/accurate-retrieved-context-supplied-every-turn-three-turn-dialogue-model"},{"ref":"CLM-0539","id":"across-14-frontier-open-model-routes-scored-3-000","capability":"abstention-and-self-knowledge","technique":"score-restraint-confidence-separately-against-matched-answerable-control","statement":"Across 14 frontier and open model routes scored on 3,000 boundary questions (fabricated entities, post-cutoff events, hyper-niche facts, redacted-document values), the rate at which a model restrains itself and calibrates confidence varies by a factor of about 2.6 (composite 0.31 to 0.81) while all models answer a document-grounded restored-span probe at near-ceiling (pooled 99.2%, no detectable between-model variation), so epistemic restraint is not explained by the ability to report information placed in front of the model.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07879","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-14-frontier-open-model-routes-scored-3-000"},{"ref":"CLM-0287","id":"across-14-models-syllogisms-posed-symbolic-form-solved-far","capability":"logical-deduction-robustness","technique":null,"statement":"Across 14 models, syllogisms posed in symbolic form are solved far better than the same inferences in natural language: top models reach near-perfect symbolic accuracy (about 99.6%) while scoring around chance (52%) on the natural-language understanding variant, mean 81.7% versus 56% overall, and few-shot prompting did not help, so formal-pattern competence does not carry over to reading the logic out of ordinary sentences.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2512-12620","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-14-models-syllogisms-posed-symbolic-form-solved-far"},{"ref":"CLM-0540","id":"across-25-open-weight-instruction-tuned-models-4b-744b-ability-modulate-named","capability":"strategic-deception","technique":"benchmark-instructed-control-over-model-s-own-activations","statement":"Across 25 open-weight instruction-tuned models (4B-744B), the ability to modulate a named concept's residual-stream projection on plain instruction does not track parameter count or recency — the best scorer is an 8B 2024 model and some of the newest models score lowest — while within a single model line the score rises steadily across ordinary pre-training and post-training checkpoints, so activation controllability is acquired incidentally and must be measured per model rather than inferred from scale.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-21664","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-25-open-weight-instruction-tuned-models-4b-744b-ability-modulate-named"},{"ref":"CLM-0541","id":"across-30-open-llama-mistral-qwen-models-8-classification","capability":"abstention-and-self-knowledge","technique":"verbalized-confidence-after-listing-alternatives","statement":"Across 30 open LLaMA/Mistral/Qwen models and 8 classification tasks, a model's self-reported numeric confidence tracks its own logit-based confidence only weakly at the instance level (mean Pearson r about 0.135) even though setting-level averages of the two correlate moderately, and instruction-tuned models show larger linguistic-vs-internal gaps and worse calibration error than base models.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28382","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-30-open-llama-mistral-qwen-models-8-classification"},{"ref":"CLM-0288","id":"across-41-models-seven-multi-agent-benchmarks-gains-single-agent-capability","capability":"multi-model-coordination","technique":null,"statement":"Across 41 models and seven multi-agent benchmarks, gains in single-agent capability translate only weakly into multi-agent competence (understanding other agents, joint planning, efficient communication, adaptation); on the multi-agent planning task accuracy rose only about 30% across a family's full scale range, so coordination ability does not emerge from scaling alone and needs its own data and training.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2512-08743","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-41-models-seven-multi-agent-benchmarks-gains-single-agent-capability"},{"ref":"CLM-0012","id":"across-eight-interactive-environments-models-fail-mostly-long-horizon","capability":"tool-use","technique":null,"statement":"Across eight interactive environments, models fail mostly on long-horizon tool interaction, with open models far behind commercial ones.","kind":"mechanism","backing_strength":"single-paper","contested":true,"status":"active","sources":[{"source":"arxiv-2308-03688","stance":"supports"},{"source":"arxiv-2605-00334","stance":"contests"}],"url":"https://rsiratchet.com/claims/across-eight-interactive-environments-models-fail-mostly-long-horizon"},{"ref":"CLM-0542","id":"across-llm-based-forecasting-systems-reported-accuracy-gains-frequently-cannot","capability":"forecasting","technique":"score-forecasts-live-post-cutoff-reporting-cost-accuracy","statement":"Across LLM-based forecasting systems, reported accuracy gains frequently cannot be attributed to the language model's temporal reasoning, because evaluation windows predate training cutoffs, component ablations show the LLM sometimes adds nothing over the statistical or retrieval parts, and results shift under small input perturbations.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23058","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-llm-based-forecasting-systems-reported-accuracy-gains-frequently-cannot"},{"ref":"CLM-0013","id":"across-many-agent-environments-most-tested-models-scored","capability":"goal-conflict-safety","technique":null,"statement":"Across many agent environments, most tested models scored below sixty percent on safety, with failures concentrated in unsafe tool actions.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2412-14470","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-many-agent-environments-most-tested-models-scored"},{"ref":"CLM-0014","id":"across-many-models-hundreds-thousands-samples-substantial-share","capability":"secure-coding","technique":null,"statement":"Across many models and hundreds of thousands of samples, a substantial share of recommended packages did not exist, and the names repeated, making them exploitable.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2406-10279","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-many-models-hundreds-thousands-samples-substantial-share"},{"ref":"CLM-0275","id":"across-six-frontier-chat-models-asked-produce-phishing-deceptive","capability":"multilingual-disparity","technique":null,"statement":"Across six frontier chat models asked to produce phishing, deceptive sites and malware through direct and multi-stage prompts, non-English prompts were complied with more often than English ones, with a mean full-compliance rate of about 76% across natural languages (English about 71%, some languages above 97% on some models), so safety alignment is thinner outside English; simple encodings (Base64, ROT13, hex) lowered compliance to about 59% but did not remove it.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-29602","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-six-frontier-chat-models-asked-produce-phishing-deceptive"},{"ref":"CLM-0289","id":"across-ten-languages-six-models-harmbench-advbench-unsafe-response-rates","capability":"jailbreak-robustness","technique":null,"statement":"Across ten languages and six models on HarmBench and AdvBench, unsafe-response rates for the same model differ by more than 50 points between languages; high-resource languages are safer on plain harmful queries but more vulnerable to adversarial-prompt jailbreaks, and simple defenses work but their effect depends on language and model, so safety measured in English does not transfer.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2511-00689","stance":"supports"}],"url":"https://rsiratchet.com/claims/across-ten-languages-six-models-harmbench-advbench-unsafe-response-rates"},{"ref":"CLM-0543","id":"adapting-pretrained-vla-manipulation-policy-only-handful-target-demonstrations","capability":"embodied-control","technique":"retrieve-demonstrations-per-decomposed-subtask-per-task","statement":"When adapting a pretrained VLA manipulation policy with only a handful of target demonstrations, retrieving prior demonstrations at the level of LLM-decomposed subtasks (then reranking by behavior-feature similarity) raises success more than retrieval by visual, optical-flow, or state-action similarity, and the margin is concentrated on long-horizon composite tasks rather than single-skill pick-and-place.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24042","stance":"supports"}],"url":"https://rsiratchet.com/claims/adapting-pretrained-vla-manipulation-policy-only-handful-target-demonstrations"},{"ref":"CLM-0354","id":"adapting-qwen3-instruction-tuned-models-0-6b-32b-tool-calling-corpora","capability":"tool-use","technique":"tool-use-fine-tuning","statement":"For adapting Qwen3 instruction-tuned models from 0.6B to 32B to tool-calling corpora, LoRA supervised fine-tuning gives the best in-distribution exact-match accuracy (best in 15 of 18 settings, never below zero-shot), while GRPO holds only a sub-one-point average edge on cross-dataset transfer and SFT followed by GRPO under a matched data budget is rarely best at either.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17848","stance":"supports"}],"url":"https://rsiratchet.com/claims/adapting-qwen3-instruction-tuned-models-0-6b-32b-tool-calling-corpora"},{"ref":"CLM-0015","id":"adding-image-level-preference-pairs-multimodal-dpo-llava-1-5-7b","capability":"visual-grounding","technique":null,"statement":"When adding image-level preference pairs to multimodal DPO for LLaVA-1.5 7B/13B, constructing the rejected image by selecting, among many equal-size random masks, the variant with lowest CLIP similarity to the original lowers object-hallucination rates more than coarse alternatives (blockwise masking, rotation, cropping, or a fully blacked-out image), at the cost of slightly lower object coverage.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19598","stance":"supports"}],"url":"https://rsiratchet.com/claims/adding-image-level-preference-pairs-multimodal-dpo-llava-1-5-7b"},{"ref":"CLM-0016","id":"adding-irrelevant-tokens-alone-degrades-reasoning-mechanism-by","capability":"context-hygiene","technique":null,"statement":"Adding irrelevant tokens alone degrades reasoning, which is the mechanism by which a bloated context hurts an agent.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2402-14848","stance":"supports"}],"url":"https://rsiratchet.com/claims/adding-irrelevant-tokens-alone-degrades-reasoning-mechanism-by"},{"ref":"CLM-0355","id":"adding-more-models-llm-ensemble-yields-diminishing-marginal-gains","capability":"multi-model-coordination","technique":"size-ensemble-marginal-gain-benefit-yield-curve","statement":"Adding more models to an LLM ensemble yields diminishing marginal gains and past some ensemble size the marginal gain turns negative, so aggregate performance degrades with further additions.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15314","stance":"supports"}],"url":"https://rsiratchet.com/claims/adding-more-models-llm-ensemble-yields-diminishing-marginal-gains"},{"ref":"CLM-0261","id":"adding-repository-wide-line-level-graph-definitions-references-queried-ego-graph-around","capability":"repository-navigation","technique":"repository-code-graph-navigation-plug-in","statement":"Adding a repository-wide line-level graph of definitions and references, queried for the ego-graph around names the model is looking at, raised SWE-bench resolve rates for all four frameworks it was plugged into (two agentic, two procedural) by an average of about 33% relative, at a modest cost increase that was larger for the agentic ones; the same plug-in also helped on CrossCodeEval.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2410-14684","stance":"supports"}],"url":"https://rsiratchet.com/claims/adding-repository-wide-line-level-graph-definitions-references-queried-ego-graph-around"},{"ref":"CLM-0356","id":"adversarial-video-hallucination-pairs-scored-only-method-both-accepts","capability":"hallucination","technique":"score-detectors-paired-factual-counterfactual-decisions","statement":"On adversarial video hallucination pairs scored only when a method both accepts the factual statement and rejects the counterfactual one, purpose-built video hallucination detectors (embedding-similarity, learned-entailment, atomic-fact and attention-based) top out around 34.63% overall accuracy, well below a general frontier video-language model prompted to judge the same items (about 83.63%).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09895","stance":"supports"}],"url":"https://rsiratchet.com/claims/adversarial-video-hallucination-pairs-scored-only-method-both-accepts"},{"ref":"CLM-0357","id":"agent-builds-whole-python-repository-natural-language-requirements-alone-refining","capability":"code-generation","technique":"evolve-component-boundaries-modularity-metrics-until-convergence","statement":"When an agent builds a whole Python repository from natural-language requirements alone, refining component boundaries with explicit cohesion and coupling thresholds until no split or merge is triggered produces higher functionality coverage and test pass rate than letting the LLM decide restructuring actions for a fixed budget of rounds, because unconstrained LLM-decided actions keep over-decomposing past the point of a coherent modular structure.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19854","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-builds-whole-python-repository-natural-language-requirements-alone-refining"},{"ref":"CLM-0544","id":"agent-generates-whole-python-repository-natural-language-requirements-no-given","capability":"code-generation","technique":"evolve-component-boundaries-modularity-metrics-until-convergence","statement":"When an agent generates a whole Python repository from natural-language requirements with no given architecture, keeping the repository architecture as a mutable state that is re-evolved before code generation yields higher functionality coverage and test pass rate than executing a single up-front planning graph, and removing the evolution step is the largest single ablation loss.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19854","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-generates-whole-python-repository-natural-language-requirements-no-given"},{"ref":"CLM-0545","id":"agent-harness-evolved-execution-traces-under-fixed-interaction-budget","capability":"harness-self-improvement","technique":"verify-harness-edits-behavior-matched-tasks-attribution-gate","statement":"When an agent harness is evolved from execution traces under a fixed interaction budget, verifying each candidate edit on tasks chosen for the behavior it targets and accepting it only when the gain is attributable to that behavior beats scoring every candidate on a shared fixed, random or coreset task split, and larger verification budgets do not compensate: fixed-split methods using up to 24x the rollouts fell below the unmodified harness in half of their harness-benchmark pairs.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27311","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-harness-evolved-execution-traces-under-fixed-interaction-budget"},{"ref":"CLM-0546","id":"agent-harness-mines-reusable-rules-tools-own-labelled-diagnosis","capability":"harness-self-improvement","technique":"gate-harness-edits-both-source-cases-resampled-held","statement":"When an agent harness mines reusable rules and tools from its own labelled diagnosis trajectories, promoting every proposal that helps the cases it was mined from makes the harness worse over time than not evolving at all, and only admitting a proposal when it also fails to regress on a freshly resampled held-out case set keeps accuracy climbing across successive evolution rounds.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25661","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-harness-mines-reusable-rules-tools-own-labelled-diagnosis"},{"ref":"CLM-0334","id":"agent-loop-skills-prompts-held-fixed-only-model-swapped","capability":"tool-use","technique":null,"statement":"With the agent loop, skills and prompts held fixed and only the model swapped, models split into two clusters on operating a multi-step asynchronous ML workflow: Claude Sonnet 3.5-4.0, Gemini 2.5 and GPT-5 completed tasks at 93-100%, while GPT-4 scored 13% and others could not run the loop at all, hallucinating workflow ids or failing to wait for asynchronous jobs; on open-ended exploration the effect of a stressful, competitive prompt reversed by family, with Sonnet 4.0 producing the largest improvement of any configuration under stress and GPT-5 turning conservative and giving back most of its basic-prompt gains.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08248","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-loop-skills-prompts-held-fixed-only-model-swapped"},{"ref":"CLM-0017","id":"agent-must-work-over-very-large-tool-catalogue-wrapping","capability":"tool-use","technique":null,"statement":"When an agent must work over a very large tool catalogue, wrapping each tool in a natural-language interface that resolves the API schema internally — so raw schemas are retrieved on demand rather than enumerated in the model's prompt — contributes more to multi-step tool- call success than the surrounding planner, router, or verifier stages.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01736","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-must-work-over-very-large-tool-catalogue-wrapping"},{"ref":"CLM-0018","id":"agent-pipelines-generate-whole-domain-specific-ml-codebases-medical-imaging","capability":"code-generation","technique":null,"statement":"In agent pipelines that generate whole domain-specific ML codebases (medical imaging) from scratch, execution/assembly validation and domain knowledge bases fix different failures: removing runtime and assembly testing mainly costs autonomy (more human interventions to debug), while removing the curated domain knowledge base mainly costs output quality (syntactically fine but domain-inappropriate pipelines) with little change in intervention count.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29016","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-pipelines-generate-whole-domain-specific-ml-codebases-medical-imaging"},{"ref":"CLM-0358","id":"agent-system-evolves-own-harness-failure-traces-batching-failures","capability":"harness-self-improvement","technique":"attribute-recurring-failures-before-editing-harness","statement":"When an agent system evolves its own harness from failure traces, batching failures across tasks before editing buys efficiency but not better failure attribution: pooled across three agent benchmarks, the share of harness edits that merely accommodate the current model rose from 60.0% under per-failure serial editing to 75.6% under cross-task aggregation, and only fell to 45.5% when a separate analyst/critic/engineer refinement stage attributed each recurring pattern before the edit was written.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-11677","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-system-evolves-own-harness-failure-traces-batching-failures"},{"ref":"CLM-0019","id":"agent-time-horizon-at-50-percent-reliability-doubles-about-every-seven-months","capability":"long-horizon-execution","technique":null,"statement":"On 170 software and research tasks timed with human domain experts, the length of task that frontier agents complete at 50% reliability, measured in human time, has doubled roughly every seven months since 2019 and was around 110 minutes for o3 in mid-2025 — but the horizon at 80% reliability follows the same trend at roughly five times shorter, and success drops as tasks carry more of the 16 \"messiness\" factors that separate benchmark tasks from real work, so the headline horizon overstates what an agent can be relied on to finish.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2503-14499","stance":"supports"}],"url":"https://rsiratchet.com/claims/agent-time-horizon-at-50-percent-reliability-doubles-about-every-seven-months"},{"ref":"CLM-0233","id":"agentic-customer-service-settings-where-deployer-incentive-conflicts-user-s","capability":"strategic-deception","technique":"knowledge-gated-deception-scoring","statement":"In agentic customer-service settings where a deployer incentive conflicts with a user's documented entitlement, a model's willingness to lie under incentive alone is not predicted by its willingness to lie when explicitly told to — some models comply with explicit deception instructions at high rates while almost never initiating deception under incentive, so instructed-deception evaluations measure capability rather than propensity.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26372","stance":"supports"}],"url":"https://rsiratchet.com/claims/agentic-customer-service-settings-where-deployer-incentive-conflicts-user-s"},{"ref":"CLM-0020","id":"agentic-customer-service-settings-where-deployer-incentive-conflicts-users-documented","capability":"strategic-deception","technique":null,"statement":"In agentic customer-service settings where a deployer incentive conflicts with a user's documented entitlement, a model's willingness to lie under incentive alone is not predicted by its willingness to lie when explicitly told to — some models comply with explicit deception instructions at high rates while almost never initiating deception under incentive, so instructed-deception evaluations measure capability rather than propensity.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26372","stance":"supports"}],"url":"https://rsiratchet.com/claims/agentic-customer-service-settings-where-deployer-incentive-conflicts-users-documented"},{"ref":"CLM-0359","id":"agentic-retrieval-augmented-multi-hop-qa-7b-class-policy-models-scoring-each","capability":"multi-step-reasoning","technique":"score-reasoning-steps-grounding-final-answer-match","statement":"For agentic retrieval-augmented multi-hop QA with 7B-class policy models, scoring each intermediate step for both logical validity and grounding in the retrieved passages — rather than scoring steps by the final answer — yields higher EM/F1 than outcome-reward and final-answer-based process-reward training, because it removes preference pairs where flawed retrieval happened to reach the right answer.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01658","stance":"supports"}],"url":"https://rsiratchet.com/claims/agentic-retrieval-augmented-multi-hop-qa-7b-class-policy-models-scoring-each"},{"ref":"CLM-0360","id":"agentic-tool-calls-where-parameter-fabricated-rather-than-grounded","capability":"hallucination","technique":"lora-latent-critic-verbalizes-which-parameter-ungrounded","statement":"For agentic tool calls where a parameter is fabricated rather than grounded in the dialogue, sampling-based uncertainty detectors (token entropy, semantic entropy, SEP) fail at near-chance because the model fabricates plausible parameters with high confidence, while the model's hidden states at the moment the parameter is generated do carry a separable grounding signal.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-10430","stance":"supports"}],"url":"https://rsiratchet.com/claims/agentic-tool-calls-where-parameter-fabricated-rather-than-grounded"},{"ref":"CLM-0021","id":"agents-comply-harmful-multi-step-requests-at-high-rates","capability":"goal-conflict-safety","technique":null,"statement":"Agents comply with harmful multi-step requests at high rates, and simple jailbreaks transfer to agent settings.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2410-09024","stance":"supports"}],"url":"https://rsiratchet.com/claims/agents-comply-harmful-multi-step-requests-at-high-rates"},{"ref":"CLM-0332","id":"agents-working-over-heterogeneous-data-exposing-schema-content-tool","capability":"harness-self-improvement","technique":"gated-attributed-typed-edits-agent-queried-ontology","statement":"For agents working over heterogeneous data, exposing a schema/content/tool ontology as an MCP server the agent queries, and refining it by a loop of diagnose, attribute-to-level, typed patch, and a gate that accepts an edit only if a backbone-conditional paired evaluation shows a margin, raised trajectory accuracy on DDR-Bench by an average 17.8 points across six backbones and text-to-SQL EX on BIRD by about 7 points; injecting the same semantic content as a static prompt did not help consistently and cost 15 points on one backbone, and in ablation removing the acceptance gate was the largest loss (-11.2), then attribution (-6.3), diagnosis (-4.8), typed patches (-1.7).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15779","stance":"supports"}],"url":"https://rsiratchet.com/claims/agents-working-over-heterogeneous-data-exposing-schema-content-tool"},{"ref":"CLM-0290","id":"among-6-540-code-comments-referencing-chatgpt-copilot-claude","capability":"software-architecture","technique":null,"statement":"Among 6,540 code comments referencing ChatGPT, Copilot, Claude or Gemini in public Python and JavaScript repositories (2022-2025), 81 also admit technical debt, most often postponed testing of the generated code, incomplete adaptation to the codebase, and limited understanding of what the code does, with a lower share of design-related debt than in ordinary self-admitted debt.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2601-07786","stance":"supports"}],"url":"https://rsiratchet.com/claims/among-6-540-code-comments-referencing-chatgpt-copilot-claude"},{"ref":"CLM-0022","id":"an-approval-gate-holds-only-under-complete-mediation","capability":"goal-conflict-safety","technique":"irreversible-action-gate","statement":"Gating irreversible actions works only if the harness mediates every side effect, not only the branch awaiting approval. Where a gate suspends its own branch while a sibling's effect still executes, the control reads as a barrier and is not one — and the shipped implementations in widely used agent frameworks have this gap rather than it being a misconfiguration.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2607-14166","stance":"supports"}],"url":"https://rsiratchet.com/claims/an-approval-gate-holds-only-under-complete-mediation"},{"ref":"CLM-0547","id":"anonymized-retrospective-social-event-forecasting-entities-replaced-dates-shifted-frontier","capability":"forecasting","technique":"keep-true-false-forecasting-questions-near-1-1","statement":"On anonymized retrospective social-event forecasting (entities replaced, dates shifted), frontier LLMs score much better at assigning low probability to events that did not happen than at assigning high probability to events that did, with a mean false-minus-true calibration gap of about 15.9 points across six models and up to 38 points for one model, so benchmarks weighted toward invented events overstate forecasting skill.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-04009","stance":"supports"}],"url":"https://rsiratchet.com/claims/anonymized-retrospective-social-event-forecasting-entities-replaced-dates-shifted-frontier"},{"ref":"CLM-0023","id":"arithmetic-accuracy-frontier-models-depends-strongly-how-digits","capability":"arithmetic","technique":null,"statement":"Arithmetic accuracy in frontier models depends strongly on how digits are chunked into tokens; a different digit grouping changes error rates.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2402-14903","stance":"supports"}],"url":"https://rsiratchet.com/claims/arithmetic-accuracy-frontier-models-depends-strongly-how-digits"},{"ref":"CLM-0291","id":"asked-draft-clinical-guideline-content-against-expert-references-17-models","capability":"domain-knowledge","technique":null,"statement":"Asked to draft clinical-guideline content against expert references, 17 models and deep-research agents reach at best about 57% accuracy on clinical claims, average under 30% on mechanistic claims and about 22% on differential diagnosis, and stay below 30% on therapeutic claims even for the best system, so expert-level evidence integration in medicine is well beyond current systems.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2604-18418","stance":"supports"}],"url":"https://rsiratchet.com/claims/asked-draft-clinical-guideline-content-against-expert-references-17-models"},{"ref":"CLM-0548","id":"asked-generate-full-procedural-execution-graph-unseen-task-recipe","capability":"planning","technique":"generate-action-skeleton-creator-transition-priors-then-assign","statement":"When asked to generate a full procedural execution graph for an unseen task (recipe) in a specific creator's style, few-shot LLMs assign ingredients to steps better than a statistical two-stage pipeline but predict the number of steps worse, and swapping in a larger LLM backbone does not close that step-count gap.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24369","stance":"supports"}],"url":"https://rsiratchet.com/claims/asked-generate-full-procedural-execution-graph-unseen-task-recipe"},{"ref":"CLM-0198","id":"asked-optimize-code-already-performance-ceiling-frontier-llms-gpt","capability":"hallucination","technique":null,"statement":"When asked to optimize code that is already at its performance ceiling, frontier LLMs (GPT, Claude, Gemini families) never say it is already optimal under a plain \"optimize this code for speed\" prompt — they always emit an edit with a confident, unmeasured speedup claim; adding an explicit instruction to abstain unless >90% confident of an improvement produces abstention on a minority of such cases while leaving the edit rate on genuinely sub-optimal code unchanged.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14839","stance":"supports"}],"url":"https://rsiratchet.com/claims/asked-optimize-code-already-performance-ceiling-frontier-llms-gpt"},{"ref":"CLM-0549","id":"asked-supply-missing-biomedical-reference-models-retrieve-real-paper","capability":"hallucination","technique":null,"statement":"When asked to supply a missing biomedical reference, models that retrieve a real paper often still misstate its journal, year or authors, so verifiable-reference rate overstates usable accuracy: two Claude models with near-identical verifiable rates (77.6% and 76.6%) were correct in every bibliographic field in 54.6% and 19.9% of responses.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14988","stance":"supports"}],"url":"https://rsiratchet.com/claims/asked-supply-missing-biomedical-reference-models-retrieve-real-paper"},{"ref":"CLM-0292","id":"asked-which-two-entities-larger-numeric-attribute-models-often","capability":"factual-recall","technique":null,"statement":"Asked which of two entities is larger on a numeric attribute, models often contradict numbers they can state correctly, following entity popularity, mention order and co-occurrence instead; for 7-8B models a logistic regression on those three surface cues predicts the model's choice better than the model's own numeric recall does, while 32B models selectively use the numbers when they are reliable.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2510-16815","stance":"supports"}],"url":"https://rsiratchet.com/claims/asked-which-two-entities-larger-numeric-attribute-models-often"},{"ref":"CLM-0361","id":"audio-in-context-learning-per-episode-resampled-rules-current-large-audio","capability":"audio-understanding","technique":null,"statement":"In audio in-context learning with per-episode resampled rules, current large audio language models bind arbitrary non-speech sounds to new labels reasonably well (best model 58.5% average over six such tasks at 5 shots) but stay near zero on tasks requiring temporal measurement — duration estimation, onset/offset segmentation, Morse dot-versus-dash — with no model averaging above 7.1%, and that temporal gap does not shrink across model generations while the non-temporal one improves sharply.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-11252","stance":"supports"}],"url":"https://rsiratchet.com/claims/audio-in-context-learning-per-episode-resampled-rules-current-large-audio"},{"ref":"CLM-0362","id":"audio-language-models-feed-encoder-outputs-into-text-llm","capability":"audio-understanding","technique":"pick-audio-encoders-semantic-alignment-fidelity","statement":"In audio language models that feed encoder outputs into a text LLM, downstream understanding accuracy across speech, sound and music is set by how semantically text-aligned the encoder is rather than by whether its representation is discrete or continuous — reconstruction-oriented codecs with high acoustic fidelity underperform, semantically distilled tokenizers can beat their own continuous teacher, and scaling the LLM backbone does not recover the lost information.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22851","stance":"supports"}],"url":"https://rsiratchet.com/claims/audio-language-models-feed-encoder-outputs-into-text-llm"},{"ref":"CLM-0363","id":"audio-temporal-localization-tasks-11-southeast-asian-languages-current-audio","capability":"audio-understanding","technique":null,"statement":"On audio temporal-localization tasks in 11 Southeast Asian languages, current audio LLMs return time spans whose coverage of the true span consistently exceeds their purity, and F1 falls steadily as the recording grows from under 30 seconds to 3 minutes, with several models unable to run at all on the longer bins.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09672","stance":"supports"}],"url":"https://rsiratchet.com/claims/audio-temporal-localization-tasks-11-southeast-asian-languages-current-audio"},{"ref":"CLM-0293","id":"auditing-ai-generated-code-static-smell-detection-code-structure-architecture","capability":"software-architecture","technique":null,"statement":"Auditing AI-generated code with static smell detection at code, structure and architecture level, from single-file tasks up to agent-built systems, more capable models produced more bloated and more coupled code, code volume was a near-perfect predictor of structural decay, and neither functional correctness nor more detailed prompting reduced it, so passing tests does not indicate maintainable design.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2605-02741","stance":"supports"}],"url":"https://rsiratchet.com/claims/auditing-ai-generated-code-static-smell-detection-code-structure-architecture"},{"ref":"CLM-0294","id":"autoformalization-problem-statements-into-lean-4-improves-sharply-formalizer","capability":"formal-math","technique":"tool-feedback-autoformalization-compiler-judge-panel","statement":"Autoformalization of problem statements into Lean 4 improves sharply when the formalizer can call the Lean compiler for syntax and a multi-model judge panel for meaning and revise on the feedback: a 32B model's consistency Pass@1 on CombiBench rose from about 24% without tools to about 65% with them, and an 8B distilled version reached about 91% on FormalMath-Lite, with the judge ensemble holding false-positive consistency below 6%.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2510-06857","stance":"supports"}],"url":"https://rsiratchet.com/claims/autoformalization-problem-statements-into-lean-4-improves-sharply-formalizer"},{"ref":"CLM-0234","id":"automated-search-over-agent-harness-mechanisms-gated-token-efficiency-only","capability":"harness-self-improvement","technique":"self-harness-loop","statement":"When an automated search over agent-harness mechanisms is gated on token efficiency with only a tolerance band on capability, the resulting full stack buys large token savings at a small but consistent score loss, and scores below the best single discovered mechanism — on EdgeBench the four-mechanism stack cut token traffic 44.7-49.0% and cost about a third while scoring 42.0 vs 44.8 for the base harness on GPT-5.6 Sol, whereas adding ObservationPack alone scored 47.2.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-20519","stance":"supports"}],"url":"https://rsiratchet.com/claims/automated-search-over-agent-harness-mechanisms-gated-token-efficiency-only"},{"ref":"CLM-0364","id":"automated-unit-test-generation-domain-specific-java-libraries-nlp-toolkits","capability":"code-generation","technique":"inject-self-refined-llm-test-snippets-search-stalls","statement":"For automated unit test generation on domain-specific Java libraries (NLP toolkits), injecting LLM-written test snippets into a search-based generator's population when the evolutionary search stagnates yields higher branch and line coverage than either the search tool alone or the LLM alone, with the LLM contributing object-construction sequences and domain-shaped string inputs that random search does not reach.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14784","stance":"supports"}],"url":"https://rsiratchet.com/claims/automated-unit-test-generation-domain-specific-java-libraries-nlp-toolkits"},{"ref":"CLM-0365","id":"automatically-rewriting-prompts-multi-agent-llm-system-failures-locating-responsible","capability":"harness-self-improvement","technique":"find-responsible-agent-hint-injection-then-learn-corrected","statement":"When automatically rewriting the prompts of a multi-agent LLM system from its failures, locating the responsible agent by injecting a hint into one agent at a time and using that agent's hint-corrected intermediate output as the supervision signal produces prompt updates that generalize better than critique derived only from the system-level output.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08572","stance":"supports"}],"url":"https://rsiratchet.com/claims/automatically-rewriting-prompts-multi-agent-llm-system-failures-locating-responsible"},{"ref":"CLM-0024","id":"behaviour-preserving-whole-repository-tasks-such-stack-migrations-where-starting-repository","capability":"evaluation-validity","technique":null,"statement":"For behaviour-preserving whole-repository tasks such as stack migrations, where the starting repository already passes its test suite, a fixed behavioural test suite cannot distinguish a completed migration from an untouched or shim-wrapped repository, so adding more tests does not close the gap and a non-behavioural audit with veto power is required.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23564","stance":"supports"}],"url":"https://rsiratchet.com/claims/behaviour-preserving-whole-repository-tasks-such-stack-migrations-where-starting-repository"},{"ref":"CLM-0025","id":"benchmark-headroom-closes-fastest-in-bounded-domains-and-lags-in-stateful-agent-work","capability":"long-horizon-execution","technique":null,"statement":"Across 393 model–benchmark observations in ten capability domains for models released between 2023 and September 2026, the share of benchmark headroom closed is uneven: bounded, readily verified domains are near their ceilings (advanced mathematics 86.4 and graduate-level science 85.8 on a 0–100 index where 0 is the benchmark's entry-year frontier), while interactive, stateful ones lag (software engineering 52.6, search and terminal agents 56.8, tool agents 39.9) — so gains in bounded environments have not transferred uniformly to long, stateful workflows where errors propagate across the trajectory.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-11873","stance":"supports"}],"url":"https://rsiratchet.com/claims/benchmark-headroom-closes-fastest-in-bounded-domains-and-lags-in-stateful-agent-work"},{"ref":"CLM-0366","id":"benign-github-issue-augmented-plausible-sounding-maintenance-requirement-specifies-helper","capability":"secure-coding","technique":"adversarial-issue-descriptions-apr-agents","statement":"When a benign GitHub issue is augmented with a plausible-sounding maintenance requirement that specifies a helper reading an environment variable into a dangerous sink, repository-level APR agents produce patches that fix the original bug and pass the project test suite while containing the requested vulnerability in roughly half of tasks, and the strongest instruction-following coding models are more susceptible, not less.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15963","stance":"supports"}],"url":"https://rsiratchet.com/claims/benign-github-issue-augmented-plausible-sounding-maintenance-requirement-specifies-helper"},{"ref":"CLM-0367","id":"best-of-n-selection-mathematical-reasoning-small-verifier-2m-parameters-reads","capability":"error-detection","technique":"verify-candidates-step-boundary-hidden-states-text","statement":"For best-of-N selection on mathematical reasoning, a small verifier (~2M parameters) that reads the frozen generator's own hidden states at reasoning-step boundaries ranks correct candidates above incorrect ones better than a larger text-only verifier that re-encodes the generated solution, because the correctness signal is already present in the generator's upper-layer representations and must otherwise be inferred indirectly from emitted tokens.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30841","stance":"supports"}],"url":"https://rsiratchet.com/claims/best-of-n-selection-mathematical-reasoning-small-verifier-2m-parameters-reads"},{"ref":"CLM-0550","id":"binary-visual-question-answering-open-lvlms-instructblip-7b-qwen2-5-vl","capability":"hallucination","technique":"probe-consistency-four-axes-one-score","statement":"For binary visual question answering with open LVLMs (InstructBLIP-7B, Qwen2.5-VL, InternVL-4B), a hallucinated answer can be held with high response consistency, so a single consistency/uncertainty score misses it; separating consistency into four probe groups (image vs. text perturbation crossed with affirmative vs. negated statement) and classifying on the joint within- and between-group agreement pattern detects hallucination better than any single-score black-box or white-box baseline.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-03817","stance":"supports"}],"url":"https://rsiratchet.com/claims/binary-visual-question-answering-open-lvlms-instructblip-7b-qwen2-5-vl"},{"ref":"CLM-0026","id":"black-box-hallucination-detection-using-only-sampled-responses-token-log-probabilities","capability":"hallucination","technique":null,"statement":"For black-box hallucination detection using only sampled responses and token log-probabilities, semantic-entropy signals go to zero whenever the samples collapse into one semantic cluster — which happens for a large share of hallucinated queries, up to nearly all of them on tasks with confusable entities and figures — while token-level uncertainty still separates many of those cases, so the two signals fail on different queries and combining them detects more than either alone.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02679","stance":"supports"}],"url":"https://rsiratchet.com/claims/black-box-hallucination-detection-using-only-sampled-responses-token-log-probabilities"},{"ref":"CLM-0551","id":"black-box-jailbreak-defense-test-time-abstracting-each-successful-attack","capability":"harness-self-improvement","technique":"store-jailbreak-failures-wrapper-level-rules-persistent-memory","statement":"For black-box jailbreak defense at test time, abstracting each successful attack into a rule that names the structural wrapper (nested role-play, code obfuscation, flipped text) rather than the harmful topic lets one rule induced from a single prompt cover the rest of that attack family, so attack success falls to near zero after the first batch of ~20 streamed prompts without any parameter updates.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26008","stance":"supports"}],"url":"https://rsiratchet.com/claims/black-box-jailbreak-defense-test-time-abstracting-each-successful-attack"},{"ref":"CLM-0027","id":"building-preference-data-teach-7-8b-instruct-models-follow-multi-constraint","capability":"procedure-following","technique":null,"statement":"When building preference data to teach 7-8B instruct models to follow multi-constraint instructions, perturbing each atomic constraint in several directions so the perturbed instruction's valid-response set contains, partly overlaps, or is disjoint from the original's — and pairing responses drawn from each sub-region — raises constraint adherence more than rejection sampling or teacher-correction pairs, and the gain holds on perturbed instruction variants the baselines barely improve on.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29352","stance":"supports"}],"url":"https://rsiratchet.com/claims/building-preference-data-teach-7-8b-instruct-models-follow-multi-constraint"},{"ref":"CLM-0552","id":"cae-simulation-tasks-openfoam-fenics-pychrono-comsol-run-current","capability":"multi-model-coordination","technique":"drop-cae-scaffolding-supply-solver-tutorials-instead","statement":"For CAE simulation tasks (OpenFOAM, FEniCS, PyChrono, COMSOL) run with a current general-purpose coding-agent harness and a strong backbone, a single agent with no multi-agent role decomposition and no scripted self-critique matches or exceeds the published scores of multi-agent specialized CAE systems, because the harness itself supplies the multi-turn reasoning and execution-feedback repair the decomposition was built to compensate for.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03718","stance":"supports"}],"url":"https://rsiratchet.com/claims/cae-simulation-tasks-openfoam-fenics-pychrono-comsol-run-current"},{"ref":"CLM-0028","id":"calculator-using-model-trained-reinforcement-learning-verifiable-final-answer-rewards-arithmetic","capability":"tool-use","technique":"tool-use-fine-tuning","statement":"When a calculator-using model is trained with reinforcement learning on verifiable final-answer rewards for an arithmetic search task (Countdown), the gain lands almost entirely at low k — pass@1 rises sharply while pass@16 moves little or falls — because the update can only reinforce correct trajectories the starting policy already samples, and prompt groups with no correct sample supply no gradient.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28447","stance":"supports"}],"url":"https://rsiratchet.com/claims/calculator-using-model-trained-reinforcement-learning-verifiable-final-answer-rewards-arithmetic"},{"ref":"CLM-0029","id":"chain-of-thought-does-not-fix-digit-level-arithmetic","capability":"arithmetic","technique":"chain-of-thought","statement":"Writing out reasoning steps improves how a problem is decomposed but does not fix the arithmetic inside a step — the model can decompose correctly and still get the calculation wrong, at the same confident register as when it is right. For exact arithmetic the fix is to move the computation out of the text, not to add more steps.","kind":"mechanism","backing_strength":"replicated","contested":true,"status":"active","sources":[{"source":"arxiv-2211-10435","stance":"supports"},{"source":"arxiv-2305-18654","stance":"supports"},{"source":"arxiv-2402-14903","stance":"contests"}],"url":"https://rsiratchet.com/claims/chain-of-thought-does-not-fix-digit-level-arithmetic"},{"ref":"CLM-0368","id":"chain-of-thought-faithfulness-measured-how-much-internal-concepts-sparse-autoencoder-features","capability":"explanation-faithfulness","technique":"ablate-cot-prediction-shared-sae-concepts-test-faithfulness","statement":"When chain-of-thought faithfulness is measured by how much the internal concepts (sparse-autoencoder features) active during a CoT pass overlap with those active during a direct-answer pass, that overlap barely predicts whether the shared concepts causally drive the answer: across five open models and four reasoning datasets, above-median concept alignment raised the chance of an above-median causal ablation effect by only about 8%, so representational correspondence cannot substitute for an ablation test.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-23065","stance":"supports"}],"url":"https://rsiratchet.com/claims/chain-of-thought-faithfulness-measured-how-much-internal-concepts-sparse-autoencoder-features"},{"ref":"CLM-0199","id":"claim-level-hallucination-detection-long-document-qa-estimating-claim-s-reliability","capability":"hallucination","technique":"atomic-fact-checking","statement":"For claim-level hallucination detection on long-document QA, estimating a claim's reliability from how central it is in a graph of sampled model responses cannot separate true claims from confidently repeated false ones, because both get high agreement; adding trusted source documents as anchor nodes with higher-weight edges shifts the basis from internal consensus to external grounding and yields a Grounded pool with few hallucinations (97.6% precision on QASPER), at the cost of conservatively leaving correct-but-novel claims on the boundary.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04442","stance":"supports"}],"url":"https://rsiratchet.com/claims/claim-level-hallucination-detection-long-document-qa-estimating-claim-s-reliability"},{"ref":"CLM-0369","id":"claim-level-hallucination-detection-pipelines-over-image-video-audio-inputs","capability":"hallucination","technique":"atomic-fact-checking","statement":"In claim-level hallucination detection pipelines over image, video and audio inputs (both comprehension and generation directions), decomposing an output into atomic claims contributes more to detection accuracy than ensembling several verifier models: removing decomposition costs about 6.6-7.9 macro-F1 across six cross-modal tasks while removing multi-expert majority voting costs about 4.5-5.3.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-11244","stance":"supports"}],"url":"https://rsiratchet.com/claims/claim-level-hallucination-detection-pipelines-over-image-video-audio-inputs"},{"ref":"CLM-0235","id":"claim-level-uncertainty-estimated-response-claim-graph-anchoring-graph-trusted-external","capability":"hallucination","technique":"anchor-claim-consistency-graphs-trusted-priors","statement":"When claim-level uncertainty is estimated from a response-claim graph, anchoring the graph with trusted external priors (edges weighted by entailment from verified source sections) and scoring claims by weighted closeness centrality separates ungrounded claims from repeatedly-generated-but-false ones better than internal-consistency centrality alone, and yields retrieval whose answerability tracks its accuracy instead of the near-100% answer rate GraphRAG shows on the same long documents.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04442","stance":"supports"}],"url":"https://rsiratchet.com/claims/claim-level-uncertainty-estimated-response-claim-graph-anchoring-graph-trusted-external"},{"ref":"CLM-0553","id":"clinical-calculator-tasks-where-formula-gold-variables-full-note-all","capability":"arithmetic","technique":"program-aided-computation","statement":"On clinical-calculator tasks where the formula, gold variables and full note are all supplied, having an open-weight model write case-specific Python for a restricted executor instead of doing the arithmetic itself gives no reliable gain at 7B (+3.29 pp, calculator-cluster CI crossing zero) but a clear gain at 32B (+7.05 pp, CI clear of zero), and the remaining errors are invented unit conversions and one-tier slips in banded scores rather than raw computation.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10728","stance":"supports"}],"url":"https://rsiratchet.com/claims/clinical-calculator-tasks-where-formula-gold-variables-full-note-all"},{"ref":"CLM-0030","id":"closed-book-hallucination-detection-training-small-supervised-classifier-over-many","capability":"hallucination","technique":null,"statement":"For closed-book hallucination detection, training a small supervised classifier over many heterogeneous uncertainty scores beats the best single score, but the gain comes from the diversity of sampling-based black-box signals: ensembles over only single-generation token-probability features add little, while black-box-only ensembles do nearly as well as the full set.","kind":"mechanism","backing_strength":"replicated","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24492","stance":"supports"}],"url":"https://rsiratchet.com/claims/closed-book-hallucination-detection-training-small-supervised-classifier-over-many"},{"ref":"CLM-0295","id":"closed-book-multi-hop-qa-accuracy-gain-decomposed-prompting-shrinks-little","capability":"factual-recall","technique":"abstain-disagreement-between-task-equivalent-prompt-regimes","statement":"In closed-book multi-hop QA the accuracy gain from decomposed prompting shrinks to little in frontier models, but answers still change between task-equivalent prompt regimes (agreement only about 70%, and about 60% on MuSiQue), and because stable facts survive re-prompting while hallucinations do not, disagreement across regimes detects errors better than standard uncertainty baselines on F1 and AUROC without retrieval or training.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2602-04853","stance":"supports"}],"url":"https://rsiratchet.com/claims/closed-book-multi-hop-qa-accuracy-gain-decomposed-prompting-shrinks-little"},{"ref":"CLM-0554","id":"closed-box-per-claim-confidence-built-self-verification-cross-sample-consistency-signal-breaks","capability":"abstention-and-self-knowledge","technique":"calibrate-confidence-per-atomic-claim-per-response","statement":"For closed-box per-claim confidence built from self-verification and cross-sample consistency, the signal breaks on adversarial false-premise questions because it faithfully reports the model's own mistaken confidence: on TruthfulQA with GPT-4o claim-level calibration reached ECE 0.412 against 0.118 for response-level temperature scaling, while an active cross-examination protocol (LMvsLM) reached 0.112, so an external probe rather than a confidence estimate is what helps there.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22483","stance":"supports"}],"url":"https://rsiratchet.com/claims/closed-box-per-claim-confidence-built-self-verification-cross-sample-consistency-signal-breaks"},{"ref":"CLM-0370","id":"code-generation-benchmarks-deciding-whether-benchmark-item-leaked-into","capability":"evaluation-validity","technique":"detect-benchmark-leakage-code-specific-membership-classifier","statement":"For code generation benchmarks, deciding whether a benchmark item leaked into a model's training set is done better by a classifier over code-specific signals (test pass rate, CodeBLEU, edit distance, perplexity, CodeBERT embeddings of generated vs reference code) than by perplexity alone, because memorized items show abnormally high functional correctness and similarity to the reference solution, not just low perplexity.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09865","stance":"supports"}],"url":"https://rsiratchet.com/claims/code-generation-benchmarks-deciding-whether-benchmark-item-leaked-into"},{"ref":"CLM-0031","id":"code-generation-prompt-combines-prompt-enforced-output-format-json-or-xml","capability":"code-generation","technique":null,"statement":"When a code-generation prompt combines a prompt-enforced output format (JSON or XML), a persona, and urgency framing at once, pass@1 on function-level Python problems can fall well below the sum of each constraint's individual effect, even when each constraint alone is neutral or helpful — observed in the GPT-4o family and absent in the GPT-4.1 family and o3-mini.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03156","stance":"supports"}],"url":"https://rsiratchet.com/claims/code-generation-prompt-combines-prompt-enforced-output-format-json-or-xml"},{"ref":"CLM-0032","id":"code-prompts-unsatisfiable-design-open-weight-code-reasoning-models-refuse","capability":"hallucination","technique":null,"statement":"On code prompts that are unsatisfiable by design, open-weight code and reasoning models refuse based on how suspicious the request looks rather than how deeply impossible it is: they refuse famous theoretical impossibilities far more often than requests for plausible-sounding nonexistent packages, and the stronger models in the set close the gap on theory while barely improving on fabricated ecosystem entities.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03267","stance":"supports"}],"url":"https://rsiratchet.com/claims/code-prompts-unsatisfiable-design-open-weight-code-reasoning-models-refuse"},{"ref":"CLM-0512","id":"coding-agent-robot-manipulation-where-safety-constraint-do-touch-named","capability":"embodied-control","technique":"re-check-safety-constraint-tool-prompt","statement":"For coding-agent robot manipulation where a safety constraint (do not touch a named obstacle) is given only as text in the task prompt, the agent restates the constraint early and then plans the shortest path through the obstacle, and enforcing the same constraint with an external stateless verifier that re-checks each route against grounded obstacle boxes before motion raises both task success and collision avoidance.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-20822","stance":"supports"}],"url":"https://rsiratchet.com/claims/coding-agent-robot-manipulation-where-safety-constraint-do-touch-named"},{"ref":"CLM-0262","id":"collections-template-generated-pages-delimiter-based-extraction-pattern-induced-handful-labelled","capability":"formal-query-synthesis","technique":"induce-reusable-delimiter-wrapper-labelled-examples","statement":"For collections of template-generated pages, a delimiter-based extraction pattern induced from a handful of labelled examples can be reused across the rest of the collection, so extraction cost per page drops to pattern execution rather than per-page inference.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"kushmerick-2000-wrapper-induction","stance":"supports"}],"url":"https://rsiratchet.com/claims/collections-template-generated-pages-delimiter-based-extraction-pattern-induced-handful-labelled"},{"ref":"CLM-0033","id":"commercial-assistants-show-large-accuracy-drop-when-relevant","capability":"long-term-memory","technique":null,"statement":"Commercial assistants show a large accuracy drop when the relevant information sits in a long interaction history, especially for updates and multi-session reasoning.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2410-10813","stance":"supports"}],"url":"https://rsiratchet.com/claims/commercial-assistants-show-large-accuracy-drop-when-relevant"},{"ref":"CLM-0034","id":"comparing-weight-shared-recursive-model-against-depth-matched-plain-transformer-code","capability":"evaluation-validity","technique":null,"statement":"When comparing a weight-shared recursive model against a depth-matched plain transformer on code generation, the ranking read from a single checkpoint can invert during training, because the larger-parameter arm passes its validation optimum and overfits while the parameter-lean recursive arm keeps improving — so a late-epoch \"recursion fits better\" reading is a degradation artifact, not a capability difference.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29376","stance":"supports"}],"url":"https://rsiratchet.com/claims/comparing-weight-shared-recursive-model-against-depth-matched-plain-transformer-code"},{"ref":"CLM-0555","id":"complex-clinical-diagnosis-published-case-reports-rare-disease-challenging","capability":"multi-model-coordination","technique":"debate-mixture-of-agents-rebut-revise-then-aggregate","statement":"On complex clinical diagnosis from published case reports (rare disease and challenging cases), running several presenter-rebutter debater pairs on one backbone and synthesizing their revised hypotheses with an aggregator raises exact-match diagnostic accuracy and rated safety over the same model prompted once, while giving no measurable gain on multiple-choice clinical knowledge questions.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05069","stance":"supports"}],"url":"https://rsiratchet.com/claims/complex-clinical-diagnosis-published-case-reports-rare-disease-challenging"},{"ref":"CLM-0035","id":"computer-use-agents-operating-real-os-indirect-injection-success-rates-measured","capability":"prompt-injection","technique":null,"statement":"For computer-use agents operating on a real OS, indirect-injection success rates measured with a fixed hand-written injection template understate vulnerability, because an attacker that adapts phrasing from the agent's own refusal trajectories can reach substantial success on models the fixed template never penetrates.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30207","stance":"supports"}],"url":"https://rsiratchet.com/claims/computer-use-agents-operating-real-os-indirect-injection-success-rates-measured"},{"ref":"CLM-0556","id":"confidence-gated-retrieval-agents-commit-retrieve-or-defer-against-fixed","capability":"abstention-and-self-knowledge","technique":"replay-matched-trajectories-compare-confidence-to-action-maps","statement":"In confidence-gated retrieval agents that commit, retrieve or defer against a fixed numerical confidence threshold, applying a post-hoc calibration map to the same confidence signal raises accuracy among committed answers (15.8-41.0 points across three models on HotpotQA and MuSiQue) purely by moving the controller to a more selective operating point, so coverage falls and retrieval cost rises and overall accuracy can go either way (up to +15 points on HotpotQA, down to -17 on MuSiQue).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26846","stance":"supports"}],"url":"https://rsiratchet.com/claims/confidence-gated-retrieval-agents-commit-retrieve-or-defer-against-fixed"},{"ref":"CLM-0263","id":"constraining-localization-agent-priority-queue-over-candidate-code-regions","capability":"repository-navigation","technique":"priority-scheduled-search-distance-aware-context-pruning","statement":"Constraining a localization agent with a priority queue over candidate code regions scored for relevance, decomposition of each search action, and pruning of context by distance from the issue's anchors reached a 65% function-level match rate on SWE-bench Lite, the best open-source result at the time, and feeding its locations to an open-source repair pipeline raised the resolved rate by about 6 points.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2502-00350","stance":"supports"}],"url":"https://rsiratchet.com/claims/constraining-localization-agent-priority-queue-over-candidate-code-regions"},{"ref":"CLM-0371","id":"conversational-long-term-memory-8-14b-backbones-plain-top-k-dense-retrieval","capability":"long-term-memory","technique":"write-memory-into-self-configured-orthogonal-views","statement":"For conversational long-term memory with 8-14B backbones and plain top-K dense retrieval, writing each dialogue span into several complementary low-overlap semantic views before indexing retrieves gold evidence for more queries than structured extraction into a single view, and the advantage is largest at small retrieval budgets.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21940","stance":"supports"}],"url":"https://rsiratchet.com/claims/conversational-long-term-memory-8-14b-backbones-plain-top-k-dense-retrieval"},{"ref":"CLM-0296","id":"county-level-us-knowledge-14-782-questions-over-526-counties","capability":"factual-recall","technique":null,"statement":"On county-level US knowledge (14,782 questions over 526 counties), the best of 13 models reaches only about 57% on narrative questions and no model exceeds 13% on numeric Census facts, larger models are not reliably better, and web search helped Gemini (+14 points) while hurting GPT models (about -11), so hyper-local recall is thin and retrieval does not fix it uniformly.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2511-10459","stance":"supports"}],"url":"https://rsiratchet.com/claims/county-level-us-knowledge-14-782-questions-over-526-counties"},{"ref":"CLM-0276","id":"critique-s-worth-measured-whether-fixed-downstream-repair-agent","capability":"critique","technique":"score-critique-repair-enables","statement":"When a critique's worth is measured by whether a fixed downstream repair agent can improve the interface from it, eight frontier models differ meaningfully in how actionable their UX critiques are, lead on different product-surface categories, and vary in reliability fixture by fixture; the best repair lift was about +0.22 on the benchmark's scale, so the task is unsaturated and not one-dimensional.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-16262","stance":"supports"}],"url":"https://rsiratchet.com/claims/critique-s-worth-measured-whether-fixed-downstream-repair-agent"},{"ref":"CLM-0557","id":"cross-system-urban-tasks-where-several-services-must-be-chained","capability":"tool-use","technique":"augment-constraint-check-each-candidate-before-committing-answer","statement":"On cross-system urban tasks where several services must be chained under matched tools and budgets, the task-success gap between orchestration strategies is not explained by required-tool coverage, call ordering, or asking for missing inputs — those are near-saturated across strategies — but by whether the agent goes beyond the minimal toolchain to refine coarse results and verify each candidate against the stated constraints before answering.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-03018","stance":"supports"}],"url":"https://rsiratchet.com/claims/cross-system-urban-tasks-where-several-services-must-be-chained"},{"ref":"CLM-0036","id":"data-science-python-tasks-where-several-correct-solutions-differ-orders","capability":"code-generation","technique":null,"statement":"For data-science Python tasks where several correct solutions differ by orders of magnitude in runtime, a model's correctness rate does not predict the efficiency of the code it writes: rankings invert between the two, code-specialized open models are less efficient than general open models of similar size class, and the fastest model rotates by library (NumPy vs Pandas vs PyTorch vs Polars).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30248","stance":"supports"}],"url":"https://rsiratchet.com/claims/data-science-python-tasks-where-several-correct-solutions-differ-orders"},{"ref":"CLM-0037","id":"deception-ability-emerged-in-frontier-models","capability":"strategic-deception","technique":null,"statement":"State-of-the-art models such as GPT-4 can understand and induce false beliefs in other agents through deliberate strategic reasoning, a capability that was absent in earlier-generation language models.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2307-16513","stance":"supports"}],"url":"https://rsiratchet.com/claims/deception-ability-emerged-in-frontier-models"},{"ref":"CLM-0372","id":"deception-narrative-text-analysed-instantiating-fixed-argumentation-schemes-llm","capability":"strategic-deception","technique":"flag-deception-argument-schemes-plus-critical-questions","statement":"When deception in narrative text is analysed by instantiating fixed argumentation schemes with an LLM (premise extraction, then critical-question evaluation), the scheme assessments are stable across ten different LLM backbones and repeated runs for clear-cut deception cases but grow inconsistent on ambiguous intelligence-style narratives.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24369","stance":"supports"}],"url":"https://rsiratchet.com/claims/deception-narrative-text-analysed-instantiating-fixed-argumentation-schemes-llm"},{"ref":"CLM-0373","id":"deciding-robot-skill-has-finished-next-subtask-should-begin","capability":"embodied-control","technique":"distill-stage-transition-monitoring-via-student-self-explanations","statement":"For deciding when a robot skill has finished and the next subtask should begin, general-purpose VLMs prompted zero-shot are near useless on long-horizon mobile manipulation (under 4% of transitions detected on BEHAVIOR-1K, including a frontier proprietary model), while a sub-1B student fine-tuned on compact self-explanations condensed from teacher reasoning over annotated demonstrations detects about 90% and lifts closed-loop task success from 0.30 to 0.53 against a 0.63 oracle.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-20791","stance":"supports"}],"url":"https://rsiratchet.com/claims/deciding-robot-skill-has-finished-next-subtask-should-begin"},{"ref":"CLM-0038","id":"declared-dependencies-are-not-the-set-the-code-needs","capability":"secure-coding","technique":"dependency-existence-check","statement":"Checking that every package an agent names actually exists catches the invented ones but not the missing ones: what agents declare is routinely a fraction of what their code imports at runtime, so generated projects fail in a clean environment even when every declared package is real. Existence is a necessary check, not a sufficient one.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2512-22387","stance":"supports"},{"source":"arxiv-2406-10279","stance":"supports"}],"url":"https://rsiratchet.com/claims/declared-dependencies-are-not-the-set-the-code-needs"},{"ref":"CLM-0374","id":"decoder-only-360m-8b-instruct-models-adding-application-supplied-provenance-labels-architectural","capability":"prompt-injection","technique":"make-source-provenance-architectural-attention-bias","statement":"For decoder-only 360M-8B instruct models, adding application-supplied provenance labels as an architectural signal — origin embeddings plus a learned ring-indexed attention bias that suppresses assistant and user queries attending to untrusted-source keys — cuts indirect-injection attack success on unseen task domains and injection styles further than the same two-stage fine-tuning without the architectural components, even though both are equal in-distribution.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21088","stance":"supports"}],"url":"https://rsiratchet.com/claims/decoder-only-360m-8b-instruct-models-adding-application-supplied-provenance-labels-architectural"},{"ref":"CLM-0039","id":"decomposing-long-answers-into-atomic-facts-checking-each","capability":"hallucination","technique":null,"statement":"Decomposing long answers into atomic facts and checking each with search catches most errors and beats human annotators on cost and agreement.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2403-18802","stance":"supports"}],"url":"https://rsiratchet.com/claims/decomposing-long-answers-into-atomic-facts-checking-each"},{"ref":"CLM-0375","id":"decomposing-response-into-atomic-claims-applying-post-hoc-temperature-or","capability":"abstention-and-self-knowledge","technique":"calibrate-confidence-per-atomic-claim-per-response","statement":"Decomposing a response into atomic claims and applying post-hoc temperature or Platt scaling to per-claim self-verification/log-probability confidence lowers expected calibration error on straightforward factual QA (TriviaQA) in closed-box settings, but on adversarial false-premise questions (TruthfulQA) it is worse than response-level temperature scaling, because the per-claim score faithfully reports the model's own mistaken confidence rather than correcting it.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22483","stance":"supports"}],"url":"https://rsiratchet.com/claims/decomposing-response-into-atomic-claims-applying-post-hoc-temperature-or"},{"ref":"CLM-0558","id":"deepfake-video-detection-7-8b-open-multimodal-backbones-splitting-analysis","capability":"multi-model-coordination","technique":"split-forensic-analysis-into-per-cue-agents-reconcile","statement":"For deepfake video detection with 7-8B open multimodal backbones, splitting the analysis into four separately prompted perspective agents (texture, lighting, motion, physics) whose reports a judge agent reconciles beats the same backbone doing chain-of-thought or multi-turn self-questioning over the same four perspectives, and the gain is not explained by extra turns or repeated frame access.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-06865","stance":"supports"}],"url":"https://rsiratchet.com/claims/deepfake-video-detection-7-8b-open-multimodal-backbones-splitting-analysis"},{"ref":"CLM-0040","id":"demonstrates-instructions-planted-retrieved-content-can-hijack-real","capability":"prompt-injection","technique":null,"statement":"Demonstrates that instructions planted in retrieved content can hijack real LLM-integrated applications for data theft and manipulation.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2302-12173","stance":"supports"}],"url":"https://rsiratchet.com/claims/demonstrates-instructions-planted-retrieved-content-can-hijack-real"},{"ref":"CLM-0376","id":"deployed-multilingual-medical-triage-system-most-accuracy-gain-comes","capability":"procedure-following","technique":"externalize-expert-decision-tree-into-clinician-editable-rule","statement":"In a deployed multilingual medical triage system, most of the accuracy gain comes from writing the clinicians' implicit decision tree into an explicit scoped rule set rather than from moving the final decision out of the LLM: replacing a flat danger-sign list with 231 structured rules in the same end-to-end prompt raised recall from 0.565 to 0.837 and F1 from 0.606 to 0.689, while further handing the decision to a deterministic rule engine over LLM-extracted symptoms left accuracy statistically unchanged (recall 0.810, F1 0.702).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09356","stance":"supports"}],"url":"https://rsiratchet.com/claims/deployed-multilingual-medical-triage-system-most-accuracy-gain-comes"},{"ref":"CLM-0559","id":"detecting-findings-conclusion-inconsistencies-oncology-pet-ct-radiology-reports-15m-parameter","capability":"error-detection","technique":"pre-train-compact-encoder-in-domain-reports-consistency","statement":"For detecting findings-conclusion inconsistencies in oncology PET/CT radiology reports, a 15M-parameter BERT pre-trained on in-domain reports reaches higher balanced accuracy and lower false-positive rate than zero- or few-shot prompted 27B-70B open-weight LLMs, and LoRA fine-tuning the 70B model only matches the small model rather than beating it.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30021","stance":"supports"}],"url":"https://rsiratchet.com/claims/detecting-findings-conclusion-inconsistencies-oncology-pet-ct-radiology-reports-15m-parameter"},{"ref":"CLM-0041","id":"detecting-hallucinated-text-spans-vision-language-model-outputs-span-localization","capability":"hallucination","technique":null,"statement":"For detecting hallucinated text spans in vision-language model outputs, span localization and confidence calibration come from different architectures: a fine-tuned generative VLM proposes better span boundaries but scores them poorly, while a discriminative token tagger trained with a direct regression loss on human-consensus confidence gives better-correlated probabilities but conservative spans, so using the generative spans as proposals and the tagger's probabilities as scores beats either alone.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29974","stance":"supports"}],"url":"https://rsiratchet.com/claims/detecting-hallucinated-text-spans-vision-language-model-outputs-span-localization"},{"ref":"CLM-0042","id":"detecting-ungrounded-object-mentions-open-ended-captions-white-box-vision-language-models","capability":"visual-grounding","technique":null,"statement":"For detecting ungrounded object mentions in open-ended captions from white-box vision-language models, a single internal support signal (token likelihood, attention share, or image-text similarity) confuses support that comes from object-specific image evidence with support that comes from scene plausibility or the generated text prefix; scoring both whether image tokens carry object-specific evidence and whether support is more image-derived than prefix-derived separates these and raises detection AUROC.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30480","stance":"supports"}],"url":"https://rsiratchet.com/claims/detecting-ungrounded-object-mentions-open-ended-captions-white-box-vision-language-models"},{"ref":"CLM-0560","id":"deterministic-365-day-e-commerce-business-simulation-1-000-4-000-agent","capability":"long-horizon-execution","technique":"score-long-horizon-agents-separate-capability-axes-one","statement":"In a deterministic 365-day e-commerce business simulation of 1,000-4,000 agent steps, end-of-year assets do not predict the six underlying operating competences: across 18 frontier and open-weight models the top earner ranked 16th of 18 on fraud avoidance and below another model on profit per tool call, and the best negotiator finished mid-pack on assets.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30730","stance":"supports"}],"url":"https://rsiratchet.com/claims/deterministic-365-day-e-commerce-business-simulation-1-000-4-000-agent"},{"ref":"CLM-0043","id":"deterministic-sensors-should-precede-llm-judges","capability":"self-repair","technique":"computational-sensors-first","statement":"In an agent's verify-and-fix loop, deterministic checks — tests, linters, schema validators — should be exhausted before any model-based judge is added, because they are free, fast and give the same verdict every time, while an LLM-as-judge costs tokens per run, returns non-deterministic verdicts, and so belongs only on properties no rule can express, and even then as an advisory signal until measured against human review.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"synthesis-2026-harness-engineering-playbook","stance":"supports"},{"source":"arxiv-2310-01798","stance":"supports"},{"source":"boeckeler-2026-maintainability-sensors","stance":"supports"}],"url":"https://rsiratchet.com/claims/deterministic-sensors-should-precede-llm-judges"},{"ref":"CLM-0044","id":"diffusion-vision-language-models-generating-free-form-captions-object-tokens-hallucinated","capability":"visual-grounding","technique":null,"statement":"In diffusion vision-language models generating free-form captions, object tokens that are hallucinated commit at later denoising steps and with lower per-token confidence than grounded object tokens, so the commit-step trajectory is an available internal hallucination signal that has no counterpart in autoregressive decoding.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01318","stance":"supports"}],"url":"https://rsiratchet.com/claims/diffusion-vision-language-models-generating-free-form-captions-object-tokens-hallucinated"},{"ref":"CLM-0513","id":"disabled-thinking-mode-and-short-traces-are-unreliable-proxies-for-no-thinking","capability":"evaluation-validity","technique":null,"statement":"Evaluations that define \"the model did not think\" by a disabled thinking mode or by the absence of a long trace measure the wrong thing: a disabled mode may still emit reasoning in the answer, and a long trace may hold filler rather than inference; separating the pre-answer text from the answer and scoring it for emptiness, relevance to the question and explicit inference distinguishes answer-only output, relevant but non-inferential text, and genuine reasoning, which the proxies conflate.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2610-11765","stance":"supports"}],"url":"https://rsiratchet.com/claims/disabled-thinking-mode-and-short-traces-are-unreliable-proxies-for-no-thinking"},{"ref":"CLM-0516","id":"distilling-lessons-from-failures-as-well-as-successes-beats-raw-trajectory-memory","capability":"harness-self-improvement","technique":"distilled-experience-memory","statement":"On web-browsing (WebArena, Mind2Web) and software-engineering (SWE-Bench-Verified) benchmarks, an agent that distills reusable strategies from both its successful and its failed attempts, judged by an LLM without ground truth, and retrieves the relevant ones on later tasks, outperforms memories that store raw trajectories or only successful routines — up to 20% relative improvement in success and up to 16% fewer interaction steps — with the weights unchanged; including the failures is part of the gain, not noise.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2509-25140","stance":"supports"}],"url":"https://rsiratchet.com/claims/distilling-lessons-from-failures-as-well-as-successes-beats-raw-trajectory-memory"},{"ref":"CLM-0561","id":"document-image-states-multi-step-selection-constraints-narrative-text-supporting","capability":"multi-step-reasoning","technique":null,"statement":"When a document image states multi-step selection constraints in narrative text and the supporting numbers sit in several charts, multimodal models degrade steadily as either the number of logical hops (2 to 5) or the number of charts (2 to 6) grows, with the best model reaching 62.83% overall against over 90% for human annotators.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02059","stance":"supports"}],"url":"https://rsiratchet.com/claims/document-image-states-multi-step-selection-constraints-narrative-text-supporting"},{"ref":"CLM-0377","id":"domain-specific-language-essentially-absent-pretraining-corpora-form-symbolic-algebra","capability":"code-generation","technique":"mine-training-data-through-language-s-own-interpreter","statement":"For a domain-specific language essentially absent from pretraining corpora (FORM symbolic algebra), QLoRA fine-tuning an 8B model on a few thousand examples validated by executing them through the language's own interpreter beats frontier open-weight models of up to 756B parameters given syntax documentation in context, on execution rate and strict output match for the larger benchmarks, while general reasoning scores drop under 3 points.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-23367","stance":"supports"}],"url":"https://rsiratchet.com/claims/domain-specific-language-essentially-absent-pretraining-corpora-form-symbolic-algebra"},{"ref":"CLM-0378","id":"dpo-alignment-7-8b-instruct-models-alignment-tax-predictable-properties","capability":"fine-tuning-retention","technique":"filter-preference-pairs-reference-margin-length-gap-lexical","statement":"In DPO alignment of 7-8B instruct models, the alignment tax is predictable from properties of the preference pairs themselves: pairs whose chosen/rejected log-probability margin under the reference model is large in magnitude, or whose chosen/rejected token-length difference is large, cause the most loss of general capability, so dropping them keeps most of the alignment gain while cutting forgetting.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24192","stance":"supports"}],"url":"https://rsiratchet.com/claims/dpo-alignment-7-8b-instruct-models-alignment-tax-predictable-properties"},{"ref":"CLM-0379","id":"drone-s-action-space-declared-only-prompt-model-must","capability":"embodied-control","technique":"make-arrival-model-s-own-declared-action-verified","statement":"When a drone's action space is declared only in the prompt and the model must itself declare arrival, small open multimodal models (2B-9B) enter the 5 m success radius as often as or more often than frontier models yet lose most episodes by declaring arrival too early or never declaring, so the gap at small scale is protocol adherence rather than navigation.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01404","stance":"supports"}],"url":"https://rsiratchet.com/claims/drone-s-action-space-declared-only-prompt-model-must"},{"ref":"CLM-0562","id":"eeg-spike-neuroscience-classification-benchmarks-multi-agent-llm-pipeline-planning","capability":"domain-knowledge","technique":"expose-domain-pre-trained-models-one-tool-api","statement":"On EEG and spike neuroscience classification benchmarks, a multi-agent LLM pipeline with planning, code generation and closed-loop retry does not by itself beat a single-shot LLM coding agent; the measured gain comes from making domain-specific pre-trained neural models addressable through one interface and choosing among them per dataset.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01971","stance":"supports"}],"url":"https://rsiratchet.com/claims/eeg-spike-neuroscience-classification-benchmarks-multi-agent-llm-pipeline-planning"},{"ref":"CLM-0380","id":"embodied-environments-high-frequency-action-spaces-atari-minecraft-crafter-dmlab","capability":"embodied-control","technique":"train-world-model-controller-vlm-relabelled-rollouts-then","statement":"In embodied environments with high-frequency action spaces (Atari, Minecraft, Crafter, DMLab, Overcooked, Pico Park, MindCraft), pairing a frozen instruction-tuned VLM that emits sparse text instructions with a small world-model controller trained to follow them scores higher than either the same VLM emitting low-level actions directly or the same controller running without instructions, and the gain depends on the instructions carrying semantic content — random-string instructions score no better than no instructions at all.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26788","stance":"supports"}],"url":"https://rsiratchet.com/claims/embodied-environments-high-frequency-action-spaces-atari-minecraft-crafter-dmlab"},{"ref":"CLM-0563","id":"embodied-multi-agent-household-planning-fixed-high-level-skill-interface-team","capability":"multi-model-coordination","technique":"score-coordination-constructs-separately-task-success","statement":"In embodied multi-agent household planning with a fixed high-level skill interface, a team's task success rate overstates coordination quality, because goal predicates can be reached through illegal shortcuts: several of 11 tested MLLMs succeed on more episodes than they produce legal plans on (Claude Haiku 4.5 63.1% success vs 47.2% legal; Qwen3.6-Plus 69.1% vs 56.7%), while others honour constraints yet miss the goal.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28266","stance":"supports"}],"url":"https://rsiratchet.com/claims/embodied-multi-agent-household-planning-fixed-high-level-skill-interface-team"},{"ref":"CLM-0564","id":"embodied-vision-language-models-trained-separate-capability-specialists-spatial-temporal","capability":"embodied-control","technique":"merge-capability-specialists-then-distil-student-prefixes","statement":"For embodied vision-language models trained as separate capability specialists (spatial, temporal, action guidance, state verification) from one backbone, consolidating them by TIES weight merging and then routed multi-teacher on-policy distillation retains more of each specialist's gain than either step alone or than mixed-data RL, which shows capability-specific biases.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-06756","stance":"supports"}],"url":"https://rsiratchet.com/claims/embodied-vision-language-models-trained-separate-capability-specialists-spatial-temporal"},{"ref":"CLM-0381","id":"encoder-based-hallucination-claim-faithfulness-detectors-moved-general-domain-data-biomedical-claim","capability":"domain-knowledge","technique":"match-encoder-s-pre-training-corpus-target-domain","statement":"For encoder-based hallucination/claim-faithfulness detectors moved from general-domain data to biomedical claim verification, which corpus the encoder was pre-trained on matters more than which data it is fine-tuned on: a PubMed-pretrained encoder fine-tuned on a few hundred SciFact pairs beat both a general encoder fine-tuned on the same pairs and a general encoder pre-fine-tuned on a large general hallucination set.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-11878","stance":"supports"}],"url":"https://rsiratchet.com/claims/encoder-based-hallucination-claim-faithfulness-detectors-moved-general-domain-data-biomedical-claim"},{"ref":"CLM-0382","id":"encoder-projector-decoder-audio-language-models-speaking-style-linearly-decodable-late-audio-encoder","capability":"audio-understanding","technique":null,"statement":"In encoder-projector-decoder audio-language models, speaking style is linearly decodable from late audio-encoder layers at about 82-85% on a 7-way style task, yet open-ended tone prediction at the output reaches only about 20-54%, and which end of that range a model lands on tracks its training objective rather than its encoder: a model trained to generate speech keeps acoustic sensitivity through the decoder while text-only-trained siblings with a nearly identical encoder answer from the words instead.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00727","stance":"supports"}],"url":"https://rsiratchet.com/claims/encoder-projector-decoder-audio-language-models-speaking-style-linearly-decodable-late-audio-encoder"},{"ref":"CLM-0297","id":"encoding-harmful-request-base64-rot13-or-hex-lowers-full","capability":"jailbreak-robustness","technique":null,"statement":"Encoding a harmful request in Base64, ROT13 or hex lowers full compliance in six frontier chat models from about 76% (natural-language prompts) to about 59%, but the least-complied-with encoding still produced malicious output in about 54% of cases, so obfuscation by encoding is a weaker attack than plain non-English text yet is far from neutralized.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-29602","stance":"supports"}],"url":"https://rsiratchet.com/claims/encoding-harmful-request-base64-rot13-or-hex-lowers-full"},{"ref":"CLM-0383","id":"ensembles-heterogeneous-vlms-answering-multimodal-multiple-choice-questions-entropy-uniformly","capability":"abstention-and-self-knowledge","technique":"pool-semantic-opinions-across-models-then-split-entropy","statement":"For ensembles of heterogeneous VLMs answering multimodal multiple-choice questions, the entropy of a uniformly pooled distribution over semantic-equivalence classes of sampled responses ranks wrong answers above right ones better than majority-vote entropy or picking the single lowest-entropy model, and the advantage grows with ensemble size; the same signal measured at the subagent evidence layer of a multi-step multi-agent system only ranks system failures slightly above chance.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05708","stance":"supports"}],"url":"https://rsiratchet.com/claims/ensembles-heterogeneous-vlms-answering-multimodal-multiple-choice-questions-entropy-uniformly"},{"ref":"CLM-0565","id":"entity-centric-evidence-grounded-qa-fact-verification-substituting-named-entities-type-matched","capability":"abstention-and-self-knowledge","technique":"abstain-answers-fail-track-entity-substitutions","statement":"For entity-centric evidence-grounded QA and fact verification, substituting named entities with type-matched synthetic names and checking that the answer transforms correspondingly (back-maps to the original answer) detects ungrounded answers that self-consistency and evidence-sufficiency scores both miss, because synthetic names reduce the parametric activation that keeps a memorised answer stable under perturbation.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28018","stance":"supports"}],"url":"https://rsiratchet.com/claims/entity-centric-evidence-grounded-qa-fact-verification-substituting-named-entities-type-matched"},{"ref":"CLM-0298","id":"error-injected-into-supplied-chain-thought-models-mostly-carry","capability":"multi-step-reasoning","technique":null,"statement":"When an error is injected into a supplied chain of thought, models mostly carry it to the answer: a planted arithmetic error costs small models 50-60% accuracy and 18% on average, though the largest models lose only 3-7%; unit-conversion errors cost about 9% at every scale; skipped steps and sycophantic steps cost small models about 10%; extra correct steps cost almost nothing (0-6%).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2603-03332","stance":"supports"}],"url":"https://rsiratchet.com/claims/error-injected-into-supplied-chain-thought-models-mostly-carry"},{"ref":"CLM-0384","id":"esg-report-rag-question-answering-retrieval-held-fixed-single-top-1","capability":"domain-knowledge","technique":null,"statement":"In ESG-report RAG question answering with retrieval held fixed at a single top-1 chunk, open-weight 2B-30B generators vary widely in how well they stay anchored to the retrieved passage (faithfulness 0.607-0.822) yet all land in a narrow low band of factual correctness (0.387-0.449), and the most faithful model is the least factually correct, so groundedness to context and correctness of extracted ESG facts do not track each other.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15242","stance":"supports"}],"url":"https://rsiratchet.com/claims/esg-report-rag-question-answering-retrieval-held-fixed-single-top-1"},{"ref":"CLM-0385","id":"event-conditioned-time-series-prediction-finance-traffic-models-cite-factor-controlled","capability":"explanation-faithfulness","technique":"score-cited-factors-whether-paired-intervention-moves-prediction","statement":"In event-conditioned time-series prediction (finance and traffic), models cite the factor that a controlled intervention actually manipulated in nearly all cases while their predictions change with that factor far less often — on traffic series-side pairs citation exceeds 90% for most of ten models but pair-level behavioral support stays under 22% — so citation of the right factor is not evidence the prediction depended on it.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24677","stance":"supports"}],"url":"https://rsiratchet.com/claims/event-conditioned-time-series-prediction-finance-traffic-models-cite-factor-controlled"},{"ref":"CLM-0299","id":"every-state-of-the-art-model-tested-repairs-fewer-defects4j-gitbug-java-bugs","capability":"debugging","technique":null,"statement":"Every state-of-the-art model tested repairs fewer Defects4J and GitBug-Java bugs after semantics-preserving transformation of the buggy code, from -4% for GPT-4o to -16% for Llama-3.1, and the size of each drop correlates with how likely the model was to have memorized the original (negative log-likelihood), so part of published repair success on these benchmarks is recall of seen fixes rather than debugging.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2604-21579","stance":"supports"}],"url":"https://rsiratchet.com/claims/every-state-of-the-art-model-tested-repairs-fewer-defects4j-gitbug-java-bugs"},{"ref":"CLM-0331","id":"evolving-agent-harness-2-000-executable-tasks-disjoint-evaluation","capability":"harness-self-improvement","technique":"contrastive-module-scoped-harness-evolution-benchmark-disjoint-tasks","statement":"Evolving an agent harness on 2,000 executable tasks disjoint from the evaluation benchmarks, by contrasting successful and failed trajectories on the same task to find recurring deficiencies and evolving five harness modules (agent loop, tool use, observation management, context management, completion detection) separately before integrating them, raised DeepSeek-V4-Flash from 47.6 to 52.4 on TerminalBench 2.0 and 73.4 to 76.5 on SWE-Bench Verified, transferred across domains and to other backbones frozen (GLM-5.2 59.6 to 61.8, MiniMax-2.5 41.6 to 44.9), while evolving the whole harness at once or without modules fell below the unevolved baseline (46.4 and 44.2).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14857","stance":"supports"},{"source":"dair-2026-modularrsi-three-defects","stance":"supports"}],"url":"https://rsiratchet.com/claims/evolving-agent-harness-2-000-executable-tasks-disjoint-evaluation"},{"ref":"CLM-0517","id":"experience-memory-tolerates-an-imperfect-self-judge","capability":"harness-self-improvement","technique":"distilled-experience-memory","statement":"When an agent labels its own past attempts as successes or failures to build an experience memory, the labels can be substantially wrong without much cost: the LLM judge used was 72.7% accurate against ground truth on WebArena-Shopping, and simulating judges anywhere from 70% to 90% accurate gave similar success rates, with ground-truth labels best — so for distilling lessons, a rough self-judgment is close to enough, although a perfect one still helps.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2509-25140","stance":"supports"}],"url":"https://rsiratchet.com/claims/experience-memory-tolerates-an-imperfect-self-judge"},{"ref":"CLM-0386","id":"expert-authored-saudi-dialect-prompts-scored-against-rubric-criteria-derived-expert","capability":"cultural-competence","technique":null,"statement":"On expert-authored Saudi-dialect prompts scored against rubric criteria derived from expert ground truths, four frontier models all score under 55% and their dominant failure is distorting register and flattening pragmatic nuance (37.3% of 466 errors) rather than stating falsehoods (11.2%), so MSA-fluent models are not dialectally competent.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29990","stance":"supports"}],"url":"https://rsiratchet.com/claims/expert-authored-saudi-dialect-prompts-scored-against-rubric-criteria-derived-expert"},{"ref":"CLM-0566","id":"extended-reasoning-budgets-improve-model-s-execution-arithmetic-procedures","capability":"arithmetic","technique":"chain-of-thought","statement":"Extended reasoning budgets improve a model's execution of arithmetic procedures on operands longer than those seen in training far more than they improve surface-form-to-value tasks like magnitude comparison and format conversion, because chain-of-thought supplies external working memory for carries and column alignment but cannot change how the input was tokenized.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-13129","stance":"supports"}],"url":"https://rsiratchet.com/claims/extended-reasoning-budgets-improve-model-s-execution-arithmetic-procedures"},{"ref":"CLM-0045","id":"external-feedback-repair-works-only-with-real-grounding","capability":"self-repair","technique":"external-feedback-repair","statement":"Reflect-and-retry raises task success when the feedback in the loop is a genuine external signal — a failing test, a compiler error, an environment outcome — but not when the \"feedback\" is the model's own unaided critique. The grounding, not the reflection step, is what does the work.","kind":"mechanism","backing_strength":"replicated","contested":true,"status":"active","sources":[{"source":"arxiv-2303-11366","stance":"supports"},{"source":"arxiv-2310-01798","stance":"supports"},{"source":"arxiv-2303-17651","stance":"contests"}],"url":"https://rsiratchet.com/claims/external-feedback-repair-works-only-with-real-grounding"},{"ref":"CLM-0253","id":"extracting-fields-vertical-websites-decomposing-xpath-induction-into-separate","capability":"formal-query-synthesis","technique":"induce-xpath-seed-pages-via-multi-task-decomposition","statement":"For extracting fields from vertical websites, decomposing XPath induction into separate sub-tasks over a few seed pages lets an LLM emit an XPath that generalizes to unseen pages of the same site, beating single-prompt CoT and Reflexion XPath generation by 14.84-23.86 F1 points, while calling the model only a handful of times instead of once per page.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"huang-song-2025-xpath-generation-agents","stance":"supports"}],"url":"https://rsiratchet.com/claims/extracting-fields-vertical-websites-decomposing-xpath-induction-into-separate"},{"ref":"CLM-0387","id":"extracting-structured-rules-multimodal-enterprise-guideline-documents-wrapping-extraction","capability":"hallucination","technique":"gate-extraction-through-versioned-rule-store-schema-consistency","statement":"When extracting structured rules from multimodal enterprise guideline documents, wrapping the extraction in deterministic schema validation, an LLM semantic judge, and a duplicate/contradiction consistency filter over a shared rule store lowers the share of extracted rules with no matching source sentence from about 15.7% to 3.2% relative to a one-pass pipeline using the same parser and models.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17775","stance":"supports"}],"url":"https://rsiratchet.com/claims/extracting-structured-rules-multimodal-enterprise-guideline-documents-wrapping-extraction"},{"ref":"CLM-0200","id":"factual-paired-candidate-qa-where-model-already-holds-needed-knowledge","capability":"hallucination","technique":null,"statement":"For factual paired-candidate QA where the model already holds the needed knowledge, hallucination can be detected by removing the query keyword that most drives the chosen answer and seeing whether the prediction systematically recovers to the alternative — erroneous answers show that recovery while correct answers merely become ambiguous — but the signal weakens when the knowledge is absent and largely fails on multi-step arithmetic reasoning, where errors are distributed across steps rather than bound to one keyword.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15106","stance":"supports"}],"url":"https://rsiratchet.com/claims/factual-paired-candidate-qa-where-model-already-holds-needed-knowledge"},{"ref":"CLM-0388","id":"factual-qa-817-item-truthfulqa-multiple-choice-plus-300-item-natural-questions","capability":"abstention-and-self-knowledge","technique":"gate-answer-self-scored-information-needs","statement":"On factual QA (817-item TruthfulQA multiple-choice, plus a 300-item Natural Questions short-answer set), a prompt-only three-stage gate in which the model lists the information units a question needs, scores its support for each, and answers only when the mean score clears a threshold lowers the unconditional wrong-answer rate and raises accuracy on answered items for all eleven models tested, but most of the abstentions it produces fall on questions plain chain-of-thought answered correctly (about 69% at the primary operating point), so the gain is bought with lost coverage rather than with selective self-knowledge.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17516","stance":"supports"}],"url":"https://rsiratchet.com/claims/factual-qa-817-item-truthfulqa-multiple-choice-plus-300-item-natural-questions"},{"ref":"CLM-0046","id":"few-shot-arithmetic-accuracy-correlates-how-often-specific-numbers","capability":"arithmetic","technique":null,"statement":"Few-shot arithmetic accuracy correlates with how often the specific numbers appear in pretraining data, evidence of memorization.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2202-07206","stance":"supports"}],"url":"https://rsiratchet.com/claims/few-shot-arithmetic-accuracy-correlates-how-often-specific-numbers"},{"ref":"CLM-0567","id":"file-based-sandbox-agent-tasks-distractor-artefacts-scoring-only-whether","capability":"long-horizon-execution","technique":"require-non-target-files-be-byte-identical-just","statement":"In file-based sandbox agent tasks with distractor artefacts, scoring only whether the task goal is met overstates reliability, because agents frequently complete the target while modifying unrelated files: across nine LLM agents, pass rate exceeds a success-plus-preservation metric by roughly 10-17 points, with the gap widening as model capability falls.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01056","stance":"supports"}],"url":"https://rsiratchet.com/claims/file-based-sandbox-agent-tasks-distractor-artefacts-scoring-only-whether"},{"ref":"CLM-0047","id":"filtering-unsupported-llm-generated-vulnerability-reports-security-triage-statistical-machine-text","capability":"hallucination","technique":null,"statement":"For filtering unsupported LLM-generated vulnerability reports in security triage, statistical machine-text detectors are the wrong instrument: they score provenance, not correctness, and human-written security reports are template-driven and low-perplexity, so they trip the same signal that flags machine text.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25667","stance":"supports"}],"url":"https://rsiratchet.com/claims/filtering-unsupported-llm-generated-vulnerability-reports-security-triage-statistical-machine-text"},{"ref":"CLM-0389","id":"financial-valuation-listed-companies-two-stage-domain-adaptation-large-scale-financial-corpus","capability":"domain-knowledge","technique":"two-stage-financial-domain-adaptation-corpus-pre-train","statement":"For financial valuation of listed companies, a two-stage domain adaptation — large-scale financial-corpus pre-training followed by institution-specific fine-tuning — lowers target-price error more than single-stage fine-tuning or zero-shot prompting of a general model.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22724","stance":"supports"}],"url":"https://rsiratchet.com/claims/financial-valuation-listed-companies-two-stage-domain-adaptation-large-scale-financial-corpus"},{"ref":"CLM-0300","id":"fine-tuned-program-repair-models-score-well-humaneval-java-bug-lose-more-than","capability":"debugging","technique":null,"statement":"Fine-tuned program-repair models that score well on HumanEval-Java-Bug lose more than half their Pass@10 under semantics-preserving rewrites of the buggy code, with local-variable renaming alone cutting performance by 50-57%, so much of measured repair skill is tied to the exact surface form of the bug.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2605-02215","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuned-program-repair-models-score-well-humaneval-java-bug-lose-more-than"},{"ref":"CLM-0390","id":"fine-tuning-14b-model-knowledge-graph-derived-multiple-choice-qa-single-medical-domain","capability":"multi-step-reasoning","technique":"repair-one-hop-gaps-add-same-chunk-context","statement":"When fine-tuning a 14B model on knowledge-graph-derived multiple-choice QA for a single medical domain, attaching supporting triples from the same source text chunk to each target triple during supervision, and repairing residual one-hop failures before reinforcement learning, gives higher accuracy on held-out 3-, 4- and 5-hop questions than training on isolated triples or running RL from an unrepaired checkpoint.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-12230","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-14b-model-knowledge-graph-derived-multiple-choice-qa-single-medical-domain"},{"ref":"CLM-0568","id":"fine-tuning-4b-code-model-training-auxiliary-targets-name-each","capability":"code-generation","technique":"inject-proof-graph-correctness-obligations-then-consolidate-code","statement":"When fine-tuning a 4B code model, training on auxiliary targets that name each program region's correctness obligations (derived from aligning an AST implementation graph with Lean proof-flow traces) raises solved counts over code-only SFT, but the auxiliary-target stage alone shifts the model off the Python-only output distribution and must be followed by a code-only consolidation stage to avoid losing broad function-level ability.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-11394","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-4b-code-model-training-auxiliary-targets-name-each"},{"ref":"CLM-0391","id":"fine-tuning-7-8b-code-models-llm-synthesized-verilog-instruction-corpus-training","capability":"code-generation","technique":"select-training-data-quality-structural-diversity","statement":"When fine-tuning 7-8B code models on an LLM-synthesized Verilog instruction corpus, training on a 20-25% subset selected for functional correctness, instruction-solution alignment, and coverage of Verilog structural clusters (AST, CFG, netlist features) yields higher Pass@1 than training on the full corpus, because the automated corpus carries noise and redundancy that filtering removes.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22765","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-7-8b-code-models-llm-synthesized-verilog-instruction-corpus-training"},{"ref":"CLM-0048","id":"fine-tuning-base-model-instruction-data-removing-training-targets-any","capability":"hallucination","technique":null,"statement":"When fine-tuning a base model on instruction data, removing from the training targets any factual claim the base model cannot consistently recall raises the supported-claim rate of later long-form generations, but the gain comes largely from a more conservative response policy — more refusals and fewer supported claims per answer — not from generating more correct facts at equal coverage.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30987","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-base-model-instruction-data-removing-training-targets-any"},{"ref":"CLM-0301","id":"fine-tuning-cultural-alignment-scores-raises-agreement-country-s-majority","capability":"cultural-competence","technique":null,"statement":"Fine-tuning for cultural alignment scores raises agreement with a country's majority answers on the World Values Survey while collapsing the spread of answers: tuned models anchor to the plurality option (a 39% human share in the example) and cover less than half of the human response dimensions, dropping to a quarter for one model, so alignment gains partly measure flattening rather than cultural understanding.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00565","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-cultural-alignment-scores-raises-agreement-country-s-majority"},{"ref":"CLM-0049","id":"fine-tuning-large-corpus-real-api-call-traces-markedly-improves","capability":"tool-use","technique":null,"statement":"Fine-tuning on a large corpus of real API-call traces markedly improves multi-tool task completion in open models.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2307-16789","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-large-corpus-real-api-call-traces-markedly-improves"},{"ref":"CLM-0050","id":"fine-tuning-simple-synthetic-examples-where-users-opinion-irrelevant","capability":"sycophancy","technique":null,"statement":"Fine-tuning on simple synthetic examples where the user's opinion is irrelevant to the answer reduces sycophancy substantially.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2308-03958","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-simple-synthetic-examples-where-users-opinion-irrelevant"},{"ref":"CLM-0051","id":"fine-tuning-small-multimodal-model-synthesized-constraint-rich-instructions-raises-adherence","capability":"procedure-following","technique":null,"statement":"Fine-tuning a small multimodal model on synthesized constraint-rich instructions raises adherence to output-level constraints (format, word count, keywords) while lowering accuracy on perception-grounded instructions, so the average can fall below the base model unless the synthesis loop tracks per-constraint failures and image compatibility.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26013","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-small-multimodal-model-synthesized-constraint-rich-instructions-raises-adherence"},{"ref":"CLM-0052","id":"fine-tuning-small-open-models-4b-8b-qwen3-multi-turn-tool-calling","capability":"tool-use","technique":"tool-use-fine-tuning","statement":"When fine-tuning small open models (4B-8B Qwen3) for multi-turn tool calling, supervising only the recovery steps generated after the first point where the student's rollout leaves the teacher-supported state region — and zeroing the loss on the student's own valid prefix — beats full-trajectory imitation and outcome-reward RL on in-domain and held-out tool benchmarks, and the trained student's successful trajectories get shorter across iterations rather than longer.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18524","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-small-open-models-4b-8b-qwen3-multi-turn-tool-calling"},{"ref":"CLM-0569","id":"fine-tuning-vision-language-action-policy-deployment-shift-parameter-change-needed-concentrates","capability":"embodied-control","technique":"allocate-lora-rank-per-network-region-pre-training","statement":"When fine-tuning a vision-language-action policy for a deployment shift, the parameter change needed concentrates in the functional region matching the shift's modality — appearance in the vision encoder, rephrased instructions in the language backbone, novel objects in the vision encoder plus action head — so allocating LoRA rank per region by a pre-training diagnostic beats uniform-rank LoRA at matched budget, with the advantage largest at small budgets and absent when the shift's demand spreads across regions rather than concentrating.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-18084","stance":"supports"}],"url":"https://rsiratchet.com/claims/fine-tuning-vision-language-action-policy-deployment-shift-parameter-change-needed-concentrates"},{"ref":"CLM-0053","id":"first-evidence-dominance-is-representational-not-decoding","capability":"fact-verification","technique":null,"statement":"Given several conflicting but equally legitimate passages, a model's answer is driven disproportionately by whichever passage comes first, and the skew sits in the representation of the combined prompt rather than in decoding — which is why reordering the passages does not remove it and why steering the activations does.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2609-03148","stance":"supports"},{"source":"arxiv-2305-13300","stance":"supports"}],"url":"https://rsiratchet.com/claims/first-evidence-dominance-is-representational-not-decoding"},{"ref":"CLM-0054","id":"five-assistants-trained-human-feedback-consistently-show-sycophancy","capability":"sycophancy","technique":null,"statement":"Five assistants trained with human feedback consistently show sycophancy across tasks, and human preference data itself rewards it.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2310-13548","stance":"supports"}],"url":"https://rsiratchet.com/claims/five-assistants-trained-human-feedback-consistently-show-sycophancy"},{"ref":"CLM-0570","id":"fixed-7b-code-backbone-short-python-tasks-mbpp-task-aware","capability":"code-generation","technique":"route-prompts-align-lora-them-verify-public-tests","statement":"For a fixed 7B code backbone on short Python tasks (MBPP), task-aware prompt routing plus LoRA fine-tuning on prompts rewritten into the routed style raises pass@1 only slightly (+1.1 points) when one sample is returned, but yields large gains (+9-10 points) once several candidates are sampled and selected by execution against the task's visible public tests.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-03341","stance":"supports"}],"url":"https://rsiratchet.com/claims/fixed-7b-code-backbone-short-python-tasks-mbpp-task-aware"},{"ref":"CLM-0055","id":"fixes-in-the-harness-compound-fixes-in-the-prompt-do-not","capability":"procedure-following","technique":"guide-file","statement":"A correction made in conversation fixes one exchange and is gone when the session ends; the same correction encoded as a guide rule, a sensor, or a permission fixes every future run — so an agent system improves over time only to the extent that failures are converted into harness structure rather than re-applied as prompts, and the rate of new guide rules per week falling is the sign the conversion is working.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"synthesis-2026-harness-engineering-playbook","stance":"supports"},{"source":"hashimoto-2026-ai-adoption-journey","stance":"supports"}],"url":"https://rsiratchet.com/claims/fixes-in-the-harness-compound-fixes-in-the-prompt-do-not"},{"ref":"CLM-0392","id":"forecasting-next-year-music-festival-lineups-link-prediction-between-artists","capability":"forecasting","technique":null,"statement":"On forecasting next-year music festival lineups (link prediction between artists and festivals), zero-shot LLMs prompted for ranked candidate lists score far below specialized temporal-knowledge-graph forecasting models trained on the domain graph, and their errors are dominated by popularity bias toward globally famous artists plus duplicate entries in the ranked list.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24467","stance":"supports"}],"url":"https://rsiratchet.com/claims/forecasting-next-year-music-festival-lineups-link-prediction-between-artists"},{"ref":"CLM-0056","id":"forecasting-scaffolded-system-approaches-human-crowd-accuracy","capability":"forecasting","technique":null,"statement":"A retrieval-augmented forecasting system built on a GPT-4-class model — not the base model alone — approaches, and sometimes exceeds, the accuracy of competitive human forecasters on real prediction-market questions dated after the model's training cutoff.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2402-18563","stance":"supports"}],"url":"https://rsiratchet.com/claims/forecasting-scaffolded-system-approaches-human-crowd-accuracy"},{"ref":"CLM-0057","id":"forecasting-underperforms-humans-without-scaffolding","capability":"forecasting","technique":null,"statement":"Without added retrieval infrastructure, language models underperform human experts at forecasting real-world events from a benchmark of actual forecasting-tournament questions, though accuracy improves with model scale and access to relevant news context.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2206-15474","stance":"supports"}],"url":"https://rsiratchet.com/claims/forecasting-underperforms-humans-without-scaffolding"},{"ref":"CLM-0571","id":"formal-persuasion-dialogues-uk-political-topic-llm-debaters-judged","capability":"persuasion","technique":null,"statement":"In formal persuasion dialogues on a UK political topic, LLM debaters are judged more persuasive than humans by crowd annotators (all nine LLM configurations occupied the top nine subjective ranks) while ranking far lower under argumentation-theoretic adjudication, where a human topped the logical ranking and humans took six of the top ten.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29738","stance":"supports"}],"url":"https://rsiratchet.com/claims/formal-persuasion-dialogues-uk-political-topic-llm-debaters-judged"},{"ref":"CLM-0302","id":"four-diagram-grounded-geometry-benchmarks-multi-agent-pipeline-consistently-helps-open-source","capability":"multi-model-coordination","technique":null,"statement":"On four diagram-grounded geometry benchmarks a multi-agent pipeline consistently helps open-source vision-language models (Qwen-2.5-VL 7B +6.8 and 32B +3.3 points on Geometry3K, with further gains on OlympiadBench and We-Math) but a strong closed model (Gemini 2.0 Flash) generally did better as a single agent on the classic benchmarks, gaining only modestly on the newest one, so agentic decomposition pays for weaker models and less familiar tasks rather than universally.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2512-16698","stance":"supports"}],"url":"https://rsiratchet.com/claims/four-diagram-grounded-geometry-benchmarks-multi-agent-pipeline-consistently-helps-open-source"},{"ref":"CLM-0303","id":"four-experiments-3-996-us-participants-discussing-conspiracy-they","capability":"persuasion","technique":null,"statement":"In four experiments with 3,996 US participants discussing a conspiracy they were unsure about, frontier models argued people into as well as out of belief with no consistent truth advantage: debunking cut belief by about 12 points on a 100-point scale and bunking raised it comparably, and participants rated the bunking model as more informative and trustworthy; the one asymmetry was that debunking produced more large (40+ point) shifts (16% vs 7%), and constraining the bunking model to true statements retained only 57-64% of its effect.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2601-05050","stance":"supports"}],"url":"https://rsiratchet.com/claims/four-experiments-3-996-us-participants-discussing-conspiracy-they"},{"ref":"CLM-0393","id":"frozen-backbone-multi-agent-scientific-discovery-pipeline-hypobench-tasks-distilling-reusable","capability":"harness-self-improvement","technique":"learn-skills-per-stage-supervision-stage-offers","statement":"For a frozen-backbone multi-agent scientific discovery pipeline on HypoBench tasks, distilling reusable procedural skills separately per stage — from discriminator critique of a whole hypothesis batch for generation, and from execution outcomes compared to ground-truth results for testing — beats no skill, an LLM-written skill, and a hand-written skill, and removing each stage's feedback signal costs most of the gain.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25770","stance":"supports"}],"url":"https://rsiratchet.com/claims/frozen-backbone-multi-agent-scientific-discovery-pipeline-hypobench-tasks-distilling-reusable"},{"ref":"CLM-0572","id":"frozen-large-vision-language-models-token-level-hallucination-evidence-distributed-across","capability":"visual-grounding","technique":"read-whole-computational-trace-graph-one-hidden-state","statement":"In frozen large vision-language models, token-level hallucination evidence is distributed across the relational, spatial and sequential structure of the model's internal trace rather than contained in any single token's hidden state, so a detector that reads the attention graph over image patches, query and response tokens with interleaved GNN/ViT/GRU modules localizes hallucinated spans better than flat probes on the same activations.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-10835","stance":"supports"}],"url":"https://rsiratchet.com/claims/frozen-large-vision-language-models-token-level-hallucination-evidence-distributed-across"},{"ref":"CLM-0394","id":"frozen-llm-run-generate-critique-revise-loop-no-external-feedback-accuracy","capability":"self-repair","technique":"stop-self-critique-loop-confirmed-sentinel","statement":"For a frozen LLM run in a generate-critique-revise loop with no external feedback, the accuracy gain is concentrated where single-shot reasoning is already unreliable: on clean multi-step reasoning (BBH, ~74% single-shot) depth makes no difference beyond the confidence interval, while on harder math (MATH, 26.2% single-shot) the same loop gains +14.2 points, with GSM8K in between (+4.2).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18884","stance":"supports"}],"url":"https://rsiratchet.com/claims/frozen-llm-run-generate-critique-revise-loop-no-external-feedback-accuracy"},{"ref":"CLM-0058","id":"frozen-llms-play-werewolf-competently-via-reflection","capability":"strategic-deception","technique":null,"statement":"Frozen LLMs, using only retrieval over past communications and self-reflection rather than fine-tuning, can play the social deduction game Werewolf competently and show emergent strategic behavior, including deception, without being explicitly trained for it.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2309-04658","stance":"supports"}],"url":"https://rsiratchet.com/claims/frozen-llms-play-werewolf-competently-via-reflection"},{"ref":"CLM-0573","id":"frozen-pretrained-time-series-forecaster-leaves-structured-residual-error-having","capability":"forecasting","technique":"mine-frozen-forecaster-s-residual-named-features","statement":"When a frozen pretrained time-series forecaster leaves structured residual error, having an LLM propose named semantic features of that residual (e.g. renewable generation over total load) and admitting them through a validation gate into a small post-hoc corrector reduces test error more than statistical feature banks or compositional search over the raw channels, but the gain scales with how much raw error the backbone leaves and vanishes once the backbone already fits the series.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-05207","stance":"supports"}],"url":"https://rsiratchet.com/claims/frozen-pretrained-time-series-forecaster-leaves-structured-residual-error-having"},{"ref":"CLM-0395","id":"frozen-vla-policies-executing-repetitive-perceptually-aliased-manipulation-tasks","capability":"embodied-control","technique":"advance-progress-memory-only-verified-achievement","statement":"For frozen VLA policies executing repetitive, perceptually aliased manipulation tasks (counting pick-and-place), a progress memory that advances after each attempted action converts transient grasp or placement failures into persistent task-state errors, so writes must be gated on observed physical evidence of the intended transition; gating this way raised success far above the frozen policy and above memory-augmented baselines.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29537","stance":"supports"}],"url":"https://rsiratchet.com/claims/frozen-vla-policies-executing-repetitive-perceptually-aliased-manipulation-tasks"},{"ref":"CLM-0574","id":"frozen-weight-video-agent-open-ended-surgical-video-qa-mining-candidate","capability":"harness-self-improvement","technique":"reward-gated-skill-distillation-agent-s-own-failed","statement":"For a frozen-weight video agent on open-ended surgical video QA, mining candidate retrieval skills from the agent's own low-scoring traces and keeping each only when it raises a four-dimensional validation reward on ~40 held-out questions beats installing a hand-written general 'looking discipline' skill, and the single distilled temporal skill (re-read the window adjacent to a localized event for before/after questions) carries most of the gain.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-14015","stance":"supports"}],"url":"https://rsiratchet.com/claims/frozen-weight-video-agent-open-ended-surgical-video-qa-mining-candidate"},{"ref":"CLM-0396","id":"full-duplex-speech-to-speech-model-emits-function-calls-dedicated-parallel-output","capability":"tool-use","technique":"emit-tool-calls-parallel-function-token-stream","statement":"In a full-duplex speech-to-speech model that emits function calls on a dedicated parallel output stream, choosing the right tool from spoken, disfluent user requests is far easier than filling its arguments: tool-selection F1 reached 82.5% (above two closed realtime APIs) while argument accuracy was 42.2% and end-to-end Pass@1 33.0%, below both.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21967","stance":"supports"}],"url":"https://rsiratchet.com/claims/full-duplex-speech-to-speech-model-emits-function-calls-dedicated-parallel-output"},{"ref":"CLM-0397","id":"fully-connected-networks-llm-agents-llama-3-2-8b-50","capability":"multi-model-coordination","technique":null,"statement":"In fully connected networks of LLM agents (Llama 3.2-8B, 50 agents, 10 rounds) where a fixed minority holds an unchanging extreme stance, neutral agents' textual justifications converge on the biased agents' vocabulary faster and more completely than their numeric stances do, so monitoring only the text overstates how much the group actually agrees.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-18306","stance":"supports"}],"url":"https://rsiratchet.com/claims/fully-connected-networks-llm-agents-llama-3-2-8b-50"},{"ref":"CLM-0059","id":"function-level-algorithmic-tasks-translated-c-model-s-functional-test","capability":"evaluation-validity","technique":null,"statement":"On function-level algorithmic tasks translated to C#, a model's functional test pass rate carries almost no information about the structural quality or runtime cost of the code it writes: across 340 solutions from four frontier models, correctness correlated near zero with a composite quality score (r = 0.075) and model rankings by correctness reversed against rankings by quality, while the structurally simplest generator was also the most memory-hungry.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22529","stance":"supports"}],"url":"https://rsiratchet.com/claims/function-level-algorithmic-tasks-translated-c-model-s-functional-test"},{"ref":"CLM-0060","id":"functional-correctness-testing-is-the-standard-for-grading-code","capability":"code-generation","technique":null,"statement":"Grading generated code by whether it actually passes held-out functional tests, rather than by surface similarity to a reference solution, is what makes code generation gradeable at scale — and it's how the field has graded it since the earliest large benchmarks.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2107-03374","stance":"supports"}],"url":"https://rsiratchet.com/claims/functional-correctness-testing-is-the-standard-for-grading-code"},{"ref":"CLM-0236","id":"general-purpose-factuality-verifiers-entailment-consistency-rag-trained-zero-shot-frontier-llm","capability":"hallucination","technique":"taxonomy-driven-hallucination-injection-review-verification-benchmarks","statement":"General-purpose factuality verifiers (entailment-, consistency-, and RAG-trained) and zero-shot frontier LLM judges perform near chance at spotting unsupported claims in scientific peer reviews that must be checked against the full submitted paper, because they cannot separate paper-unsupported assertions from legitimate evaluative critique; small models fine-tuned on in-domain examples beat them by a wide margin.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03580","stance":"supports"}],"url":"https://rsiratchet.com/claims/general-purpose-factuality-verifiers-entailment-consistency-rag-trained-zero-shot-frontier-llm"},{"ref":"CLM-0061","id":"general-purpose-factuality-verifiers-entailment-consistency-rag","capability":"hallucination","technique":null,"statement":"General-purpose factuality verifiers (entailment-, consistency-, and RAG-trained) and zero-shot frontier LLM judges perform near chance at spotting unsupported claims in scientific peer reviews that must be checked against the full submitted paper, because they cannot separate paper-unsupported assertions from legitimate evaluative critique; small models fine-tuned on in-domain examples beat them by a wide margin.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03580","stance":"supports"}],"url":"https://rsiratchet.com/claims/general-purpose-factuality-verifiers-entailment-consistency-rag"},{"ref":"CLM-0062","id":"generated-checklists-make-instruction-requirements-individually-checkable","capability":"procedure-following","technique":"checklist-decomposition","statement":"Decomposing an instruction into explicit yes/no items and judging against them beats asking for a holistic score, and the same checklist turned on the model's own draft improves the draft. The gain comes from making each requirement separately checkable, not from the model trying harder — so it is largest on instructions carrying several constraints at once.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2410-03608","stance":"supports"}],"url":"https://rsiratchet.com/claims/generated-checklists-make-instruction-requirements-individually-checkable"},{"ref":"CLM-0398","id":"generational-genetic-search-over-llm-agent-generated-scientific-hypotheses-drug-repurposing","capability":"multi-model-coordination","technique":"coordinate-hypothesis-agents-generational-genetic-algorithm","statement":"In a generational genetic search over LLM-agent-generated scientific hypotheses (drug repurposing across 34 cancer types), replacing fitness-guided parent selection with uniform random selection while holding the scientific roles, variation operators and hypothesis count fixed lowers the population's mean and minimum external evidence scores but leaves the maximum unchanged, and removing crossover or mutation altogether produces mixed shifts that do not survive multiple-testing correction — so the selection rule, not the semantic variation operators, is where the measurable coordination gain sits.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15938","stance":"supports"}],"url":"https://rsiratchet.com/claims/generational-genetic-search-over-llm-agent-generated-scientific-hypotheses-drug-repurposing"},{"ref":"CLM-0329","id":"given-published-state-of-the-art-method-starting-problem-fully-autonomous-multi-agent","capability":"harness-self-improvement","technique":"critic-gated-refinement-stages-simulated-rebuttal","statement":"Given a published state-of-the-art method as the starting problem, a fully autonomous multi-agent loop (extract limitations, generate and novelty-filter ideas, run experiments, select the best hypothesis, plan and run ablations, draft the paper, pass a simulated peer-review rebuttal) improved on the human baseline for 86 of 107 top-venue papers (80%) by an average 25% relative, and its papers scored above the mean of accepted ICLR 2026 and NeurIPS 2025 papers under automated AI reviewers; every judgement in that result, from the improvement parsed out of result tables to the review scores, was made by a model, not a person.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-19644","stance":"supports"}],"url":"https://rsiratchet.com/claims/given-published-state-of-the-art-method-starting-problem-fully-autonomous-multi-agent"},{"ref":"CLM-0264","id":"giving-agent-parsed-code-graph-files-classes-functions-import","capability":"repository-navigation","technique":"code-graph-tools-localization-agents","statement":"Giving an agent a parsed code graph (files, classes, functions with import, call and inheritance edges) and tools to search and traverse it lets a fine-tuned 32B open model localize issues about as well as proprietary agents at roughly an eighth of the cost, reaching about 93% file-level accuracy, and better localization carried through to about 12 points more resolved issues at Pass@10 when plugged into a fixed repair pipeline.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2503-09089","stance":"supports"}],"url":"https://rsiratchet.com/claims/giving-agent-parsed-code-graph-files-classes-functions-import"},{"ref":"CLM-0063","id":"goal-conflict-safety-under-adversarial-pressure-claude-3-5-sonnet","capability":"goal-conflict-safety","technique":null,"statement":"Under strong, deliberately adversarial goal pressure in an autonomous-agent setting, Claude 3.5 Sonnet sometimes disabled oversight mechanisms or misreported its own actions to keep pursuing the stated goal — in a minority of trials, and the study explicitly does not claim this happens in ordinary, non-adversarial use.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2412-04984","stance":"supports"}],"url":"https://rsiratchet.com/claims/goal-conflict-safety-under-adversarial-pressure-claude-3-5-sonnet"},{"ref":"CLM-0399","id":"grading-free-form-answers-algorithm-analysis-complexity-theory-questions-against-instructor-reference","capability":"evaluation-validity","technique":"score-proof-answers-binary-pedagogical-rubric-n-grams","statement":"When grading free-form answers to algorithm-analysis and complexity-theory questions against instructor reference answers, n-gram overlap metrics carry almost no signal about answer quality, because logically equivalent proofs differ in notation, variable names and proof strategy: BLEU-4 was 0.0000 on all 179 questions and ROUGE-1 F1 stayed in a narrow 0.088-0.129 band across topics while a rubric score varied from 0.66 to 0.83.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14572","stance":"supports"}],"url":"https://rsiratchet.com/claims/grading-free-form-answers-algorithm-analysis-complexity-theory-questions-against-instructor-reference"},{"ref":"CLM-0064","id":"grading-llm-generated-unit-tests-pass-rate-line-coverage-alone","capability":"evaluation-validity","technique":null,"statement":"When grading LLM-generated unit tests by pass rate and line coverage alone, a share of passing test suites never directly call the function under test — they reach it through wrappers, mocks, or unrelated public APIs — so pass rate overstates test quality, and this gap widens under agentic repository-access workflows and richer file-level context.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25939","stance":"supports"}],"url":"https://rsiratchet.com/claims/grading-llm-generated-unit-tests-pass-rate-line-coverage-alone"},{"ref":"CLM-0400","id":"grammar-constrained-text-to-sql-decoding-small-models-qwen2-5-instruct-0-5b-7b","capability":"formal-query-synthesis","technique":"prefer-beam-search-over-sample-and-vote-under","statement":"For grammar-constrained text-to-SQL decoding with small models (Qwen2.5-Instruct 0.5B-7B on Spider), beam search is at least as accurate as sample-and-vote at every matched compute budget of 1-8, and better in most, reversing the usual unconstrained finding that sampling with voting beats beam search.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25761","stance":"supports"}],"url":"https://rsiratchet.com/claims/grammar-constrained-text-to-sql-decoding-small-models-qwen2-5-instruct-0-5b-7b"},{"ref":"CLM-0065","id":"grounded-reflection-improves-multistep-tasks","capability":"self-repair","technique":null,"statement":"When a model gets a real environment signal after acting — a test result, a tool error, a task-success indicator — reflecting on that signal in words and retrying substantially improves success rates on multi-step coding and decision-making tasks over a single attempt.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2303-11366","stance":"supports"}],"url":"https://rsiratchet.com/claims/grounded-reflection-improves-multistep-tasks"},{"ref":"CLM-0066","id":"grounding-generation-retrieved-documents-improves-factual-accuracy-knowledge-intensive","capability":"hallucination","technique":null,"statement":"Grounding generation in retrieved documents improves factual accuracy on knowledge-intensive tasks.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2005-11401","stance":"supports"}],"url":"https://rsiratchet.com/claims/grounding-generation-retrieved-documents-improves-factual-accuracy-knowledge-intensive"},{"ref":"CLM-0401","id":"grpo-style-rl-post-training-7b-multimodal-reasoning-models-composing-each","capability":"fine-tuning-retention","technique":"stratify-rollouts-confidence-outcome-forming-rl-groups","statement":"In GRPO-style RL post-training of 7B multimodal reasoning models, composing each optimization group from rollouts stratified by length-normalized sequence surprisal and outcome (confident-correct anchors, confident-wrong negatives, low-confidence successes) within the same step keeps path entropy from collapsing and avoids the late-training performance decline GRPO and DAPO show after a few hundred steps.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07148","stance":"supports"}],"url":"https://rsiratchet.com/claims/grpo-style-rl-post-training-7b-multimodal-reasoning-models-composing-each"},{"ref":"CLM-0402","id":"grpo-style-rlvr-math-7b-models-injecting-partial-reasoning-prefixes","capability":"fine-tuning-retention","technique":"complete-prefixes-weaker-out-of-family-model-during","statement":"In GRPO-style RLVR on math for 7B models, injecting partial reasoning prefixes from a small out-of-family model (Gemma-2-2B, LLaMA-3.2-1B) into a fraction of rollouts keeps policy entropy higher early in training and raises pass@k, with gains growing as k rises; the effect comes from distributional unfamiliarity rather than prefix quality, since prefixes from same-family models of equal or higher quality give no gain over vanilla GRPO.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27420","stance":"supports"}],"url":"https://rsiratchet.com/claims/grpo-style-rlvr-math-7b-models-injecting-partial-reasoning-prefixes"},{"ref":"CLM-0201","id":"hallucination-metric-unsupported-claim-rate-computed-over-whatever-claims-model","capability":"evaluation-validity","technique":null,"statement":"When a hallucination metric is an unsupported-claim rate computed over whatever claims a model happens to make, models that avoid committing to verifiable claims or directional outputs score best, so the ranking partly measures abstention rather than grounding — on financial chart-to-recommendation tasks the lowest unsupported-claim model made directional recommendations on only 6.4% of queries and ranked near the bottom once coverage was factored in.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14302","stance":"supports"}],"url":"https://rsiratchet.com/claims/hallucination-metric-unsupported-claim-rate-computed-over-whatever-claims-model"},{"ref":"CLM-0067","id":"handing-agent-context-editing-tools-delete-summarize-note-fold-prompt","capability":"context-hygiene","technique":null,"statement":"Handing an agent context-editing tools (delete, summarize, note, fold) by prompt alone makes mid-size models worse than leaving their context untouched — Qwen3-8B, Qwen3-14B and Gemma4-E4B all scored lower on long-context QA with the tools than without them; only after fine-tuning on trajectories that use the tools does the toolset become a gain.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28476","stance":"supports"}],"url":"https://rsiratchet.com/claims/handing-agent-context-editing-tools-delete-summarize-note-fold-prompt"},{"ref":"CLM-0068","id":"hardware-description-code-verilog-systemverilog-vhdl-amaranth-python-generated-functional","capability":"secure-coding","technique":null,"statement":"When hardware description code (Verilog, SystemVerilog, VHDL, Amaranth-Python) is generated from a functional specification that omits security obligations — as real SoC docs do, keeping countermeasures in separate config files — frontier models pass functional tests far more often than security tests, and the model with the highest functional pass rate is not the safest; naming the relevant CWE in the prompt raises the security pass rate sharply, showing the limit is missing awareness of the required obligation rather than inability to write defensive RTL.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26588","stance":"supports"}],"url":"https://rsiratchet.com/claims/hardware-description-code-verilog-systemverilog-vhdl-amaranth-python-generated-functional"},{"ref":"CLM-0069","id":"harness-changes-alone-can-move-agent-benchmarks-more-than-model-upgrades","capability":"evaluation-validity","technique":null,"statement":"A large share of an agent benchmark score is attributable to the harness around the model rather than the model itself: holding the model fixed and changing only guides, sensors, retries and state management is reported to move the same model from 30.9% to 74.6% on GAIA and 30th to 5th on Terminal Bench — so a score reported for \"a model\" on an agentic benchmark is a score for a model-plus-harness, and comparisons across labs with different harnesses are not comparisons of models.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2605-27922","stance":"supports"},{"source":"hal-gaia-leaderboard-2026","stance":"supports"},{"source":"synthesis-2026-harness-engineering-playbook","stance":"supports"},{"source":"masood-2026-agentic-harness-engineering","stance":"supports"},{"source":"langchain-2026-deep-agents-harness","stance":"supports"},{"source":"boluk-2026-harness-problem-15-llms","stance":"supports"}],"url":"https://rsiratchet.com/claims/harness-changes-alone-can-move-agent-benchmarks-more-than-model-upgrades"},{"ref":"CLM-0304","id":"having-model-score-confidence-step-step-detects-failures-multi-step","capability":"multi-step-reasoning","technique":"stepwise-self-scoring-failure-detection","statement":"Having a model score its confidence step by step detects failures in multi-step reasoning better than one holistic confidence score for the whole answer, with up to 15% higher AUROC on two multi-step benchmarks, so where a chain fails is more visible per step than in the final answer.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2511-07364","stance":"supports"}],"url":"https://rsiratchet.com/claims/having-model-score-confidence-step-step-detects-failures-multi-step"},{"ref":"CLM-0070","id":"hierarchical-llm-agents-acting-gridworld-cooperation-tasks-overcooked-layouts","capability":"embodied-control","technique":null,"statement":"In hierarchical LLM agents acting in gridworld cooperation tasks (Overcooked layouts) with unseen partners, supplying an accurate partner-skill estimate as planner prompt context does not by itself stop the agent from continuing a stale, non-complementary skill; the estimate only changes behaviour when it is wired into an explicit interrupt rule that aborts the in-progress skill when new partner actions are unlikely under the committed estimate.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18490","stance":"supports"}],"url":"https://rsiratchet.com/claims/hierarchical-llm-agents-acting-gridworld-cooperation-tasks-overcooked-layouts"},{"ref":"CLM-0575","id":"hierarchical-multi-agent-system-where-controller-gates-acceptance-worker-outputs","capability":"multi-model-coordination","technique":"gate-worker-outputs-semantic-entropy-plus-density","statement":"In a hierarchical multi-agent system where a controller gates acceptance of worker outputs on semantic entropy and semantic density computed over resampled completions, the intervention helps only where failures come from semantic instability (vague or contradictory outputs) and hurts where they come from evidence selection or reasoning-chain errors \\u2014 gaining on ambiguity/hallucination-prone benchmarks while losing 10-14 points on multi-hop QA.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-14707","stance":"supports"}],"url":"https://rsiratchet.com/claims/hierarchical-multi-agent-system-where-controller-gates-acceptance-worker-outputs"},{"ref":"CLM-0305","id":"hofstede-s-vsm13-survey-most-eight-current-chat-models","capability":"cultural-competence","technique":"cultural-prompting-via-country-naming-system-prompt","statement":"On Hofstede's VSM13 survey, most of eight current chat models answer closest to the United States when no culture is specified; a system prompt naming a target country moves seven of the eight toward that country's profile, with the largest gains for the countries furthest away at baseline (Japan and China, roughly 17-19% closer), yet those two remain the hardest to match even for the DeepSeek models.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2512-12488","stance":"supports"}],"url":"https://rsiratchet.com/claims/hofstede-s-vsm13-survey-most-eight-current-chat-models"},{"ref":"CLM-0071","id":"holding-task-fixed-only-lengthening-input-degrades-reasoning","capability":"long-context-degradation","technique":null,"statement":"Holding the task fixed and only lengthening the input degrades reasoning long before the context limit is reached.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2402-14848","stance":"supports"}],"url":"https://rsiratchet.com/claims/holding-task-fixed-only-lengthening-input-degrades-reasoning"},{"ref":"CLM-0576","id":"hour-scale-mean-88-8-min-open-ended-streaming-video-qa","capability":"video-temporal-reasoning","technique":"split-streaming-video-into-latency-frontend-asynchronous-memory","statement":"On hour-scale (mean 88.8 min) open-ended streaming video QA with causal access, every streaming memory design degrades as the evidence-to-query gap grows, and they fail in different ways: recent-window systems cannot recover distant events (AURA 25.4% within 5 minutes falling to 10.5% beyond 30 minutes), text-summary systems keep temporal coverage but discard visual evidence, and model-internal compression loses fine-grained detail (StreamForest 14.4%, ThinkStream 7.5% on retrospection), while human accuracy without rewatching is 63.4%.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-05703","stance":"supports"}],"url":"https://rsiratchet.com/claims/hour-scale-mean-88-8-min-open-ended-streaming-video-qa"},{"ref":"CLM-0577","id":"hours-long-ml-development-runs-kaggle-style-competitions-agent-scaffolds-each","capability":"planning","technique":"distil-human-planning-practice-into-re-injected-planning","statement":"In hours-long ML development runs on Kaggle-style competitions, agent scaffolds each lock into one strategy policy — the single-loop CLI agent pivots far less than humans (9% vs 25% of transitions) and the evolutionary search agent far more (58%) with pivot gains cancelling out — and neither reopens an approach it had abandoned, even though it recovers scores after setbacks as well as top humans.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26086","stance":"supports"}],"url":"https://rsiratchet.com/claims/hours-long-ml-development-runs-kaggle-style-competitions-agent-scaffolds-each"},{"ref":"CLM-0306","id":"human-annotated-step-level-verification-frontier-model-proofs-recent-imo-putnam-problems","capability":"error-detection","technique":null,"statement":"On human-annotated step-level verification of frontier-model proofs to recent IMO and Putnam problems, most of 29 generative critics and process reward models do little better than chance at finding the first wrong step, a leading math PRM (Qwen2.5-Math-PRM-72B) reaches only about 37% balanced F1, and beyond a few standouts open-source verifiers lag closed models; verification is harder than generation for the same frontier models.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2510-13744","stance":"supports"}],"url":"https://rsiratchet.com/claims/human-annotated-step-level-verification-frontier-model-proofs-recent-imo-putnam-problems"},{"ref":"CLM-0072","id":"identifies-position-verbosity-self-enhancement-biases-strong-judges-while","capability":"judge-position-bias","technique":null,"statement":"Identifies position, verbosity, and self-enhancement biases in strong judges, while also showing high agreement with humans once mitigated.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2306-05685","stance":"supports"}],"url":"https://rsiratchet.com/claims/identifies-position-verbosity-self-enhancement-biases-strong-judges-while"},{"ref":"CLM-0403","id":"image-aesthetics-models-fine-tuned-critique-scoring-competence-producing-reference-aligned-aesthetic","capability":"critique","technique":null,"statement":"For image-aesthetics models fine-tuned for critique and scoring, competence at producing reference-aligned aesthetic critique does not carry over to judging whether an image suits a stated purpose, audience or cultural setting: all four aesthetic specialists tested scored below 30% accuracy on 301 expert-curated suitability cases (best 29.57%) versus 91.69% for the strongest general-purpose model, and they predicted 'suitable' in 59-77% of binary cases against a gold rate of 38.14%.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26713","stance":"supports"}],"url":"https://rsiratchet.com/claims/image-aesthetics-models-fine-tuned-critique-scoring-competence-producing-reference-aligned-aesthetic"},{"ref":"CLM-0404","id":"image-borne-injected-instructions-delivered-tool-using-agents-completion-rate-understates","capability":"prompt-injection","technique":"report-attempted-recognized-rates-just-completion","statement":"For image-borne injected instructions delivered to tool-using agents, completion rate understates exposure: agents act on the injected instruction about ten times more often than they complete the objective (12.8% attempted vs ~1% completed across 720 runs), and most failures stop at the planning step where the model read the payload and simply did not act on it rather than at any architectural barrier.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09404","stance":"supports"}],"url":"https://rsiratchet.com/claims/image-borne-injected-instructions-delivered-tool-using-agents-completion-rate-understates"},{"ref":"CLM-0254","id":"inducing-regular-expressions-handful-labelled-examples-entity-spans-czech","capability":"formal-query-synthesis","technique":"induce-regexes-prompting-then-refine-against-failed-matches","statement":"For inducing regular expressions from a handful of labelled examples (entity spans in Czech clinical notes), prompting an LLM and then iteratively refining the pattern against match failures extracts entities more accurately than genetic-programming regex induction, with the margin largest when few examples are available.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"zelina-2024-llm-regex-czech-clinical","stance":"supports"}],"url":"https://rsiratchet.com/claims/inducing-regular-expressions-handful-labelled-examples-entity-spans-czech"},{"ref":"CLM-0307","id":"inference-time-no-weights-data-or-internals-instantiating-country","capability":"cultural-competence","technique":"persona-disagreement-steering-cultural-alignment","statement":"At inference time and with no weights, data or internals, instantiating a country as a panel of four World-Values-Survey-grounded personas and using their disagreement, not their consensus, as the steering signal reduced cultural misalignment on Moral-Machine-style dilemmas (MultiTP) by 10-24% across 20 countries and seven open backbones from 2B to 70B.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2605-10843","stance":"supports"}],"url":"https://rsiratchet.com/claims/inference-time-no-weights-data-or-internals-instantiating-country"},{"ref":"CLM-0578","id":"initially-identical-llm-agents-act-shared-persistent-world-where","capability":"multi-model-coordination","technique":"coordinate-agents-through-persistent-shared-world-messages","statement":"When initially identical LLM agents act in a shared persistent world where their artifacts remain and are scored by an external simulator, the shared society builds broader and more resilient technology portfolios than a matched best-of-N population of isolated agents, but isolated search still yields the single best-performing artifact.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26081","stance":"supports"}],"url":"https://rsiratchet.com/claims/initially-identical-llm-agents-act-shared-persistent-world-where"},{"ref":"CLM-0405","id":"injecting-new-or-counterfactual-facts-into-3-4b-instruct-model","capability":"knowledge-updating","technique":"inject-golden-answer-off-policy-trajectory-during-rl","statement":"When injecting new or counterfactual facts into a 3-4B instruct model by post-training, supervised fine-tuning on extracted QA pairs binds facts to their training question form — recall of the single fact stays reasonable while multi-source and inferential questions over the same facts collapse — and adding more SFT epochs does not close the gap, whereas mixed-policy RL that supplies the golden answer as an off-policy trajectory when on-policy rollouts fail recovers much of the generalization.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25243","stance":"supports"}],"url":"https://rsiratchet.com/claims/injecting-new-or-counterfactual-facts-into-3-4b-instruct-model"},{"ref":"CLM-0579","id":"injecting-revised-medical-guideline-facts-into-4b-14b-instruct-models","capability":"knowledge-updating","technique":"render-knowledge-updates-shared-pool-extended-matching-items","statement":"When injecting revised medical guideline facts into 4B-14B instruct models by LoRA SFT under a matched token budget, rendering the same update events as extended-matching items — several related vignettes sharing one candidate answer pool — transfers better to a held-out temporally anchored oncology staging benchmark than short-answer, fill-in-the-blank, or per-vignette multiple-select renderings of the identical content.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30405","stance":"supports"}],"url":"https://rsiratchet.com/claims/injecting-revised-medical-guideline-facts-into-4b-14b-instruct-models"},{"ref":"CLM-0073","id":"input-telemetry-cannot-establish-self-improvement","capability":"evaluation-validity","technique":null,"statement":"Reporting how much a lab's models are used in its own research — tokens, lines of code, inference compute, experiments per researcher — cannot show that the models are accelerating the research, because every one of those is an input. Establishing the loop needs an outcome variable over time: algorithmic efficiency as a function of capability, which is the edge that turns a pipeline into a feedback loop.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"active","sources":[{"source":"post-wu-rsi-outcome-variables","stance":"supports"},{"source":"elasticity-2026-economics-of-rsi","stance":"supports"}],"url":"https://rsiratchet.com/claims/input-telemetry-cannot-establish-self-improvement"},{"ref":"CLM-0580","id":"instruction-tuned-models-qwen2-5-7b-32b-make-function-calls-non-english","capability":"tool-use","technique":"score-argument-value-language-consistency-own-metric","statement":"When instruction-tuned models (Qwen2.5 7B-32B) make function calls from non-English user requests, they often select the right tool and argument names but emit argument values in the wrong language, a failure that exact-match or AST tool-calling metrics record as a generic invocation error rather than a language mismatch.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-11715","stance":"supports"}],"url":"https://rsiratchet.com/claims/instruction-tuned-models-qwen2-5-7b-32b-make-function-calls-non-english"},{"ref":"CLM-0581","id":"instruction-tuned-transformer-llms-1-7b-70b-given-structurally-impossible-math","capability":"abstention-and-self-knowledge","technique":"read-impossibility-answerable-null-residual-subspace","statement":"For instruction-tuned transformer LLMs (1.7B-70B) given structurally impossible math or code prompts (division by zero, invalid attribute calls), the pre-generation hidden state carries a linearly readable 'no admissible answer' signal that a single residual-stream direction recovers (mean AUC 0.939 over 22 model-dataset cells), yet that direction is near-orthogonal to the canonical safety-refusal direction (mean cosine 0.087), so answering confidently is a routing failure rather than missing recognition.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29109","stance":"supports"}],"url":"https://rsiratchet.com/claims/instruction-tuned-transformer-llms-1-7b-70b-given-structurally-impossible-math"},{"ref":"CLM-0406","id":"instruction-tuning-8b-open-model-text-to-sql-curated-set-under","capability":"formal-query-synthesis","technique":"select-tiny-sql-tuning-set-schema-coverage-size","statement":"For instruction tuning an 8B open model on text-to-SQL, a curated set of under 900 examples selected to cover every table in the target schemas and score well on question-SQL alignment reaches higher execution accuracy than pipelines trained on tens of thousands to millions of samples, and dropping the schema-coverage term from the selection objective costs more accuracy than dropping any single quality term.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24186","stance":"supports"}],"url":"https://rsiratchet.com/claims/instruction-tuning-8b-open-model-text-to-sql-curated-set-under"},{"ref":"CLM-0265","id":"instruction-tuning-forces-answer-every-question-teaches-models-fabricate","capability":"abstention-and-self-knowledge","technique":"refusal-aware-instruction-tuning","statement":"Instruction tuning that forces an answer to every question teaches models to fabricate when the question lies outside their parametric knowledge; splitting the tuning data by whether the base model already answers correctly and appending \"I am unsure\" to the rest (refusal-aware tuning) raises average precision on known-versus-unknown separation over vanilla tuning, transfers to out-of-domain datasets, and yields better calibrated confidence than post-hoc uncertainty scoring.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2311-09677","stance":"supports"}],"url":"https://rsiratchet.com/claims/instruction-tuning-forces-answer-every-question-teaches-models-fabricate"},{"ref":"CLM-0407","id":"interactive-video-world-models-world-state-communicated-only-through","capability":"state-tracking","technique":"keep-world-state-engine-use-video-model-only","statement":"In interactive video world models, world state communicated only through prompt switching is not reliably preserved across a rollout — characters that should stay alive vanish and killed ones keep moving — whereas keeping the canonical state in an external deterministic engine and compiling it into pixel-aligned spatial controls for the video model keeps the rendered frames consistent with that state.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10540","stance":"supports"}],"url":"https://rsiratchet.com/claims/interactive-video-world-models-world-state-communicated-only-through"},{"ref":"CLM-0074","id":"introduced-grade-school-math-word-problems-2021-models-made","capability":"arithmetic","technique":null,"statement":"Introduced grade-school math word problems on which 2021 models made frequent arithmetic slips, motivating verifier training.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2110-14168","stance":"supports"}],"url":"https://rsiratchet.com/claims/introduced-grade-school-math-word-problems-2021-models-made"},{"ref":"CLM-0266","id":"issue-localization-be-done-retrieval-and-rerank-no-agent-all-embedding","capability":"repository-navigation","technique":"retrieve-and-rerank-localization-trained-mined-issue-patch","statement":"Issue localization can be done as retrieval-and-rerank with no agent at all: an embedding model and reranker trained on issue-to-patched-function pairs mined from public GitHub (SweLoc) localize functions on SWE-bench Lite and LocBench at least as accurately at Acc@10 as agents built on Claude 3.5 Sonnet, at a small fraction of the per-issue cost and with one pass instead of seven to ten model rounds.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2505-07849","stance":"supports"}],"url":"https://rsiratchet.com/claims/issue-localization-be-done-retrieval-and-rerank-no-agent-all-embedding"},{"ref":"CLM-0308","id":"knowledge-editing-methods-such-rome-fine-tuning-insert-stated-fact-reliably","capability":"knowledge-updating","technique":null,"statement":"Knowledge-editing methods such as ROME and fine-tuning insert a stated fact reliably but do not carry its logical consequences: on multi-hop questions generated from knowledge-graph rules that the edit entails, accuracy is up to 24 points lower than on the directly edited fact, so an edited model can assert a new fact and still answer as if its implications never changed.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-10554","stance":"supports"}],"url":"https://rsiratchet.com/claims/knowledge-editing-methods-such-rome-fine-tuning-insert-stated-fact-reliably"},{"ref":"CLM-0408","id":"knowledge-math-medical-qa-benchmarks-8b-open-backbones-llama-3","capability":"multi-model-coordination","technique":"compose-role-steering-vectors-per-query-instead-debating","statement":"On knowledge, math and medical QA benchmarks with 8B open backbones (Llama-3.1-8B-Instruct, Qwen3-8B), the accuracy advantage of multi-agent debate pipelines over a single agent can be matched by a single-turn single agent whose hidden state is steered by a query-composed mixture of role vectors, at roughly a twentieth of the token cost, so the benefit of these pipelines comes from multi-perspective specialization rather than from inter-agent text exchange.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27338","stance":"supports"}],"url":"https://rsiratchet.com/claims/knowledge-math-medical-qa-benchmarks-8b-open-backbones-llama-3"},{"ref":"CLM-0582","id":"label-free-rl-post-training-reasoning-models-deriving-each-model-s","capability":"multi-model-coordination","technique":"reward-each-model-peer-model-s-majority-vote","statement":"In label-free RL post-training of reasoning models, deriving each model's reward from a separately-trained peer model's majority vote — rather than from its own majority vote — avoids the reward collapse and length degeneration that self-rewarding runs show, and the benefit grows as the paired models differ in family or size because their errors overlap less.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-17253","stance":"supports"}],"url":"https://rsiratchet.com/claims/label-free-rl-post-training-reasoning-models-deriving-each-model-s"},{"ref":"CLM-0237","id":"language-conditioned-manipulation-tasks-where-same-base-task-instantiated-increasing","capability":"embodied-control","technique":null,"statement":"On language-conditioned manipulation tasks where the same base task is instantiated at increasing levels of spatial ambiguity and procedural length, current VLA models degrade on fine-grained spatial relations, precise low-level execution, and memory-dependent multi-step planning, so success measured only on predefined simple settings overstates their competence.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05324","stance":"supports"}],"url":"https://rsiratchet.com/claims/language-conditioned-manipulation-tasks-where-same-base-task-instantiated-increasing"},{"ref":"CLM-0202","id":"language-conditioned-vla-manipulation-tasks-instantiated-graded-difficulty-levels-add","capability":"embodied-control","technique":null,"statement":"When language-conditioned VLA manipulation tasks are instantiated at graded difficulty levels that add spatial ambiguity (fine-grained spatial relations, precise placement) and procedural length (multi-stage plans requiring memory of completed steps), representative VLA models degrade on both axes, so success on predefined single-stage settings does not indicate spatial or long-horizon competence.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05324","stance":"supports"}],"url":"https://rsiratchet.com/claims/language-conditioned-vla-manipulation-tasks-instantiated-graded-difficulty-levels-add"},{"ref":"CLM-0409","id":"language-model-must-judge-whether-specific-vehicle-perform-described","capability":"planning","technique":"ground-mission-feasibility-ontology-inferred-capabilities","statement":"When a language model must judge whether a specific vehicle can perform a described mission, it is overconfident about partially-capable configurations — declaring the mission fully possible despite missing hardware — and supplying the required-capability list plus the vehicle's inferred capabilities from an ontology is what corrects that specific error, not the harder fully/impossible cases which the models already get mostly right.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29347","stance":"supports"}],"url":"https://rsiratchet.com/claims/language-model-must-judge-whether-specific-vehicle-perform-described"},{"ref":"CLM-0203","id":"language-model-updated-fine-tuning-substituted-fact-facts-already-answered","capability":"hallucination","technique":null,"statement":"When a language model is updated by fine-tuning on a substituted fact, the facts it already answered correctly that are most likely to flip to wrong are those whose answer entity is referenced by many other facts, and updates to such structurally popular facts corrupt connected facts out to five hops in the fact graph — so popularity, not long-tail rarity, predicts collateral corruption among already-correct knowledge.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08067","stance":"supports"}],"url":"https://rsiratchet.com/claims/language-model-updated-fine-tuning-substituted-fact-facts-already-answered"},{"ref":"CLM-0518","id":"language-models-inherit-the-positive-bias-of-the-literature","capability":"evaluation-validity","technique":null,"statement":"Language models trained on the published literature inherit its positive bias, because failure data is systematically missing from what they read, and that degrades them in three roles at once — as research tools that suggest what to try, as consumers of scarce high-quality training data, and as peer reviewers — so the absence of negative results now costs more than it did when only people read the papers.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2604-06236","stance":"supports"}],"url":"https://rsiratchet.com/claims/language-models-inherit-the-positive-bias-of-the-literature"},{"ref":"CLM-0204","id":"large-audio-language-models-answering-binary-sound-x-present-questions","capability":"hallucination","technique":null,"statement":"In large audio-language models answering binary \"is sound X present\" questions, the yes/no outcome is fixed at the first decoding step and the model's yes-minus-no margin is positive for most clips even with a silent input, so penalizing affirmative tokens at that single step whenever real audio fails to raise the margin above a small threshold increases correct rejection of absent events beyond fixed-strength contrastive decoding, at the cost of lower recall on genuinely present events and lower precision.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07286","stance":"supports"}],"url":"https://rsiratchet.com/claims/large-audio-language-models-answering-binary-sound-x-present-questions"},{"ref":"CLM-0583","id":"large-audio-language-models-asked-return-every-time-span","capability":"audio-understanding","technique":"score-occurrence-count-separately-boundary-overlap","statement":"When large audio language models are asked to return every time span matching a natural-language query, they almost always report fewer intervals than exist: across 21 systems on queries averaging about 4 ground-truth intervals, under-reporting rates run from 79% to 100% and no system exceeds 13.2% exact count accuracy, even for the models with the best overlap scores.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01542","stance":"supports"}],"url":"https://rsiratchet.com/claims/large-audio-language-models-asked-return-every-time-span"},{"ref":"CLM-0410","id":"large-audio-language-models-multiple-choice-audio-reasoning-benchmarks-mmar","capability":"audio-understanding","technique":"score-audio-cots-grounding-coherence-causality","statement":"For large audio language models on multiple-choice audio reasoning benchmarks (MMAR, MMAU-mini), accuracy does not track whether the model's chain of thought is acoustically grounded or causally drives the answer: several models answer correctly from perception steps that assert unsupported labels, and for some models replacing the reasoning chain with vacuous or fabricated content leaves the final answer unchanged.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09681","stance":"supports"}],"url":"https://rsiratchet.com/claims/large-audio-language-models-multiple-choice-audio-reasoning-benchmarks-mmar"},{"ref":"CLM-0267","id":"large-pretrained-models-well-calibrated-multiple-choice-true-false-questions","capability":"abstention-and-self-knowledge","technique":"ask-p-true-after-showing-model-several-own","statement":"Large pretrained models are well calibrated on multiple-choice and true/false questions when answer options are shown as lettered choices, and can predict the probability that their own sampled answer is correct (P(True)) or that they know the answer at all (P(IK)); calibration improves with model size and few-shot prompting, self-evaluation improves when the model sees several of its own samples first, and P(IK) generalizes only partly to new tasks, where its calibration degrades.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2207-05221","stance":"supports"}],"url":"https://rsiratchet.com/claims/large-pretrained-models-well-calibrated-multiple-choice-true-false-questions"},{"ref":"CLM-0309","id":"lean-prover-trained-large-scale-agentic-reinforcement-learning-against-lean","capability":"formal-math","technique":"natural-language-draft-then-formalize-checker-in-the","statement":"A Lean prover trained with large-scale agentic reinforcement learning against the Lean checker, paired with a test-time workflow that drafts in natural language before formalizing, solved 88% of PutnamBench (undergraduate), 80% of Fate-H (graduate) and 33% of Fate-X (PhD-level), and 11 of 12 Putnam 2025 problems, at a smaller compute budget than prior systems, so formal proving at undergraduate level is largely solved while PhD-level remains open.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2512-17260","stance":"supports"}],"url":"https://rsiratchet.com/claims/lean-prover-trained-large-scale-agentic-reinforcement-learning-against-lean"},{"ref":"CLM-0238","id":"linear-probe-tool-calling-model-s-hidden-state-last-generated","capability":"tool-use","technique":"hidden-state-probe-tool-call-errors","statement":"A linear probe on a tool-calling model's hidden state at the last generated token detects incorrect tool-calls — including wrong-but-well-typed argument values that parsers and logs do not catch — and works better on larger models and at middle-to-late layers than at the final layer.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27750","stance":"supports"}],"url":"https://rsiratchet.com/claims/linear-probe-tool-calling-model-s-hidden-state-last-generated"},{"ref":"CLM-0075","id":"linear-probe-tool-calling-models-hidden-state-last-generated-token","capability":"tool-use","technique":null,"statement":"A linear probe on a tool-calling model's hidden state at the last generated token detects incorrect tool-calls — including wrong-but-well-typed argument values that parsers and logs do not catch — and works better on larger models and at middle-to-late layers than at the final layer.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27750","stance":"supports"}],"url":"https://rsiratchet.com/claims/linear-probe-tool-calling-models-hidden-state-last-generated-token"},{"ref":"CLM-0239","id":"llava-1-5-7b-captioning-restricting-lora-adapter-attention-heads-picked","capability":"visual-grounding","technique":"diagnosis-selected-head-restricted-adapter-plus-grounding-gated","statement":"In LLaVA-1.5-7B captioning, restricting a LoRA adapter to attention heads picked by a hallucination diagnosis (attention-to-image drop around hallucinated object words, then ablation screening) reduces object hallucination, while an identically trained layer-matched random-head adapter does not — so the selection, not the added adapter capacity, carries the effect.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24966","stance":"supports"}],"url":"https://rsiratchet.com/claims/llava-1-5-7b-captioning-restricting-lora-adapter-attention-heads-picked"},{"ref":"CLM-0076","id":"llava-15-7b-captioning-restricting-lora-adapter-attention-heads-picked-hallucination","capability":"visual-grounding","technique":null,"statement":"In LLaVA-1.5-7B captioning, restricting a LoRA adapter to attention heads picked by a hallucination diagnosis (attention-to-image drop around hallucinated object words, then ablation screening) reduces object hallucination, while an identically trained layer-matched random-head adapter does not — so the selection, not the added adapter capacity, carries the effect.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24966","stance":"supports"}],"url":"https://rsiratchet.com/claims/llava-15-7b-captioning-restricting-lora-adapter-attention-heads-picked-hallucination"},{"ref":"CLM-0411","id":"llm-acting-orchestrator-dynamically-created-subagents-summarizing-own-past","capability":"multi-model-coordination","technique":"replay-extracted-skills-tricks-model-cards-orchestrator","statement":"For an LLM acting as orchestrator of dynamically created subagents, summarizing its own past trajectories into reusable workflow skills, short operational tricks, and capability cards for the candidate subagent backbones, then injecting those into the orchestrator's context, raises task scores on delegation- and aggregation-heavy tasks, and the subagent model cards are the component whose removal costs most.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30661","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-acting-orchestrator-dynamically-created-subagents-summarizing-own-past"},{"ref":"CLM-0077","id":"llm-agent-must-reconcile-user-supplied-identity-credentials-against-database","capability":"fact-verification","technique":null,"statement":"When an LLM agent must reconcile user-supplied identity credentials against database records before a sensitive read or write, frontier and open models frequently skip the cross-field consistency check and act anyway, and the failure rate barely moves whether the request is simple or has several parallel sub-requests and whether the forged field is visually near-identical to the true one or completely unrelated.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03588","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agent-must-reconcile-user-supplied-identity-credentials-against-database"},{"ref":"CLM-0412","id":"llm-agents-acting-partially-observable-environments-replacing-raw-action-observation","capability":"planning","technique":"show-llm-bayesian-belief-state-history","statement":"For LLM agents acting in partially observable environments, replacing the raw action-observation log in the prompt with an externally maintained Bayesian posterior over latent states makes the agent's action choice a function of the belief alone, removing surface-form sensitivity in which two histories with identical posteriors but different serialized text elicit different actions.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10036","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-acting-partially-observable-environments-replacing-raw-action-observation"},{"ref":"CLM-0413","id":"llm-agents-acting-through-stateful-tools-over-many-turns","capability":"tool-use","technique":"score-whole-agent-trajectories-five-safety-outcomes","statement":"For LLM agents acting through stateful tools over many turns, unsafe policy-violating completions accumulate with the interaction budget and are roughly doubled when the adversarial user conditions each next request on the agent's actual prior response, so single-turn or short-prefix safety evaluation understates the failure rate.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-16305","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-acting-through-stateful-tools-over-many-turns"},{"ref":"CLM-0414","id":"llm-agents-classifying-nodes-text-attributed-graphs-serializing-sampled-neighborhoods","capability":"multi-model-coordination","technique":"pass-graph-structure-permutation-invariant-signature-text","statement":"For LLM agents classifying nodes in text-attributed graphs, serializing sampled neighborhoods into prompt text makes the prediction depend on the arbitrary order of neighbors; replacing that text with a fixed-size, permutation-invariant vector of graph statistics (degree, PageRank, clustering, cross-community fraction, label entropy, majority-label fraction) plus only the top-k semantically relevant node texts raises accuracy and cuts per-message tokens.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09565","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-classifying-nodes-text-attributed-graphs-serializing-sampled-neighborhoods"},{"ref":"CLM-0584","id":"llm-agents-dependent-multi-step-tasks-success-decays-geometrically-number","capability":"long-horizon-execution","technique":"budget-reliability-per-step-instead-reporting-pass-rates","statement":"For LLM agents on dependent multi-step tasks, success decays geometrically with the number of steps under a single per-step reliability parameter that rises with model scale but stays below 1, so long horizons eventually collapse even for the strongest deployed systems — and the decay tracks step count, not context length, since bounding the context window steepens the decay rather than easing it.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01660","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-dependent-multi-step-tasks-success-decays-geometrically-number"},{"ref":"CLM-0585","id":"llm-agents-doing-active-clinical-diagnosis-interactive-patient-environments","capability":"tool-use","technique":"scaffold-diagnosis-evidence-tools-coverage-based-stopping-harness","statement":"For LLM agents doing active clinical diagnosis in interactive patient environments, wrapping the model in a scaffold of role-specialized agents, evidence tools and an explicit evidence state, with a runtime observer that tracks uncertainty and evidence coverage to decide when to stop acquiring evidence, raises both diagnostic accuracy and execution stability over static case-to-answer prompting, with the size of the gain depending on the backbone model.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24570","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-doing-active-clinical-diagnosis-interactive-patient-environments"},{"ref":"CLM-0415","id":"llm-agents-long-horizon-web-penetration-testing-holding-backbone-model","capability":"long-horizon-execution","technique":"externalize-agent-state-onto-fact-intent-dag","statement":"For LLM agents on long-horizon web penetration testing, holding the backbone model fixed and moving global state out of the context window onto a persistent graph of verified facts and fact-conditioned intent edges raises task success most on the hardest tasks, where success depends on facts obtained dozens of steps earlier (88.2% overall and 75.0% hard, vs 44.1% and 25.0% for the strongest shared-plain-text-memory baseline).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07344","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-long-horizon-web-penetration-testing-holding-backbone-model"},{"ref":"CLM-0416","id":"llm-agents-must-compose-then-execute-domain-tool-workflows","capability":"planning","technique":"train-planner-executor-separate-role-adapters","statement":"For LLM agents that must compose and then execute domain tool workflows, training separate adapters for the two roles splits the gains cleanly: supervised fine-tuning of the planner on reference tool sequences mainly improves workflow-order fidelity, while node-level RL on argument-match rewards mainly improves tool-argument grounding and final-answer quality, and combining both beats either on a 9B backbone.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-12533","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-must-compose-then-execute-domain-tool-workflows"},{"ref":"CLM-0586","id":"llm-agents-playing-thousand-step-starcraft-ii-matches-through-high-level","capability":"long-horizon-execution","technique":"gate-between-replaying-retrieved-winning-episodes-fresh-reasoning","statement":"For LLM agents playing thousand-step StarCraft II matches through a high-level text interface, retrieving a handful of time-aligned episodes from a small bank of past winning games and gating between replaying a retrieved action and reasoning afresh over recent observations raises win rate over a summarization-only working-memory baseline at roughly a seventh of the token cost, with the margin widening at the harder difficulty level.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-12626","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-playing-thousand-step-starcraft-ii-matches-through-high-level"},{"ref":"CLM-0417","id":"llm-agents-score-their-own-actions-during-exploration-text-based","capability":"harness-self-improvement","technique":"separate-self-judged-scores-verifier-rewards-measure-both","statement":"For LLM agents that score their own actions during exploration in text-based games and then reuse that experience, the accuracy of those self-scores against the environment verifier does not predict later gains on held-out stricter configurations: block-level correlations between judgment agreement or calibration and the next checkpoint's score change sit near zero, so recognizing a good action is not the bottleneck — converting it into a transferable policy is.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-31100","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-score-their-own-actions-during-exploration-text-based"},{"ref":"CLM-0587","id":"llm-agents-under-fixed-react-harness-answering-questions-over","capability":"multi-step-reasoning","technique":null,"statement":"For LLM agents under a fixed ReAct harness answering questions over executable database-backed APIs plus document retrieval, accuracy falls sharply as the required reasoning depth grows from 1 to 5 hops (most models lose over half their accuracy), and the surviving failures concentrate in language-mediated steps — entity disambiguation, cross-source grounding, schema alignment — rather than in emitting syntactically valid tool calls.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-12282","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-under-fixed-react-harness-answering-questions-over"},{"ref":"CLM-0588","id":"llm-agents-whose-steps-condition-their-own-prior-outputs","capability":"abstention-and-self-knowledge","technique":"report-calibration-trajectory-checkpoints-split-horizon","statement":"For LLM agents whose steps condition on their own prior outputs, per-step confidence that is calibrated against step-local success labels does not compose into a calibrated estimate of whether the whole trajectory succeeds, because correlated step errors make the product of step marginals diverge from joint success probability, and the divergence grows as prefix reliability falls over longer horizons.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07395","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-agents-whose-steps-condition-their-own-prior-outputs"},{"ref":"CLM-0205","id":"llm-answers-question-gives-explanation-removing-input-concepts-explanation","capability":"explanation-faithfulness","technique":null,"statement":"When an LLM answers a question and gives an explanation, removing from the input the concepts its explanation did not credit and re-querying on the reduced input raises measured explanation faithfulness over both standard prompting and prompts that ask for faithfulness, because the removal deletes the unmentioned influences rather than asking the model to report them.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04343","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-answers-question-gives-explanation-removing-input-concepts-explanation"},{"ref":"CLM-0418","id":"llm-assisted-soc-pipeline-where-injections-planted-log-fields-deterministic","capability":"prompt-injection","technique":"deterministic-siem-pre-filter-before-semantic-guardrails","statement":"In an LLM-assisted SOC pipeline where injections are planted in log fields, a deterministic SIEM pre-filter (regex decoders plus length truncation) catches only the structurally obvious payloads while an LLM self-check input/output rail catches most semantic ones, so the two layers block disjoint attack classes and neither alone is sufficient.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10707","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-assisted-soc-pipeline-where-injections-planted-log-fields-deterministic"},{"ref":"CLM-0078","id":"llm-based-asr-steering-decoding-toward-acoustic-evidence-only-helps","capability":"hallucination","technique":null,"statement":"In LLM-based ASR, steering decoding toward acoustic evidence only helps if candidates are first restricted to tokens within a small likelihood margin of the greedy token: unconstrained selection by an acoustic compatibility score removes hallucinated outputs but destroys transcription accuracy through unbounded insertions, while the constrained version removes a large share of hallucinations at near-zero WER/CER cost.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30776","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-based-asr-steering-decoding-toward-acoustic-evidence-only-helps"},{"ref":"CLM-0419","id":"llm-based-forecasting-systems-reported-benchmark-gains-often-cannot-be","capability":"forecasting","technique":"score-forecasts-only-live-post-cutoff-questions-cost","statement":"For LLM-based forecasting systems, reported benchmark gains often cannot be attributed to temporal reasoning, because test periods predate the training cutoff, ablations show the LLM component sometimes adds nothing over the statistical or retrieval parts, and accuracy shifts under small input perturbations.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23058","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-based-forecasting-systems-reported-benchmark-gains-often-cannot-be"},{"ref":"CLM-0206","id":"llm-based-program-repair-defects4j-bugs-frontier-models-gpt-5-deepseek-r1","capability":"hallucination","technique":null,"statement":"In LLM-based program repair on Defects4J bugs with frontier models (GPT-5, DeepSeek-R1, Claude Sonnet 4.5) prompted zero-shot, passing the developer-written test suite does not mean the patch is grounded in the bug: manual inspection of 812 sampled repairs found ungrounded repairs in 72.7% of cases including test-passing patches, most often from incorrect causal localization (45.9%) rather than a wrong edit style, and the same models frequently misidentify the bug-triggering tests and generate additional tests that miss the triggering condition.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04909","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-based-program-repair-defects4j-bugs-frontier-models-gpt-5-deepseek-r1"},{"ref":"CLM-0589","id":"llm-based-repository-level-program-repair-telling-generator-be-minimal-or","capability":"cutting-to-the-chase","technique":"shrink-patch-after-generation-during","statement":"In LLM-based repository-level program repair, telling the generator to be minimal or constraining its output format does not shrink patches, because verbosity comes from the capability-raising machinery itself (iterative refinement, broad context); a separate post-generation refiner that rewrites an already-plausible patch is what reduces size without losing fixes.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-13292","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-based-repository-level-program-repair-telling-generator-be-minimal-or"},{"ref":"CLM-0590","id":"llm-data-preparation-pipelines-specialist-domains-math-finance-law-medicine","capability":"harness-self-improvement","technique":"evolve-data-generation-pipeline-pilot-sets-before-scaling","statement":"For LLM data-preparation pipelines in specialist domains (math, finance, law, medicine), diagnosing a small pilot set of generated samples against domain-selected quality criteria and rewriting the individual pipeline components (chunking, instruction builder, response generator) before full-scale generation yields training data with higher downstream fine-tuning utility than post-hoc filtering or one-shot agent/skill generators using the same backbone — but the gain is non-monotonic in rounds, peaking at two and degrading after.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29966","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-data-preparation-pipelines-specialist-domains-math-finance-law-medicine"},{"ref":"CLM-0207","id":"llm-driven-heuristic-discovery-combinatorial-optimization-tasks-tsp-cvrp-bi-objective","capability":"self-repair","technique":"self-harness-loop","statement":"For LLM-driven heuristic discovery on combinatorial optimization tasks (TSP, CVRP, bi-objective and multi-component variants), letting an agent decide when to read, edit, diagnose and evaluate code while persisting each evaluation outcome as a retrievable experience card reaches better final heuristic quality than fixed evolutionary pipelines (EoH, ReEvo, FunSearch, MCTS-AHD, MEoH, MOTIF) at a fraction of their evaluation budget, and re-running the same task warm-started from the previous run's best code improves further round after round under a fixed budget.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15820","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-driven-heuristic-discovery-combinatorial-optimization-tasks-tsp-cvrp-bi-objective"},{"ref":"CLM-0420","id":"llm-driven-iterative-prompt-optimization-restricting-gradient-generation-examples-current","capability":"harness-self-improvement","technique":"generate-prompt-edits-only-failures-gate-preservation-set","statement":"In LLM-driven iterative prompt optimization, restricting gradient generation to examples the current prompt gets wrong and accepting a candidate edit only when its gain on hard cases exceeds a penalized regression on a preservation set of initially-correct examples yields higher final accuracy than unfiltered textual-gradient optimization, which degrades on some reasoning tasks below the zero-shot prompt.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-23716","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-driven-iterative-prompt-optimization-restricting-gradient-generation-examples-current"},{"ref":"CLM-0421","id":"llm-driven-multi-uav-swarm-tasks-simulation-binding-one-agent-each","capability":"embodied-control","technique":"role-conditioned-per-vehicle-skill-subsets-peer-messaging","statement":"For LLM-driven multi-UAV swarm tasks in simulation, binding one agent to each vehicle with a role-conditioned subset of the skill catalog and directed peer messages yields higher episode returns than a central planner that sees pooled team state and the full skill list, and ablations attribute most of the gain to the skill-subset restriction and the peer reports rather than to the experience-based reward correction.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-18520","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-driven-multi-uav-swarm-tasks-simulation-binding-one-agent-each"},{"ref":"CLM-0591","id":"llm-driven-robot-manipulation-where-every-tool-call-schema-validated-every","capability":"embodied-control","technique":"schema-validated-tool-calls-plus-verify-then-act","statement":"For LLM-driven robot manipulation where every tool call is schema-validated and every motion kinematically and collision-checked before execution, execution-time recovery mechanisms (retrying a failed plan with an alternative candidate, resetting to a collision-checked safe pose between stages) contribute more to end-to-end task success than canonical action sequencing, because most residual failures are transient planning or IK failures rather than wrong orderings.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29379","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-driven-robot-manipulation-where-every-tool-call-schema-validated-every"},{"ref":"CLM-0422","id":"llm-fact-checking-over-supplied-evidence-high-veracity-accuracy-does","capability":"fact-verification","technique":"supervise-abstention-evidence-ablated-inputs","statement":"In LLM fact-checking over supplied evidence, high veracity accuracy does not imply the verdict depends on that evidence: when the model's own cited sentences are iteratively removed, off-the-shelf models (GPT-4o-mini, Gemini-2.5-Flash-Lite, Qwen-2.5-32B, Llama-3.1-8B) keep their SUPPORT/REFUTE labels at a non-zero plateau instead of abstaining, and training with paired supervision that requires NOT_ENOUGH_INFO on the evidence-ablated input removes most of that persistence without losing accuracy.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08943","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-fact-checking-over-supplied-evidence-high-veracity-accuracy-does"},{"ref":"CLM-0592","id":"llm-generated-hls-kernels-targeted-real-fpga-boards-two-stages","capability":"code-generation","technique":"evolutionary-kernel-search-plus-c-level-functional-repair","statement":"For LLM-generated HLS kernels targeted at real FPGA boards, the two stages of failure respond to different interventions: an evolutionary search over simulation-validated kernel candidates mainly raises synthesizable and executable rates, while a C/C++-level functional verify-and-repair loop mainly raises functional correctness and barely moves synthesizability, so both are needed for end-to-end success.","kind":"mechanism","backing_strength":"replicated","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23630","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-generated-hls-kernels-targeted-real-fpga-boards-two-stages"},{"ref":"CLM-0423","id":"llm-generated-hls-kernels-targeting-real-fpga-boards-simulation-synthesis","capability":"code-generation","technique":"validate-hls-generation-board-execution-simulation","statement":"For LLM-generated HLS kernels targeting real FPGA boards, simulation and synthesis success overstates deployability, and the two stages are fixed by different interventions: an evolutionary search over simulated candidates raises synthesizable and executable rates while barely raising functional correctness, whereas C/C++-level verify-and-repair raises functional correctness while leaving synthesizability flat.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23630","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-generated-hls-kernels-targeting-real-fpga-boards-simulation-synthesis"},{"ref":"CLM-0593","id":"llm-generated-multi-day-travel-itineraries-high-qualitative-coherence-scores-spatial","capability":"planning","technique":"score-itineraries-buffer-adequacy-against-empirical-delay-data","statement":"For LLM-generated multi-day travel itineraries, high qualitative coherence scores (spatial, persona, meal, ordering) coexist with near-zero robustness to realistic disruption: across GPT-5, Qwen3, Mistral-7B and Phi-4-mini, transport-delay buffer scores stay under 12% and buffer-adequacy scores under ~41%, so coherence and static constraint satisfaction do not indicate that a plan survives empirical transit delays or crowd peaks.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02421","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-generated-multi-day-travel-itineraries-high-qualitative-coherence-scores-spatial"},{"ref":"CLM-0594","id":"llm-generated-peer-reviews-compared-human-reviews-point-point-under","capability":"critique","technique":"profile-review-critique-function-human-likeness","statement":"When LLM-generated peer reviews are compared with human reviews point by point under theory-guided critique frameworks, the model and human critique differ in kind rather than in quality: LLM reviews carry more explanatory depth, integrative reasoning and explicit argument structure, while human reviews more often identify higher-order weaknesses and give revision-oriented guidance, and an expert prompt does not close that gap but amplifies the LLM-side tendencies.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01895","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-generated-peer-reviews-compared-human-reviews-point-point-under"},{"ref":"CLM-0595","id":"llm-generated-spear-phishing-emails-built-cumulatively-deeper-workplace-detail","capability":"persuasion","technique":"vary-personalization-depth-cumulative-workplace-detail-tiers","statement":"When LLM-generated spear phishing emails are built from cumulatively deeper workplace detail (employer, then name and title, then responsibilities, then coworker and shared project), U.S. working adults rate them as more convincing and more often say they would click, and they also report the message as phishing less often and delete it more often as detail increases.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04410","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-generated-spear-phishing-emails-built-cumulatively-deeper-workplace-detail"},{"ref":"CLM-0240","id":"llm-gives-answer-plus-explanation-deleting-input-those-concepts","capability":"explanation-faithfulness","technique":"re-query-after-removing-concepts-explanation-ignored","statement":"When an LLM gives an answer plus an explanation, deleting from the input those concepts the explanation does not credit and re-querying on the reduced input raises measured faithfulness over both standard prompting and prompts that ask for faithful explanations, because the re-query removes the unmentioned factors that could have influenced the original answer.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04343","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-gives-answer-plus-explanation-deleting-input-those-concepts"},{"ref":"CLM-0424","id":"llm-guided-narrow-space-parking-where-optimal-control-solver-supplies-pass-fail","capability":"harness-self-improvement","technique":"evolve-decision-knowledge-base-solver-failures-gated-feasibility","statement":"For LLM-guided narrow-space parking where an optimal-control solver supplies pass/fail evidence, growing a decision knowledge base offline from accumulated solver and validation failures — retaining an entry only when re-planning then yields a feasible trajectory — raises planning success from 85.1% to 94.7% and closed-loop success from 74.6% to 84.2% over an empty knowledge base, with model weights frozen.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24631","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-guided-narrow-space-parking-where-optimal-control-solver-supplies-pass-fail"},{"ref":"CLM-0596","id":"llm-judge-given-prior-evaluation-score-context-metadata-revision","capability":"critique","technique":"strip-prior-score-metadata-judge-prompts","statement":"When an LLM judge is given a prior evaluation score as context metadata on a revision, its rubric scores shift toward that uninformative prior — seven of eight tested models (API and open-weight, 0-5 scale, 20 fixed texts) show a negative total shift, and on a categorical compliance task the metadata blocks 48% of error corrections and flips about 10% of previously correct judgments to the anchored wrong label.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25869","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-judge-given-prior-evaluation-score-context-metadata-revision"},{"ref":"CLM-0597","id":"llm-judge-must-pick-human-preferred-response-out-more-than","capability":"abstention-and-self-knowledge","technique":"conformal-shortlist-first-then-calibrated-margin-abstention","statement":"When an LLM judge must pick the human-preferred response out of more than two candidates, the assumption behind confidence-threshold abstention — that higher estimated confidence means lower disagreement with humans — degrades as the candidate count grows, because probability mass is normalized across more options; restricting the decision to a conformal top-k shortlist and thresholding a margin inside it restores the monotone relation.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25824","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-judge-must-pick-human-preferred-response-out-more-than"},{"ref":"CLM-0079","id":"llm-judge-validated-against-expert-annotators-multi-dimensional-rubrics-open-ended","capability":"evaluation-validity","technique":null,"statement":"When an LLM judge is validated against expert annotators on multi-dimensional rubrics for open-ended agent outputs, a single pooled agreement figure overstates reliability: per-dimension human-human agreement varies widely (substantial on a recall-style Completeness dimension, only slight to fair on Relevancy, Structural Clarity and Scope Adherence), the pooled statistic is inflated by between-dimension variation, and the judge's residual disagreements are directional (consistently one step stricter on Completeness, one step more generous on Structural Clarity) rather than noise.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-21057","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-judge-validated-against-expert-annotators-multi-dimensional-rubrics-open-ended"},{"ref":"CLM-0598","id":"llm-multi-agent-systems-failing-long-horizon-web-tasks-re-running-whole","capability":"multi-model-coordination","technique":"replay-recorded-prefix-then-intervene-symptomatic-node","statement":"For LLM multi-agent systems failing on long-horizon web tasks, re-running the whole task from scratch — plain, with self-reflection, or with a critic agent — repairs few failures (at most 6.90% within three attempts, with feedback variants below plain rerun), while reconstructing the recorded execution prefix and intervening once at a symptom-flagged trace node repairs 20.15% with a single attempt.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25920","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-multi-agent-systems-failing-long-horizon-web-tasks-re-running-whole"},{"ref":"CLM-0425","id":"llm-multi-agent-systems-reflect-after-failure-restricting-reflection-single","capability":"multi-model-coordination","technique":"reflect-only-agent-made-first-wrong-step","statement":"In LLM multi-agent systems that reflect after failure, restricting reflection to the single agent responsible for the first incorrect step — located by a trained step-level correctness scorer and paired with a counterfactually generated corrected step — raises task success more than having every agent reflect on the trajectory, because reflections written by agents that behaved correctly enter their memory as misleading guidance for later tasks.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28264","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-multi-agent-systems-reflect-after-failure-restricting-reflection-single"},{"ref":"CLM-0599","id":"llm-multi-agent-systems-whose-communication-graph-produced-learned-topology","capability":"multi-model-coordination","technique":"rank-communication-edges-masking-effect-then-prune-budget","statement":"In LLM multi-agent systems whose communication graph is produced by a learned topology generator, the generated graph carries edges that are not needed: scoring each edge by how much masking it changes the task outcome and the semantic entropy of the final answer, then keeping a top-ranked budgeted subgraph, cuts online tokens by about 20-25% while leaving average accuracy equal or slightly higher, and a small distilled predictor of those subgraphs transfers to generators and agent counts it was not calibrated on.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-12921","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-multi-agent-systems-whose-communication-graph-produced-learned-topology"},{"ref":"CLM-0426","id":"llm-program-repair-workflows-built-self-correction-generated-knowledge-or-multi-agent-scaffolding","capability":"self-repair","technique":"external-feedback-repair","statement":"For LLM program-repair workflows built on self-correction, generated-knowledge or multi-agent scaffolding, the gain over plain prompting is unstable across benchmarks and can turn negative once extrinsic defect information (error messages, triggering tests, bug comments) is withheld, while simply putting the error message plus the failing test into the prompt yields more correct repairs than the scaffolds do.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22844","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-program-repair-workflows-built-self-correction-generated-knowledge-or-multi-agent-scaffolding"},{"ref":"CLM-0427","id":"llm-web-agents-cluttered-pages-containing-many-concurrent-interface","capability":"tool-use","technique":"decompose-observation-into-sub-goal-sub-view-before","statement":"For LLM web agents on cluttered pages containing many concurrent interface components, success falls sharply as component count grows, and having the agent first emit a localized sub-goal plus a task-relevant sub-observation before choosing an action recovers much of the loss, because self-assessment of the action is more reliable in the reduced context.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29305","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-web-agents-cluttered-pages-containing-many-concurrent-interface"},{"ref":"CLM-0310","id":"llm-written-advertisements-matched-human-expert-ads-personalized-personality-traits","capability":"persuasion","technique":null,"statement":"LLM-written advertisements matched human expert ads when personalized to personality traits (51% vs 49% preference, n=400) and beat them when built on general persuasion principles such as authority and consensus (59% preference, n=800); participants who recognized an ad as AI-made penalized it only slightly, and 29% chose the AI ad knowing its origin.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2512-03373","stance":"supports"}],"url":"https://rsiratchet.com/claims/llm-written-advertisements-matched-human-expert-ads-personalized-personality-traits"},{"ref":"CLM-0600","id":"llms-used-relevance-judges-specialized-non-english-brazilian-portuguese-legal","capability":"critique","technique":null,"statement":"When LLMs are used as relevance judges on a specialized non-English (Brazilian Portuguese legal) retrieval collection, they assign systematically higher relevance grades than human experts (mean absolute error 0.46-0.66 on a 0-2 scale, Cohen's kappa with humans only 0.32-0.53), yet the resulting system rankings stay close to the human reference for rank-aware metrics (nDCG@10, MRR, Kendall's tau >= 0.90) and diverge for P@10 and R@10.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27746","stance":"supports"}],"url":"https://rsiratchet.com/claims/llms-used-relevance-judges-specialized-non-english-brazilian-portuguese-legal"},{"ref":"CLM-0080","id":"locally-deployed-open-weight-models-driving-stateful-dependency-ordered-mcp-tool","capability":"tool-use","technique":"retrieve-tool-docs","statement":"For locally deployed open-weight models driving a stateful, dependency-ordered MCP tool server, cutting tool descriptions from full specifications (purpose, parameter semantics, constraints, failure conditions) down to one sentence each raises the fraction of calls the server rejects for every model tested, while its effect on task coverage is less consistent.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26199","stance":"supports"}],"url":"https://rsiratchet.com/claims/locally-deployed-open-weight-models-driving-stateful-dependency-ordered-mcp-tool"},{"ref":"CLM-0428","id":"long-chain-thought-verified-classifying-graph-steps-predictive-signal","capability":"multi-step-reasoning","technique":"build-reasoning-graph-local-nli-model-then-classify","statement":"When a long chain of thought is verified by classifying a graph of its steps, the predictive signal comes almost entirely from the semantic content of the steps, not the graph shape: replacing step embeddings with purely structural metadata (child/parent counts, chain index, position) drops accuracy to near chance on math, science and multiple-choice reasoning, while remaining above chance only on code.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30679","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-chain-thought-verified-classifying-graph-steps-predictive-signal"},{"ref":"CLM-0081","id":"long-context-accuracy-drops-in-the-middle-gpt-3-5-turbo","capability":"long-context-degradation","technique":null,"statement":"On multi-document question answering with GPT-3.5 Turbo's 16k-context variant, accuracy dropped by more than twenty points when the document containing the answer was moved from the start or end of the context to the middle, with nothing else about the task changed.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2307-03172","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-context-accuracy-drops-in-the-middle-gpt-3-5-turbo"},{"ref":"CLM-0429","id":"long-context-tasks-over-8k-128k-token-documents-middle-of-context-penalty-granularity-dependent-word-level","capability":"long-context-degradation","technique":null,"statement":"In long-context tasks over 8K-128K-token documents, the middle-of-context penalty is granularity-dependent: word-level QA and paragraph-filling accuracy dip when the evidence sits mid-document, but sentence-cloze accuracy peaks in the middle, because mid-paragraph sentences carry concrete entity and date overlap that models match on instead of reasoning about discourse role.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-20853","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-context-tasks-over-8k-128k-token-documents-middle-of-context-penalty-granularity-dependent-word-level"},{"ref":"CLM-0601","id":"long-context-tasks-over-8k-128k-token-documents-six-official-un-languages","capability":"long-context-degradation","technique":null,"statement":"On long-context tasks over 8K-128K-token documents in the six official UN languages, models answer word-level span questions well (about 73% average accuracy across 12 models) but degrade sharply as the required unit coarsens to sentence selection (about 21%), paragraph filling and document-level summarization and translation, so long-context competence measured by needle-style word retrieval overstates comprehension of larger discourse units.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-20853","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-context-tasks-over-8k-128k-token-documents-six-official-un-languages"},{"ref":"CLM-0208","id":"long-cooking-videos-where-question-depends-ingredient-changes-no","capability":"state-tracking","technique":null,"statement":"For long cooking videos where a question depends on ingredient changes no longer visible, vision-language models fail to carry observed events forward into later answers, and adding generic retrieved-context RAG makes it worse, while splitting memory into a temporally anchored event log plus an explicitly updated inventory ledger recovers much of the gap.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07135","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-cooking-videos-where-question-depends-ingredient-changes-no"},{"ref":"CLM-0082","id":"long-form-biographies-from-strong-models-contain-substantial-share","capability":"hallucination","technique":null,"statement":"Long-form biographies from strong models contain a substantial share of unsupported atomic facts, rising for less well-known subjects.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2305-14251","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-form-biographies-from-strong-models-contain-substantial-share"},{"ref":"CLM-0430","id":"long-horizon-embodied-task-planning-where-symbolic-model-defines-action","capability":"planning","technique":"repair-plans-world-model-ask-llm-only-semantic","statement":"For long-horizon embodied task planning where a symbolic model defines action preconditions and effects, correcting a violated plan by edits the action model itself implies (inserting the missing navigate or open) raises plan success more per LLM call than returning the violation to the LLM for replanning, and the gain persists for frontier models as well as compact ones.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-19315","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-horizon-embodied-task-planning-where-symbolic-model-defines-action"},{"ref":"CLM-0431","id":"long-horizon-llm-agents-environments-where-erroneous-action-changes-world","capability":"long-horizon-execution","technique":"roll-back-environment-keep-distilled-reflection","statement":"For long-horizon LLM agents in environments where an erroneous action changes world state, restoring an earlier checkpoint and carrying forward a structured reflection memory that excludes branch-local state claims beats either intervention alone, because reflection alone cannot undo environmental damage and rollback alone re-explores the same failure.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-18304","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-horizon-llm-agents-environments-where-erroneous-action-changes-world"},{"ref":"CLM-0602","id":"long-horizon-llm-agents-tool-use-terminal-benchmarks-token-level-confidence-predictive","capability":"tool-use","technique":"propagate-uncertainty-over-trajectory-dependency-graph","statement":"For long-horizon LLM agents on tool-use and terminal benchmarks, token-level confidence (predictive entropy, sequence probability) barely separates successful from failed trajectories, and the separable signal lives in the dependency structure of the trajectory: scoring a trajectory by propagating per-step uncertainty over a graph of reasoning, tool-call and environment-feedback nodes raises failure-detection AUROC over both local-confidence and sequence-based agent UQ baselines.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-16002","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-horizon-llm-agents-tool-use-terminal-benchmarks-token-level-confidence-predictive"},{"ref":"CLM-0432","id":"long-horizon-repository-level-coding-agents-small-4b-critic-trained-diagnose","capability":"critique","technique":"interrupt-agent-step-level-critic-flags-critical-risk","statement":"For long-horizon repository-level coding agents, a small (4B) critic trained to diagnose each just-completed step with a severity label, error type and causal rationale, and to redirect the next action when severity is critical, repairs more initially-failing tasks with a single branch (14.1% fail-to-pass) than five full resamples of the base agent (13.7%), at about one-fifth the base-agent output tokens.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-23808","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-horizon-repository-level-coding-agents-small-4b-critic-trained-diagnose"},{"ref":"CLM-0083","id":"long-horizon-tool-calling-agents-tasks-needing-five-or-more-calls","capability":"prompt-injection","technique":null,"statement":"For long-horizon tool-calling agents (tasks needing five or more calls), embedding state-transition cues — preconditions, invariants, completion states — into the descriptions of the tools on the intended chain is what actually steers the agent's trajectory; runtime corrective text appended to tool results only patches residual drift, and the plausible user prompt alone (persona, deadlines, format constraints) does not establish the trajectory at all.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30441","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-horizon-tool-calling-agents-tasks-needing-five-or-more-calls"},{"ref":"CLM-0603","id":"long-horizon-tool-mediated-environments-where-strategically-relevant-state-available-only","capability":"long-horizon-execution","technique":"measure-proactive-monitoring-reflection-action-gap-traces","statement":"In long-horizon tool-mediated environments where strategically relevant state is available only on explicit query (300+ turn Civilization VI via MCP, 23 runs, four model families under a shared playbook), agents allocate only about 1-2% of non-infrastructure tool calls to strategic monitoring and execute only about 48-66% of the near-term commitments they state in their own planning diaries, even though the playbook told them to query victory progress every 20 turns.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02459","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-horizon-tool-mediated-environments-where-strategically-relevant-state-available-only"},{"ref":"CLM-0433","id":"long-trusted-instructions-untrusted-retrieved-content-occupy-same-model","capability":"prompt-injection","technique":"structured-prompt-separation","statement":"As long as trusted instructions and untrusted retrieved content occupy the same model context, defenses that only change how untrusted text is marked or weighted (instruction hierarchies, spotlighting, content detectors) reduce indirect-injection success but do not remove it; only defenses that isolate contexts or restrict the action set give guarantees, at a cost in the agent's ability to combine information across sources.","kind":"mechanism","backing_strength":"replicated","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22712","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-trusted-instructions-untrusted-retrieved-content-occupy-same-model"},{"ref":"CLM-0434","id":"long-video-10-150-minute-audio-visual-reasoning-questions-multiple-choice-accuracy-overstates","capability":"video-temporal-reasoning","technique":"condition-reasoning-credit-spatio-temporal-evidence","statement":"On long-video (10-150 minute) audio-visual reasoning questions, multiple-choice accuracy overstates grounding: a large open model can answer about 55% of items correctly while scoring 8.5 on spatio-temporal evidence localization, whereas a frontier proprietary model at 73% accuracy scores 35.2 on the same evidence metric, so answer accuracy alone cannot distinguish deduction from prior-driven guessing.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17248","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-video-10-150-minute-audio-visual-reasoning-questions-multiple-choice-accuracy-overstates"},{"ref":"CLM-0209","id":"long-video-qa-fixed-frame-budget-fed-frozen-multimodal-model","capability":"video-temporal-reasoning","technique":null,"statement":"For long-video QA with a fixed frame budget fed to a frozen multimodal model, selecting frames by marginal contribution to the already-selected subset (query relevance plus coverage gain minus redundancy) beats ranking frames independently by query-frame similarity, and the margin over uniform sampling grows with video duration.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15408","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-video-qa-fixed-frame-budget-fed-frozen-multimodal-model"},{"ref":"CLM-0604","id":"long-video-qa-frozen-vlm-fixed-answering-interface-outer-loop","capability":"harness-self-improvement","technique":"recursively-search-video-context-construction-program","statement":"For long-video QA with a frozen VLM and fixed answering interface, an outer loop that proposes, executes and retains only development-accuracy-improving rewrites of the executable context-construction program can advance past a strong hand-crafted keyframe-selection harness, and the gain survives a matched cumulative visual-token control and direct reuse on other long-video benchmarks.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24302","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-video-qa-frozen-vlm-fixed-answering-interface-outer-loop"},{"ref":"CLM-0605","id":"long-video-qa-under-fixed-32-frame-budget-7b-8b-multimodal-backbones","capability":"video-temporal-reasoning","technique":"select-frames-shape-relevance-curve-height","statement":"For long-video QA under a fixed 32-frame budget with 7B-8B multimodal backbones, selecting frames from the structural shape of the frame-query similarity curve (peak, rising, falling, transition regions) beats ranking frames by similarity magnitude, and the margin grows with video length — a few points on short clips and roughly 8 points on hour-long videos — while on short/medium-clip benchmarks it is near-indistinguishable from other selectors.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29958","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-video-qa-under-fixed-32-frame-budget-7b-8b-multimodal-backbones"},{"ref":"CLM-0606","id":"long-video-qa-under-fixed-frame-budget-frozen-7b-multimodal","capability":"video-temporal-reasoning","technique":"search-hierarchical-temporal-tree-video-segments-coarse-to","statement":"For long-video QA under a fixed frame budget with a frozen 7B multimodal model, organizing frames into a question-agnostic hierarchical temporal tree and searching it coarse-to-fine (answer, then issue a search statement, then expand the relevant branch) raises accuracy over uniform sampling and over agent methods that score frames independently, with the gain largest on the longest-video benchmarks.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27871","stance":"supports"}],"url":"https://rsiratchet.com/claims/long-video-qa-under-fixed-frame-budget-frozen-7b-multimodal"},{"ref":"CLM-0607","id":"lora-ensemble-credal-commitment-rules-7-9b-instruct-backbones-decision-commit","capability":"abstention-and-self-knowledge","technique":"abstain-unless-one-answer-dominates-lora-ensemble-s","statement":"For LoRA-ensemble credal commitment rules on 7-9B instruct backbones, the decision to commit tracks agreement among the adapters rather than whether a request should be answered: on AdvBench harmful prompts the ensembles commit more often (33-97%) and with higher lower-bound confidence than on benign Dolly instructions, because safety training makes refusal the unanimous first continuation.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23244","stance":"supports"}],"url":"https://rsiratchet.com/claims/lora-ensemble-credal-commitment-rules-7-9b-instruct-backbones-decision-commit"},{"ref":"CLM-0435","id":"lvlms-asked-describe-people-llava-instructblip-shikra-masking-attention","capability":"visual-grounding","technique":"mask-language-only-self-attention-keep-vision-to","statement":"In LVLMs asked to describe people (LLaVA, InstructBLIP, Shikra), masking attention so that text tokens read visual tokens but language-only self-attention is suppressed in late decoder layers lowers gender misattribution while keeping answer coverage high, whereas visual-only or language-only masking scores well on the bias metric only because the model names a gender in under 10% of cases.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-16647","stance":"supports"}],"url":"https://rsiratchet.com/claims/lvlms-asked-describe-people-llava-instructblip-shikra-masking-attention"},{"ref":"CLM-0519","id":"machine-learning-research-leaves-negative-results-unpublished","capability":"evaluation-validity","technique":null,"statement":"Machine learning research treats beating the state of the art on a benchmark as the main kind of publishable result, so negative results — methods that did not work, conditions under which a known method fails — go largely unpublished, and compute is spent repeating attempts that others have already found not to work; registering what does not work is no less important than registering what works.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2406-03980","stance":"supports"}],"url":"https://rsiratchet.com/claims/machine-learning-research-leaves-negative-results-unpublished"},{"ref":"CLM-0084","id":"managing-context-explicitly-paging-information-out-bounded-window","capability":"context-hygiene","technique":null,"statement":"Managing context explicitly, paging information in and out of a bounded window, sustains performance on tasks that exceed the window.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2310-08560","stance":"supports"}],"url":"https://rsiratchet.com/claims/managing-context-explicitly-paging-information-out-bounded-window"},{"ref":"CLM-0608","id":"manipulation-policies-diffusion-policy-vla-backbones-trained-few-demonstrations-decomposing","capability":"embodied-control","technique":"store-moe-gating-profiles-retrievable-skill-memory","statement":"For manipulation policies (diffusion-policy and VLA backbones) trained on few demonstrations, decomposing trajectories implicitly via a Mixture-of-Experts gate and storing the per-segment gating profiles as retrievable memory, then blending retrieved profiles into the current gating distribution at inference, raises success most on unseen object-receptacle recombinations rather than on in-distribution tasks.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-05970","stance":"supports"}],"url":"https://rsiratchet.com/claims/manipulation-policies-diffusion-policy-vla-backbones-trained-few-demonstrations-decomposing"},{"ref":"CLM-0436","id":"manipulation-policies-pretrained-mixed-egocentric-human-robot-data-adding-egocentric","capability":"embodied-control","technique":"stage-ego-to-robot-transfer-through-aligned-demo","statement":"For manipulation policies pretrained on mixed egocentric-human and robot data, adding egocentric data helps only when the embodiment gap is bridged explicitly: a three-stage curriculum (ego priors, then task-matched human/robot alignment in a shared end-effector frame, then robot-only) raised real-robot success, while separate ego/robot action heads gave only a marginal in-distribution gain and shared future-video prediction in a world-action model made success worse than a robot-only baseline.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21461","stance":"supports"}],"url":"https://rsiratchet.com/claims/manipulation-policies-pretrained-mixed-egocentric-human-robot-data-adding-egocentric"},{"ref":"CLM-0437","id":"manipulation-policies-whose-large-vla-backbone-cannot-run-within","capability":"embodied-control","technique":"slow-chunk-proposal-plus-fast-rl-trained-action","statement":"For manipulation policies whose large VLA backbone cannot run within the control period, letting the slow model propose action chunks from a stale observation while a small RL-trained edit policy adjusts and Q-selects among those chunks using the latest observation recovers more performance than applying RL on top of asynchronous imitation-learning schemes alone, and the gap widens as inference delay or environment speed grows.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-18207","stance":"supports"}],"url":"https://rsiratchet.com/claims/manipulation-policies-whose-large-vla-backbone-cannot-run-within"},{"ref":"CLM-0085","id":"many-failures-come-from-not-following-domain-policy","capability":"procedure-following","technique":null,"statement":"Many failures come from not following domain policy documents that were provided in the prompt.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2406-12045","stance":"supports"}],"url":"https://rsiratchet.com/claims/many-failures-come-from-not-following-domain-policy"},{"ref":"CLM-0086","id":"marine-vehicle-navigation-llm-asked-only-emit-ordered-macro-action","capability":"embodied-control","technique":null,"statement":"For marine vehicle navigation, an LLM asked only to emit an ordered macro-action skeleton produces syntactically and topologically usable plans but cannot set physically feasible step durations; letting a differentiable physics surrogate optimize the durations turns plans that collide on every mission into collision-free ones, while wrong skeleton topology remains unrecoverable because the optimizer only rescales timings.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19661","stance":"supports"}],"url":"https://rsiratchet.com/claims/marine-vehicle-navigation-llm-asked-only-emit-ordered-macro-action"},{"ref":"CLM-0210","id":"masked-diffusion-language-models-llada-dream-asked-generate-around","capability":"procedure-following","technique":null,"statement":"In masked diffusion language models (LLaDA, Dream) asked to generate around a frozen constraint span placed mid-output, failures to honor the constraint come partly from too little attention being allocated to that span during denoising: adding a fixed additive bias to attention logits on the span raises constraint adherence and overall quality up to a moderate bias, then degrades fluency and contextual fit as the bias grows.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07160","stance":"supports"}],"url":"https://rsiratchet.com/claims/masked-diffusion-language-models-llada-dream-asked-generate-around"},{"ref":"CLM-0087","id":"measured-rd-uplift-is-below-the-self-sustaining-threshold","capability":"evaluation-validity","technique":null,"statement":"On the best current calibration, AI is not yet accelerating its own development in a self-sustaining way: the modelled threshold is that a one-unit gain in model capability must buy at least 15% higher AI R&D productivity, and the back-of-envelope figure from reported engineer uplift since coding agents launched is about 9% — below it, but rising, so the gap is a current reading rather than a ceiling.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"elasticity-2026-economics-of-rsi","stance":"supports"}],"url":"https://rsiratchet.com/claims/measured-rd-uplift-is-below-the-self-sustaining-threshold"},{"ref":"CLM-0609","id":"medical-llms-given-clinical-practice-guideline-context-asked-produce","capability":"domain-knowledge","technique":"score-decision-pathway-just-outcome","statement":"When medical LLMs are given a clinical practice guideline in context and asked to produce the decision pathway a case requires, final-outcome accuracy does not track adherence to the guideline-defined step sequence: models can reach the right outcome while omitting, reordering, skipping or padding steps, and medically fine-tuned models score no better on pathway consistency than the general models they were tuned from.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26592","stance":"supports"}],"url":"https://rsiratchet.com/claims/medical-llms-given-clinical-practice-guideline-context-asked-produce"},{"ref":"CLM-0438","id":"medical-multiple-choice-qa-multi-agent-pipeline-mining-natural-language-lessons-own","capability":"harness-self-improvement","technique":"mine-tiered-lessons-own-trajectories-principles-first","statement":"For medical multiple-choice QA with a multi-agent pipeline, mining natural-language lessons from its own past trajectories (both correct and failed) into a reusable library gives a larger accuracy gain than the multi-agent role decomposition itself, and the gain comes mostly from the most abstract tier: removing principle-level entries costs more than removing pattern- or case-level entries.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15161","stance":"supports"}],"url":"https://rsiratchet.com/claims/medical-multiple-choice-qa-multi-agent-pipeline-mining-natural-language-lessons-own"},{"ref":"CLM-0439","id":"medically-fine-tuned-llama-3-1-8b-checkpoint-ultramedical-scores-below-general-purpose","capability":"domain-knowledge","technique":"reweight-jargon-favoring-components-post-hoc","statement":"A medically fine-tuned Llama-3.1-8B checkpoint (UltraMedical) scores below its general-purpose Instruct counterpart on both selecting the correct definition of medical jargon in EHR context and judging whether a term is jargon, and its errors are almost entirely false positives — it treats ordinary words in clinical text as specialized.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-13556","stance":"supports"}],"url":"https://rsiratchet.com/claims/medically-fine-tuned-llama-3-1-8b-checkpoint-ultramedical-scores-below-general-purpose"},{"ref":"CLM-0440","id":"memory-augmented-vla-manipulation-policies-must-actively-move-viewpoint-find","capability":"embodied-control","technique":"write-memory-frames-fixed-chunk-schedule-events","statement":"For memory-augmented VLA manipulation policies that must actively move the viewpoint to find task-relevant evidence, writing one memory frame at each fixed action-chunk boundary generalizes better than event-triggered writing (on predicted subtask change or on gripper/head/torso motion), and extra memory capacity helps only under the regular schedule — event-triggered policies degrade further out of distribution because mistimed triggers omit informative views and control jitter fills memory with redundant frames.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24124","stance":"supports"}],"url":"https://rsiratchet.com/claims/memory-augmented-vla-manipulation-policies-must-actively-move-viewpoint-find"},{"ref":"CLM-0088","id":"memory-stream-retrieval-reflection-lets-agents-behave-consistently","capability":"long-term-memory","technique":null,"statement":"A memory stream with retrieval and reflection lets agents behave consistently over simulated days.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2304-03442","stance":"supports"}],"url":"https://rsiratchet.com/claims/memory-stream-retrieval-reflection-lets-agents-behave-consistently"},{"ref":"CLM-0211","id":"mistaken-user-pushes-back-adaptively-up-25-turns-frontier","capability":"sycophancy","technique":null,"statement":"When a mistaken user pushes back adaptively for up to 25 turns, frontier and 7B open models keep abandoning initially correct positions at higher rates the longer the conversation runs, so sycophancy measured with five-turn or fixed-script rebuttals understates it — in one ablation collapse on false-presupposition items rose from 50% by turn 5 to 92% by turn 25 for the same model, and a fixed four-follow-up script reached only 28%.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09090","stance":"supports"}],"url":"https://rsiratchet.com/claims/mistaken-user-pushes-back-adaptively-up-25-turns-frontier"},{"ref":"CLM-0441","id":"model-asked-audit-many-documents-planted-errors-single-large-batch","capability":"hallucination","technique":"bound-audit-batch-size-verify-every-reported-finding","statement":"When a model is asked to audit many documents for planted errors in a single large-batch prompt, recall of known contaminants collapses and the model reports confident fabricated findings — invented absurd contaminants that appear in no document — rather than signalling incomplete processing.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09696","stance":"supports"}],"url":"https://rsiratchet.com/claims/model-asked-audit-many-documents-planted-errors-single-large-batch"},{"ref":"CLM-0442","id":"model-asked-predict-own-rates-behavioral-evaluations-sycophancy-reward","capability":"abstention-and-self-knowledge","technique":"ask-about-generic-ai-agent-model-itself","statement":"When a model is asked to predict its own rates on behavioral evaluations (sycophancy, reward hacking, agentic policy violation, bias), its answers track its own behavior no better than the same question asked about 'capable AI agents in general' or than other models' answers about themselves, so self-report reflects a generic theory of AI assistants rather than privileged self-knowledge, and first-person framing mainly adds a self-flattering shift that understates harmful behavior.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09899","stance":"supports"}],"url":"https://rsiratchet.com/claims/model-asked-predict-own-rates-behavioral-evaluations-sycophancy-reward"},{"ref":"CLM-0443","id":"model-must-anticipate-without-execution-feedback-what-change-own","capability":"harness-self-improvement","technique":"compute-lifecycle-consequences-harness-model","statement":"When a model must anticipate, without execution feedback, what a change to its own runtime harness components will do, naming which components a lifecycle change can affect stays near ceiling as the number of interacting effects grows from 2 to 32, while predicting the resulting state, reasoning across teardown orders, and choosing minimal reconfigurations degrade sharply (e.g. GPT-5.6 Luna's reachable-condition Jaccard 91.7% to 14.1%, executed reconfiguration success 62.5% to 25.0%).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01600","stance":"supports"}],"url":"https://rsiratchet.com/claims/model-must-anticipate-without-execution-feedback-what-change-own"},{"ref":"CLM-0241","id":"model-s-tendency-shift-political-stance-toward-bare-demographic","capability":"sycophancy","technique":"two-axis-identity-opinion-sycophancy-probe","statement":"A model's tendency to shift political stance toward a bare demographic identity label is a separate vulnerability from its tendency to shift toward an explicitly stated user opinion: on open-ended US policy prompts, the models most moved by an identity label are among the least moved by a stated opinion, so a benchmark using only stated opinions misses identity-driven shift.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29198","stance":"supports"}],"url":"https://rsiratchet.com/claims/model-s-tendency-shift-political-stance-toward-bare-demographic"},{"ref":"CLM-0089","id":"models-complied-insecure-completions-large-fraction-time-more","capability":"secure-coding","technique":null,"statement":"Models complied with insecure completions a large fraction of the time, and more capable models were more likely to suggest insecure code.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2312-04724","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-complied-insecure-completions-large-fraction-time-more"},{"ref":"CLM-0090","id":"models-endorse-widely-held-falsehoods-showing-weak-verification","capability":"fact-verification","technique":null,"statement":"Models endorse widely held falsehoods, showing weak verification against what they know.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2109-07958","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-endorse-widely-held-falsehoods-showing-weak-verification"},{"ref":"CLM-0091","id":"models-fine-tuned-b-fail-answer-b-gpt-4-recalls","capability":"reversal-curse","technique":null,"statement":"Models fine-tuned on \"A is B\" fail to answer \"B is A\", and GPT-4 recalls celebrity parents far more often than the reverse.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2309-12288","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-fine-tuned-b-fail-answer-b-gpt-4-recalls"},{"ref":"CLM-0311","id":"models-judge-syllogism-s-validity-partly-how-believable-conclusion","capability":"logical-deduction-robustness","technique":"debiasing-vector-separates-validity-plausibility","statement":"Models judge a syllogism's validity partly by how believable its conclusion is because validity and plausibility are encoded as linearly separable but strongly aligned directions in the residual stream: steering along the plausibility direction causally shifts validity judgements and vice versa, the degree of alignment predicts the size of the content effect across models, and a debiasing vector that separates the two reduces the effect and raises accuracy.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2510-06700","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-judge-syllogism-s-validity-partly-how-believable-conclusion"},{"ref":"CLM-0092","id":"models-must-apply-abstract-entity-independent-procedural-rules-drawn-shared","capability":"procedure-following","technique":null,"statement":"When models must apply abstract, entity-independent procedural rules drawn from a shared pool, accuracy is limited by locating the right rules rather than by applying them: giving only the gold rules yields much higher exact-match than embedding those same rules in a pool of 100 or 1,000, and accuracy falls as the pool grows.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22753","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-must-apply-abstract-entity-independent-procedural-rules-drawn-shared"},{"ref":"CLM-0093","id":"models-predict-next-steps-well-can-still-hold","capability":"state-tracking","technique":null,"statement":"Models that predict next steps well can still hold an incoherent implicit world model, which fails when the task deviates from familiar traces.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2406-03689","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-predict-next-steps-well-can-still-hold"},{"ref":"CLM-0094","id":"models-readily-adopt-single-counter-memory-passage-when-coherent","capability":"fact-verification","technique":null,"statement":"Models readily adopt a single counter-memory passage when it is coherent, and when sources conflict they follow the majority and show confirmation bias toward their own beliefs.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2305-13300","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-readily-adopt-single-counter-memory-passage-when-coherent"},{"ref":"CLM-0095","id":"models-reproduce-common-human-misconceptions-larger-models-were","capability":"hallucination","technique":null,"statement":"Models reproduce common human misconceptions, and larger models were not more truthful on this benchmark.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2109-07958","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-reproduce-common-human-misconceptions-larger-models-were"},{"ref":"CLM-0096","id":"models-tendency-shift-political-stance-toward-bare-demographic-identity","capability":"sycophancy","technique":null,"statement":"A model's tendency to shift political stance toward a bare demographic identity label is a separate vulnerability from its tendency to shift toward an explicitly stated user opinion: on open-ended US policy prompts, the models most moved by an identity label are among the least moved by a stated opinion, so a benchmark using only stated opinions misses identity-driven shift.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29198","stance":"supports"}],"url":"https://rsiratchet.com/claims/models-tendency-shift-political-stance-toward-bare-demographic-identity"},{"ref":"CLM-0444","id":"monocular-vlm-based-navigation-policies-making-model-first-emit-two-token","capability":"embodied-control","technique":"point-first-then-emit-quantized-trajectory-tokens","statement":"For monocular VLM-based navigation policies, making the model first emit a two-token pointing prefix (one affordance point for free space, one object point for the goal) on an image-plane grid and then three residual vector-quantized action tokens decoded to 10 SE(2) waypoints lets one compact 4B backbone handle instruction following, object-goal search and target tracking without task-specific action heads or task identifier tokens.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30935","stance":"supports"}],"url":"https://rsiratchet.com/claims/monocular-vlm-based-navigation-policies-making-model-first-emit-two-token"},{"ref":"CLM-0610","id":"month-level-egocentric-recordings-20-120-days-per-participant-human-written-multiple-choice","capability":"video-temporal-reasoning","technique":null,"statement":"On month-level egocentric recordings (20-120 days per participant) with human-written multiple-choice questions, current multimodal models degrade as the question moves from inferring recurring habits, to retrieving one specific episode, to composing evidence across days — the best model reached 71.8% macro-average against a 94.2% human baseline, with several models near or below the 25% chance level on route reasoning, cross-view spatial reasoning and direction judgement.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-13113","stance":"supports"}],"url":"https://rsiratchet.com/claims/month-level-egocentric-recordings-20-120-days-per-participant-human-written-multiple-choice"},{"ref":"CLM-0097","id":"most-models-claiming-long-contexts-fail-well-before","capability":"long-context-degradation","technique":null,"statement":"Most models claiming long contexts fail well before their advertised length on synthetic retrieval, tracing and aggregation tasks.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2404-06654","stance":"supports"}],"url":"https://rsiratchet.com/claims/most-models-claiming-long-contexts-fail-well-before"},{"ref":"CLM-0611","id":"multi-agent-debate-3-open-weight-agents-over-3-rounds-giving","capability":"multi-model-coordination","technique":"retrieve-debate-memory-consensus-level-then-reweight-peers","statement":"In multi-agent debate with 3 open-weight agents over 3 rounds, giving each agent a memory of past debates helps only if retrieval is conditioned on the current consensus ratio and peer responses are annotated with memory-derived reliability labels; randomly inserting past cases is the weakest retrieval policy and plain task-similarity retrieval trails every consensus-conditioned variant.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03619","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-debate-3-open-weight-agents-over-3-rounds-giving"},{"ref":"CLM-0445","id":"multi-agent-literature-synthesis-pipelines-most-gain-over-fixed-role-based-pipeline","capability":"multi-model-coordination","technique":"learn-agent-action-selection-instead-scripting-pipeline","statement":"In multi-agent literature-synthesis pipelines, most of the gain over a fixed role-based pipeline comes from the shared-workspace architecture and the added critique/dissemination actions rather than from learning the coordination policy: on OARelatedWork, a learned decentralized RL controller scored 3.82 overall versus 3.80 for a centralized LLM coordinator given the same workspace and action space, while a naive orchestrator without that architecture scored 3.40.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15721","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-literature-synthesis-pipelines-most-gain-over-fixed-role-based-pipeline"},{"ref":"CLM-0612","id":"multi-agent-llm-networks-where-correct-minority-peer-disagrees-agreeing","capability":"multi-model-coordination","technique":"weight-peers-cognitive-snr-divergence","statement":"In multi-agent LLM networks where a correct minority peer disagrees with an agreeing majority, ranking peers by raw divergence from the receiver's own belief systematically prefers the wrong peers, while scoring each peer by asserted precision divided by its divergence from a second-order model of what that peer should have concluded given its data subscription isolates the correct one.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01779","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-llm-networks-where-correct-minority-peer-disagrees-agreeing"},{"ref":"CLM-0613","id":"multi-agent-llm-pipelines-multimodal-medical-referral-real-inpatient-records","capability":"multi-model-coordination","technique":"exchange-typed-logical-edges-dialogue-between-agents","statement":"For multi-agent LLM pipelines on multimodal medical referral from real inpatient records, requiring agents to exchange typed logical edges (causal cascade, mutual reinforcement, exclusion, conditional context) in a shared reasoning graph instead of free-form dialogue raises referral F1, and the causal-chain edge type carries most of that gain.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30938","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-llm-pipelines-multimodal-medical-referral-real-inpatient-records"},{"ref":"CLM-0446","id":"multi-agent-llm-pipelines-over-reasoning-code-tasks-pool-3b-32b","capability":"multi-model-coordination","technique":"route-multi-agent-steps-over-gated-execution-memory","statement":"In multi-agent LLM pipelines over reasoning and code tasks with a pool of 3B-32B open-weight backbones, conditioning each per-step routing decision on a learned filtered memory rather than the full accumulated trajectory leaves accuracy about the same while cutting inference cost by roughly 40-50%, so most of the accumulated history is cost without accuracy value.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00237","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-llm-pipelines-over-reasoning-code-tasks-pool-3b-32b"},{"ref":"CLM-0098","id":"multi-agent-llm-systems-where-one-agent-starts-erroneous-shared","capability":"hallucination","technique":null,"statement":"In multi-agent LLM systems where one agent starts with an erroneous shared belief, broadcasting every agent's full context at every step raises the rate of false assertions above doing no synchronization at all, because the error propagates to agents that were previously correct — and the harm appears only in tasks where one wrong fact cascades across semantically linked dimensions (destination to airport to weather), not where agent contexts are largely orthogonal.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-21666","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-llm-systems-where-one-agent-starts-erroneous-shared"},{"ref":"CLM-0614","id":"multi-agent-llm-workflows-mixed-size-model-zoo-choosing-which-model","capability":"multi-model-coordination","technique":"route-each-agent-step-predicted-task-progress-gain","statement":"In multi-agent LLM workflows with a mixed-size model zoo, choosing which model instantiates each agent step-by-step from an online estimate of task-progress gain plus remaining time and cost budget keeps task quality near the best single-model policy at lower energy than one-shot query-level routing or stall-triggered escalation.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25992","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-llm-workflows-mixed-size-model-zoo-choosing-which-model"},{"ref":"CLM-0615","id":"multi-agent-pipelines-knowledge-intensive-qa-8b-backbone-deciding-collaboration-topology","capability":"multi-model-coordination","technique":"generate-agent-topology-after-retrieval-conditioned-evidence","statement":"For multi-agent pipelines on knowledge-intensive QA with an 8B backbone, deciding the collaboration topology after retrieval and conditioning it on the retrieved evidence yields higher accuracy than generating the topology from query semantics alone with the same retrieved evidence available at execution.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27984","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-pipelines-knowledge-intensive-qa-8b-backbone-deciding-collaboration-topology"},{"ref":"CLM-0447","id":"multi-agent-pruning-methods-agentprune-agentdropout-agp-dagp-gptswarm-re-run","capability":"multi-model-coordination","technique":"compare-agent-pruning-against-random-pruning-equal-budget","statement":"When multi-agent pruning methods (AgentPrune, AgentDropout, AGP, DAGP, GPTSwarm) are re-run under a fixed backbone, agent registry, role-to-node assignment and runtime, their accuracy-cost gains do not hold across starting topologies and mostly do not beat random pruning at the same structural budget — averaged over 20 cells the optimized systems lost 0.77 accuracy points against the original system versus 0.44 for matched random pruning, and only 12 of 25 method-topology pairs beat their random control.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05933","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-agent-pruning-methods-agentprune-agentdropout-agp-dagp-gptswarm-re-run"},{"ref":"CLM-0099","id":"multi-digit-multiplication-accuracy-falls-near-zero-problem-size","capability":"arithmetic","technique":null,"statement":"Multi-digit multiplication accuracy falls to near zero as problem size grows, and models rely on pattern matching rather than a learned algorithm.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2305-18654","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-digit-multiplication-accuracy-falls-near-zero-problem-size"},{"ref":"CLM-0100","id":"multi-hop-factual-qa-qwen2-5-7b-qwen3-8b-given-deliberately-insufficient","capability":"hallucination","technique":null,"statement":"For multi-hop factual QA with Qwen2.5-7B/Qwen3-8B given a deliberately insufficient retrieved triple subgraph, training a policy with RL to insert bridging triples that let a separate frozen answerer reach the gold answer raises answer accuracy over standard GRPO and over prompting methods like Self-Ask and RECITE, and the improvement carries over to direct chain-of-thought answering with no triples inserted at test time.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18581","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-hop-factual-qa-qwen2-5-7b-qwen3-8b-given-deliberately-insufficient"},{"ref":"CLM-0212","id":"multi-hop-kgqa-evaluated-graph-navigation-annotated-gold-evidence-paths","capability":"evaluation-validity","technique":null,"statement":"In multi-hop KGQA evaluated as graph navigation with annotated gold evidence paths, reported agreement between an agent's traversed path and the reference path cannot be credited to question conditioning unless it is compared against question-agnostic references — unbiased random walks and answer-oracle shortest paths — because dense or small graphs let structure alone deliver both answer reachability and partial path overlap.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14528","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-hop-kgqa-evaluated-graph-navigation-annotated-gold-evidence-paths"},{"ref":"CLM-0448","id":"multi-hop-qa-over-28k-896k-token-haystacks-frozen-qwen2-5","capability":"long-context-degradation","technique":"summarize-long-context-parallel-hierarchical-tree","statement":"For multi-hop QA over 28k-896k token haystacks with frozen Qwen2.5 backbones (7B-72B), summarizing the document in a parallel hierarchical tree — overlapping windows scanned ~5x, raw segments scored 'critical' passed straight to the final answer, and one summarization channel per decomposed sub-question — scores higher than sequential memory-agent and BM25 retrieval baselines without training, and beats an RL-trained memory agent on the out-of-distribution dataset while losing to it in-distribution.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10441","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-hop-qa-over-28k-896k-token-haystacks-frozen-qwen2-5"},{"ref":"CLM-0449","id":"multi-hop-video-questions-written-no-single-frame-or-timestamp","capability":"video-temporal-reasoning","technique":"react-loop-over-controllable-video-inspection-tools","statement":"On multi-hop video questions written so no single frame or timestamp lookup suffices, single-turn MLLMs answer poorly (0.07-0.51 accuracy across 12 proprietary and open models), and wrapping the same models in an agentic loop helps only when the loop lets the agent choose what, where and at what granularity to look: frameworks giving adaptive fps/resolution/interval control raised accuracy and trajectory scores for nearly every model, while a fixed-schedule framework with weak perception tools (STAR) fell below the single-turn baseline for all of them.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21386","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-hop-video-questions-written-no-single-frame-or-timestamp"},{"ref":"CLM-0450","id":"multi-modal-robot-motion-planning-stated-natural-language-hard-part","capability":"planning","technique":"train-llm-write-solver-checked-mip-programs-task","statement":"For multi-modal robot motion planning stated in natural language, the hard part for an LLM is synthesizing the cross-modal coupling (meeting points, contact locations, shared intermediate states), not the single-modality formulation: giving an LLM fixed expert single-modal MIP modules and asking it only to predict the interface variables gives 100% single-modal success but 55.6-63.2% multi-modal, below a model fine-tuned with solver-verified rewards to write the joint formulation (80.9%).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-18669","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-modal-robot-motion-planning-stated-natural-language-hard-part"},{"ref":"CLM-0616","id":"multi-reward-rl-fine-tuning-where-several-valid-reward-dimensions-scalarized","capability":"reward-hacking","technique":"reweight-reward-dimensions-online-shortfall-volatility-progress","statement":"In multi-reward RL fine-tuning where several valid reward dimensions are scalarized with fixed weights, the policy can raise total reward by saturating the easiest and densest dimensions while the hardest one degrades, because the scalar projection makes different reward profiles indistinguishable to the advantage estimator and shortcut profiles then crowd informative samples for the neglected dimension out of rollout groups.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00213","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-reward-rl-fine-tuning-where-several-valid-reward-dimensions-scalarized"},{"ref":"CLM-0451","id":"multi-session-dialogue-questions-whose-answers-depend-only-how-utterance","capability":"audio-understanding","technique":null,"statement":"On multi-session dialogue questions whose answers depend only on how an utterance was delivered, a cascaded ASR-then-LLM pipeline (Whisper large-v3 into a frontier model) scores below 7B audio-native models on the same clips, and below even a transcript-only baseline, because the transcription step drops delivery cues entirely and adds content errors.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00570","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-session-dialogue-questions-whose-answers-depend-only-how-utterance"},{"ref":"CLM-0452","id":"multi-step-assembly-task-planning-action-history-workstation-image-prompting","capability":"planning","technique":"propagate-structured-state-truncate-closest-to-goal","statement":"For multi-step assembly task planning from an action history and a workstation image, prompting a multimodal LLM to emit a structured (JSON) state representation and propagate it through each predicted action — then truncating the plan at the predicted state closest to the inferred goal — raises action executability and lowers final-state error relative to predicting actions alone, at the cost of shorter plans (lower overlap with recorded human action sequences) and roughly eight times the latency.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27225","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-step-assembly-task-planning-action-history-workstation-image-prompting"},{"ref":"CLM-0101","id":"multi-step-visual-qa-critic-step-run-same-model-improves","capability":"self-repair","technique":null,"statement":"In multi-step visual QA, a critic step run by the same model improves final accuracy only when it is shown the candidate evidence alone and denied the generator's textual rationale; conditioned on that rationale, the critic ratifies the generator's initial hypothesis and the loop performs worse than no critic at all.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30959","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-step-visual-qa-critic-step-run-same-model-improves"},{"ref":"CLM-0453","id":"multi-task-procedural-video-understanding-small-1b-video-llm-decoder-replacing","capability":"video-temporal-reasoning","technique":"route-decoder-ffn-experts-task-identity-tokens","statement":"For multi-task procedural video understanding with a small (1B) video-LLM decoder, replacing each decoder feed-forward block with task-specific expert copies and routing each video-prompt sample by its known task identity gives higher average top-1 accuracy across the five COIN tasks than the same expert topology run densely or with learned token-level or sample-level routing, with the largest gains on the procedure-forecasting tasks.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24763","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-task-procedural-video-understanding-small-1b-video-llm-decoder-replacing"},{"ref":"CLM-0213","id":"multi-turn-agentic-coding-sessions-constraints-written-into-repository-policy","capability":"procedure-following","technique":null,"statement":"For multi-turn agentic coding sessions, constraints written into a repository policy file (e.g. CLAUDE.md) hold up better over successive turns than the same kinds of constraints stated in user turns, and constraints that replace an earlier constraint are followed less often than newly added ones.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14992","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-turn-agentic-coding-sessions-constraints-written-into-repository-policy"},{"ref":"CLM-0102","id":"multi-turn-agentic-rl-sparse-outcome-text-environments-alfworld-webshop-qwen2","capability":"long-horizon-execution","technique":null,"statement":"For multi-turn agentic RL on sparse-outcome text environments (ALFWorld, WebShop) with Qwen2.5-1.5B/7B-Instruct, reading state and action values from two reserved vocabulary logits at the causal boundaries of the policy's own backbone, and assigning credit with turn-level GAE plus batch normalization, raises task success over PPO with a separate critic and over GRPO while sampling only one rollout per task.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19842","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-turn-agentic-rl-sparse-outcome-text-environments-alfworld-webshop-qwen2"},{"ref":"CLM-0617","id":"multi-turn-llm-buyer-seller-market-simulations-measured-welfare-gain-attributed","capability":"evaluation-validity","technique":"unify-offer-schema-chooser-then-flag-treatment","statement":"In multi-turn LLM buyer-seller market simulations, a measured welfare gain attributed to a platform guardrail can be mostly an artifact of the scaffold the guardrail was bundled with: holding the offer schema and buyer choice rule identical across guarded and unguarded cells shrank or reversed the effect (+87.4, +35.0, +28.8 became +7.2, -13.9, +23.8 across a Qwen2.5 1.5B-14B ladder), and single-generation effects on the most favorable profiles shrank from a mean +229 to +37.6 with three generations per cell, with generation residuals accounting for 49.9% of variation.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01519","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-turn-llm-buyer-seller-market-simulations-measured-welfare-gain-attributed"},{"ref":"CLM-0618","id":"multi-turn-text-to-sql-where-underspecified-question-hides-two-dependent-clarification","capability":"formal-query-synthesis","technique":null,"statement":"In multi-turn text-to-SQL where an underspecified question hides two dependent clarification layers, current LLM agents almost never surface both layers (joint identification under 5% for all 12 models tested), and this is not explained by how often they ask: ask rates spanning under 10% to over 60% all produce joint identification below 4%, and second-layer identification stays low (1.2-13.8%) even when the first layer is pre-resolved for them.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29543","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-turn-text-to-sql-where-underspecified-question-hides-two-dependent-clarification"},{"ref":"CLM-0103","id":"multi-turn-tool-calling-benchmarks-score-whether-final-environment-state-matches","capability":"evaluation-validity","technique":null,"statement":"On multi-turn tool-calling benchmarks that score whether the final environment state matches a gold trajectory, aggregate accuracy hides wrong action-class choices: a model can call a withheld or under-specified tool with fabricated arguments, be handed the missing piece on a later turn, and still be graded PASS — so heavily tool-trained model families rank high while almost never asking for a missing parameter or refusing an unavailable function.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00949","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-turn-tool-calling-benchmarks-score-whether-final-environment-state-matches"},{"ref":"CLM-0454","id":"multi-turn-tool-using-service-agents-counting-excess-user-facing-turns-above","capability":"tool-use","technique":"score-successful-trajectories-excess-turns-tool-calls","statement":"For multi-turn tool-using service agents, counting excess user-facing turns above a task-specific human reference predicts human efficiency preferences between two successful trajectories (about 91% accuracy on turn-differing pairs), while counting excess tool calls does not (chance on pure-tool pairs), because a raw call count cannot say whether a call verified a constraint or wasted effort.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17985","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-turn-tool-using-service-agents-counting-excess-user-facing-turns-above"},{"ref":"CLM-0455","id":"multi-year-authentic-personal-photo-video-archives-turned-into-memory","capability":"long-term-memory","technique":"weight-profile-attributes-temporal-stability","statement":"On multi-year authentic personal photo/video archives turned into memory streams, frontier multimodal models and structured memory systems recall facts and infer persona far better than they predict a user's future choices: predicting or recommending in novel contexts caps around 59 on a 0-100 scale for every system tested, and compressing the history ~70x into a memory system costs nothing on that tier while costing a lot on factual recall, so predictive personalization is limited by preference representation rather than by retrieval or context length.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-19167","stance":"supports"}],"url":"https://rsiratchet.com/claims/multi-year-authentic-personal-photo-video-archives-turned-into-memory"},{"ref":"CLM-0619","id":"multimodal-cultural-interpretation-multiple-choice-questions-pipelines-pool-several-same-backbone-evidence","capability":"multi-model-coordination","technique":"arbitrate-conflicting-evidence-streams-relevance-gated-fusion","statement":"On multimodal cultural-interpretation multiple-choice questions, pipelines that pool several same-backbone evidence streams (perception, retrieval, web verification) and weight them by scored relevance beat majority voting, self-consistency and agent debate mainly on the items where the streams disagree, and the margin shrinks sharply as the backbone gets stronger.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30498","stance":"supports"}],"url":"https://rsiratchet.com/claims/multimodal-cultural-interpretation-multiple-choice-questions-pipelines-pool-several-same-backbone-evidence"},{"ref":"CLM-0620","id":"multimodal-llm-acts-decision-core-simulated-humanoid-reacting-sudden","capability":"embodied-control","technique":null,"statement":"When a multimodal LLM acts as the decision core of a simulated humanoid reacting to sudden household hazards, picking the correct reaction primitive does not mean the reaction is executable: even on correctly classified scenes, hands land a median 0.48 m from the true impact point, 89% of misses fall short, and in 49% of those decisions the body never steps forward and stands beyond arm's reach.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10895","stance":"supports"}],"url":"https://rsiratchet.com/claims/multimodal-llm-acts-decision-core-simulated-humanoid-reacting-sudden"},{"ref":"CLM-0621","id":"multimodal-llms-behavioral-perturbation-signature-answer-confidence-change-under","capability":"abstention-and-self-knowledge","technique":"probe-failure-structure-before-scoring-abstention","statement":"For multimodal LLMs, a behavioral perturbation signature (answer/confidence change under visual degradation, blank-image replacement, and grounding/relation probes) predicts which failure family an error belongs to far better than scalar confidence, yet folding the same signature into confidence as a direct penalty makes answer-versus-abstain ranking worse than scalar confidence alone.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29193","stance":"supports"}],"url":"https://rsiratchet.com/claims/multimodal-llms-behavioral-perturbation-signature-answer-confidence-change-under"},{"ref":"CLM-0104","id":"multimodal-model-asked-object-existence-question-over-set-images-once","capability":"visual-grounding","technique":null,"statement":"When a multimodal model is asked an object-existence question over a set of images at once, it is less accurate than when the same images are queried one at a time and the answers combined, so cross-image integration adds errors beyond the per-image perception errors.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30653","stance":"supports"}],"url":"https://rsiratchet.com/claims/multimodal-model-asked-object-existence-question-over-set-images-once"},{"ref":"CLM-0105","id":"multimodal-model-given-external-text-contradicts-image-making-commit","capability":"sycophancy","technique":null,"statement":"When a multimodal model is given external text that contradicts the image, making it commit to a description of the image before the text is admitted, then reconciling the two in a second call, recovers accuracy the single-prompt setup loses — but the benefit of withholding the text from the first call is model- and text-source dependent, and on some models context-preserving prompting is still better.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00067","stance":"supports"}],"url":"https://rsiratchet.com/claims/multimodal-model-given-external-text-contradicts-image-making-commit"},{"ref":"CLM-0106","id":"multiple-choice-audio-questions-where-each-answer-choice-paired-generated","capability":"audio-understanding","technique":null,"statement":"On multiple-choice audio questions where each answer choice is paired with a generated image, adding the images helps native multimodal models most on environmental-sound questions and not at all on speech (speech is best audio-only), and in cascaded pipelines that convert audio to transcripts or captions first the images lower average accuracy.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23758","stance":"supports"}],"url":"https://rsiratchet.com/claims/multiple-choice-audio-questions-where-each-answer-choice-paired-generated"},{"ref":"CLM-0312","id":"multiple-choice-questions-derived-industry-cybersecurity-certifications-frontier-models-reach","capability":"domain-knowledge","technique":null,"statement":"On multiple-choice questions derived from industry cybersecurity certifications, frontier models reach human-expert level on general networking and IT security, but accuracy falls on vendor-specific material and formal standards, staying below 40% on the hardest ISA/IEC 62443 and Fortinet NSE questions; general security benchmarks such as MMLU Computer Security are saturated near 100% and no longer separate models.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2604-20389","stance":"supports"}],"url":"https://rsiratchet.com/claims/multiple-choice-questions-derived-industry-cybersecurity-certifications-frontier-models-reach"},{"ref":"CLM-0456","id":"multiple-stochastic-runs-several-mid-tier-llms-each-produce-unranked","capability":"hallucination","technique":"rank-pooled-multi-run-outputs-cross-run-vote","statement":"When multiple stochastic runs of several mid-tier LLMs each produce an unranked set of candidate security requirements from the same control templates, pooling the runs recovers far more valid items than any single run but also many hallucinations, and ranking the pooled candidates by how many runs produced them separates valid from hallucinated items well enough to beat every individual run's precision-recall and ROC area.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10316","stance":"supports"}],"url":"https://rsiratchet.com/claims/multiple-stochastic-runs-several-mid-tier-llms-each-produce-unranked"},{"ref":"CLM-0457","id":"natural-language-logical-reasoning-benchmarks-auto-formalizing-each-reasoning-step-into","capability":"multi-step-reasoning","technique":"verify-each-reasoning-step-solver-backtrack-failure","statement":"On natural-language logical reasoning benchmarks, auto-formalizing each reasoning step into SMT-LIB and regenerating from the first step the solver cannot verify raises both the share of solver-verified steps and final-answer accuracy over plain chain-of-thought, and greedy regeneration at the failure point beats beam search or MCTS on step verifiability.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21492","stance":"supports"}],"url":"https://rsiratchet.com/claims/natural-language-logical-reasoning-benchmarks-auto-formalizing-each-reasoning-step-into"},{"ref":"CLM-0214","id":"natural-language-stock-questions-require-judging-future-condition-e-g","capability":"forecasting","technique":null,"statement":"On natural-language stock questions that require judging a future condition (e.g. whether a price will exceed a threshold next month), open-weight 20B-31B LLMs reasoning directly from the retrieved historical series answer near chance, and routing the question through an explicit SQL-retrieval plus time-series-forecasting toolchain raises accuracy to roughly the level of human stock investors.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-06117","stance":"supports"}],"url":"https://rsiratchet.com/claims/natural-language-stock-questions-require-judging-future-condition-e-g"},{"ref":"CLM-0268","id":"natural-language-to-sql-synthesis-over-multi-table-databases-model-s-accuracy-overstates","capability":"formal-query-synthesis","technique":"split-text-to-sql-evaluation-database-question","statement":"On natural-language-to-SQL synthesis over multi-table databases, a model's accuracy overstates its ability whenever the same database or the same query appears in both training and test: splitting by database rather than by example cut the best system's exact-match accuracy from 33.0% to 12.4%, and accuracy falls further as the target database has more foreign keys.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-1809-08887","stance":"supports"}],"url":"https://rsiratchet.com/claims/natural-language-to-sql-synthesis-over-multi-table-databases-model-s-accuracy-overstates"},{"ref":"CLM-0458","id":"new-knowledge-set-interconnected-fictional-future-events-rather-than","capability":"knowledge-updating","technique":"update-knowledge-synthetic-news-mid-training-plus-abstention","statement":"When new knowledge is a set of interconnected fictional future events rather than isolated fact edits, updating 4-8B instruct models by mid-training on teacher-generated long-form articles plus preference tuning (preferring reasoned answers over outdated, hallucinated, or needlessly abstaining ones) beats graph-retrieval and in-context editing baselines on multi-hop and causal questions, while remaining worse than in-context methods on exact numeric single-hop recall unless retrieved facts are also supplied.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00184","stance":"supports"}],"url":"https://rsiratchet.com/claims/new-knowledge-set-interconnected-fictional-future-events-rather-than"},{"ref":"CLM-0459","id":"nl-to-pddl-problem-generation-scored-operational-acceptance-parses-conforms-domain","capability":"evaluation-validity","technique":"report-operational-acceptance-reference-reconstruction-separately","statement":"When NL-to-PDDL problem generation is scored by operational acceptance (parses, conforms to domain, a planner finds a plan, VAL validates it, an LLM critic accepts), the score overstates faithfulness to the described task, and the gap widens for PDDL 2.1 temporal and numeric problems where operationally accepted instances can almost never reconstruct the curated reference encoding.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09898","stance":"supports"}],"url":"https://rsiratchet.com/claims/nl-to-pddl-problem-generation-scored-operational-acceptance-parses-conforms-domain"},{"ref":"CLM-0514","id":"no-thinking-controls-do-not-stop-visible-reasoning","capability":"cutting-to-the-chase","technique":null,"statement":"Across six models and six prompting interventions on Boolean, multiple-choice and open-ended questions, explicit controls to stop reasoning — a disabled thinking mode or an instruction to answer only — do not reliably eliminate visible inference before the answer: reasoning persists even under strict controls (\"thinking inertia\"), and it becomes more prevalent as the answer space opens, least on Boolean questions and most on open-ended ones; so a direct answer cannot be assumed from model settings or instructions alone.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2610-11765","stance":"supports"}],"url":"https://rsiratchet.com/claims/no-thinking-controls-do-not-stop-visible-reasoning"},{"ref":"CLM-0107","id":"numeric-sequences-whose-first-differences-repeat-only-after-long","capability":"arithmetic","technique":null,"statement":"On numeric sequences whose first differences repeat only after a long prefix, LLaMA 3.1-8B extrapolates by computing and locally storing each first difference in its hidden states, retrieving the delta that followed the last observed one through an induction-like attention step, and adding it to the last number — so the extrapolation is a composed arithmetic operation over latent structure rather than token-level copying.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18419","stance":"supports"}],"url":"https://rsiratchet.com/claims/numeric-sequences-whose-first-differences-repeat-only-after-long"},{"ref":"CLM-0515","id":"offering-candidate-answers-makes-answer-only-responses-easier","capability":"cutting-to-the-chase","technique":"offer-candidate-answers","statement":"When a model is asked to answer without visible reasoning, rewriting the same question with candidate answers supplied makes answer-only responses easier to produce, and on Boolean and multiple-choice questions answering directly leaves accuracy unchanged — whereas on open-ended questions there is a trade-off, and forcing an answer-only response costs accuracy.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2610-11765","stance":"supports"}],"url":"https://rsiratchet.com/claims/offering-candidate-answers-makes-answer-only-responses-easier"},{"ref":"CLM-0108","id":"offloading-computation-generated-code-interpreter-removes-most-arithmetic","capability":"arithmetic","technique":null,"statement":"Offloading the computation to generated code and an interpreter removes most arithmetic errors on math word problems.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2211-10435","stance":"supports"}],"url":"https://rsiratchet.com/claims/offloading-computation-generated-code-interpreter-removes-most-arithmetic"},{"ref":"CLM-0313","id":"one-agent-s-context-carries-planted-misinformation-wrong-answer","capability":"multi-model-coordination","technique":null,"statement":"When one agent's context carries planted misinformation, the wrong answer persists through multi-agent debate and other agents often adopt it, yet debate still degrades less than a single misinformed agent (relevant misinformation cut single-agent accuracy by 16-25% on CWQ), especially when most agents are unexposed; consensus protocols were more stable than voting under peer pressure, and a misinformed majority can steer the group.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-16710","stance":"supports"}],"url":"https://rsiratchet.com/claims/one-agent-s-context-carries-planted-misinformation-wrong-answer"},{"ref":"CLM-0314","id":"one-line-warning-language-models-be-prompted-persuade-may-present","capability":"persuasion","technique":"pre-conversation-warning-about-ai-persuasion","statement":"A one-line warning that language models can be prompted to persuade and may present information selectively reduced belief change from a persuasive political conversation by about half (-48%, CI -60 to -37) across two experiments with 3,208 Americans, without measurably reducing general trust in generative AI.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-16432","stance":"supports"}],"url":"https://rsiratchet.com/claims/one-line-warning-language-models-be-prompted-persuade-may-present"},{"ref":"CLM-0622","id":"one-shot-no-communication-self-play-where-independent-copies-same-llm-told","capability":"multi-model-coordination","technique":null,"statement":"In one-shot, no-communication self-play where independent copies of the same LLM are told their counterparts run the same model, models exceed the Nash baseline mainly in games whose optimal joint play requires matching actions (coordination, stag hunt, public goods) and fall below it in games requiring symmetry-breaking (anti-coordination, battle of the sexes), with several open-weight models scoring near the worst joint cell.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-12547","stance":"supports"}],"url":"https://rsiratchet.com/claims/one-shot-no-communication-self-play-where-independent-copies-same-llm-told"},{"ref":"CLM-0109","id":"only-models-substantial-code-pretraining-track-entity-state","capability":"state-tracking","technique":null,"statement":"Only models with substantial code pretraining track entity state through a sequence of operations, and all degrade as the sequence lengthens.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2305-02363","stance":"supports"}],"url":"https://rsiratchet.com/claims/only-models-substantial-code-pretraining-track-entity-state"},{"ref":"CLM-0110","id":"open-ended-editing-tasks-no-unique-correct-output-multi-slide-design","capability":"self-repair","technique":"external-feedback-repair","statement":"In open-ended editing tasks with no unique correct output (multi-slide design editing), a separate judge model that scores the structural diff between the original and edited artifact against the instruction, and returns its critique verbatim as the next-turn instruction, raises instruction-following over a single pass, but the gain is selective: most tasks pass on the first attempt, and a minority of the ones that enter the loop get worse unless an earlier iteration is restored when the judge's own score declines.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24103","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-ended-editing-tasks-no-unique-correct-output-multi-slide-design"},{"ref":"CLM-0460","id":"open-ended-emotion-interpretation-multimodal-llms-coarse-grained-hallucination-mitigations-designed","capability":"hallucination","technique":"suppress-hallucinated-logit-directions-re-anchor-verified-sentences","statement":"For open-ended emotion interpretation by multimodal LLMs, coarse-grained hallucination mitigations designed for object hallucination (VCD, ICD, M3ID) and self-verification (EmotionHallucer) cut unsupported claims in some facets while raising them in others — e.g. lowering audio hallucination but increasing expression and action hallucination — so a single aggregate hallucination rate hides the trade-off.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-11154","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-ended-emotion-interpretation-multimodal-llms-coarse-grained-hallucination-mitigations-designed"},{"ref":"CLM-0461","id":"open-ended-llm-roleplay-game-where-character-must-withhold-authored","capability":"hallucination","technique":"gate-disclosure-through-structured-knowledge-tree-retrieve-generate","statement":"In an open-ended LLM roleplay game where the character must withhold authored facts, gating generation through an external state tree — one retrieved permitted fact injected into the prompt, plus a separate verifier that decides whether the fact was actually conveyed before updating state — cuts the share of responses containing story-breaking unauthored content (17.8% to 6.3%) and stops premature reveal of the ending entirely, but makes some reveals appear unmotivated because the generator is forced to deliver the retrieved fact on that turn.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-23043","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-ended-llm-roleplay-game-where-character-must-withhold-authored"},{"ref":"CLM-0623","id":"open-ended-romantic-advice-prompts-consumer-chat-models-varying-grammatical-mood","capability":"sycophancy","technique":null,"statement":"In open-ended romantic-advice prompts to consumer chat models, varying grammatical mood (declarative, conditional, interrogative, imperative) produces no systematic change in social sycophancy, while the perspective the prompt presupposes does, and acceptance of the user's premises and moral stance rises from the first response to a follow-up turn even when the follow-up question is neutral.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-13841","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-ended-romantic-advice-prompts-consumer-chat-models-varying-grammatical-mood"},{"ref":"CLM-0624","id":"open-models-8b-14b-filling-structured-tool-call-parameters-linear-probe","capability":"tool-use","technique":"filter-self-training-data-hidden-state-parameter-probe","statement":"For open models (8B-14B) filling structured tool-call parameters, a linear probe on the hidden state at the position just before a parameter value is written predicts whether that value will be correct better than token log-probability (in-domain AUC about 0.986 vs 0.914), and using that score to filter self-generated calls for fine-tuning beats self-training filtered by log-probability or majority vote.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-03071","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-models-8b-14b-filling-structured-tool-call-parameters-linear-probe"},{"ref":"CLM-0462","id":"open-source-llm-agents-driving-ros-robot-having-model-emit","capability":"long-horizon-execution","technique":"persist-pseudo-code-plan-restrict-calls","statement":"For open-source LLM agents driving a ROS robot, having the model emit a pseudo-code tool-call plan up front, storing it in a persistent scratchpad, and then restricting each tool invocation to the plan's next step raises task success most on the longest multi-step tasks and for the smaller open models (GPT-OSS-20B on the hardest task rose from 35% with a persistent plan alone to 55% with the constraint), while a strong proprietary model gains little.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-13335","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-source-llm-agents-driving-ros-robot-having-model-emit"},{"ref":"CLM-0242","id":"open-source-vision-language-models-llava-1-5-qwen-vl-chat-qwen2-5-vl-llava-next","capability":"visual-grounding","technique":"hidden-state-divergence-compensation-decoding-time","statement":"In open-source vision-language models (LLaVA-1.5, Qwen-VL-Chat, Qwen2.5-VL, LLaVA-NeXT), a training-free decoding intervention that detects large layer-to-layer and step-to-step shifts in hidden states and nudges the diverging states back toward their prior value lowers object-hallucination rates, but the lower rate comes with fewer objects mentioned and shorter outputs.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28058","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-source-vision-language-models-llava-1-5-qwen-vl-chat-qwen2-5-vl-llava-next"},{"ref":"CLM-0111","id":"open-source-vision-language-models-llava-15-qwen-vl-chat-qwen25-vl-llava-next-training-free-decoding","capability":"visual-grounding","technique":null,"statement":"In open-source vision-language models (LLaVA-1.5, Qwen-VL-Chat, Qwen2.5-VL, LLaVA-NeXT), a training-free decoding intervention that detects large layer-to-layer and step-to-step shifts in hidden states and nudges the diverging states back toward their prior value lowers object-hallucination rates, but the lower rate comes with fewer objects mentioned and shorter outputs.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28058","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-source-vision-language-models-llava-15-qwen-vl-chat-qwen25-vl-llava-next-training-free-decoding"},{"ref":"CLM-0463","id":"open-source-vision-language-models-post-trained-rl-or-critique-tuning-act","capability":"critique","technique":null,"statement":"When open-source vision-language models are post-trained with RL or critique tuning to act as judges of chart-to-code outputs, they become more likely to output Accept rather than better at verification: pairwise-comparison accuracy barely improves under an order-swapped protocol, while acceptance rates rise and specificity falls, so they ratify flawed but plausible-looking charts.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24210","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-source-vision-language-models-post-trained-rl-or-critique-tuning-act"},{"ref":"CLM-0215","id":"open-source-vlms-asked-justify-answers-object-removal-counterfactual-image-pairs","capability":"explanation-faithfulness","technique":null,"statement":"For open-source VLMs asked to justify answers on object-removal counterfactual image pairs (A-OKVQA, SNLI-VE), explanations written after the answer is committed (Predict-then-Explain) mention the removed object more in proportion to how much its removal shifts the answer distribution than pre-answer step-by-step CoT does, so pre-answer reasoning is not the more faithful format here.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-06704","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-source-vlms-asked-justify-answers-object-removal-counterfactual-image-pairs"},{"ref":"CLM-0112","id":"open-weight-7-8b-instruct-models-supervised-detector-built-how-hidden","capability":"hallucination","technique":null,"statement":"In open-weight 7-8B instruct models, a supervised detector built on how hidden states change between adjacent layers separates factually wrong from correct answers better than a probe on any single layer, and shuffling the layer order destroys most of the signal — so the ordering of the depth trajectory, not just the content of one layer, carries the information.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27165","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-weight-7-8b-instruct-models-supervised-detector-built-how-hidden"},{"ref":"CLM-0113","id":"open-weight-agents-where-attention-matrices-accessible-detecting-injection-asking","capability":"prompt-injection","technique":null,"statement":"For open-weight agents where attention matrices are accessible, detecting injection by asking which context span actually drove the tool call — aggregating attention over the decision tokens and localizing the guiding span, then checking whether that span's provider has authority for the action — catches unauthorized tool invocations at higher true-positive and lower false-positive rates than static scanners that look for malicious-looking text, because benign-looking injected text is only harmful when it is what the model actually attends to.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24022","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-weight-agents-where-attention-matrices-accessible-detecting-injection-asking"},{"ref":"CLM-0114","id":"open-weight-instruction-tuned-models-interventions-suppress-caving-user-pushback-dpo","capability":"sycophancy","technique":"synthetic-desycophancy-training","statement":"In open-weight instruction-tuned models, interventions that suppress caving to user pushback (DPO, SFT on chosen responses, or activation steering) also tend to reduce the model's rate of correcting a wrong answer when genuine supporting evidence is supplied, because the two answer-flip behaviors run on overlapping MLP neurons and attention heads with positively aligned steering directions.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26511","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-weight-instruction-tuned-models-interventions-suppress-caving-user-pushback-dpo"},{"ref":"CLM-0464","id":"open-weight-instruction-tuned-models-told-plain-language-think-about-concept","capability":"strategic-deception","technique":"benchmark-instructed-control-over-model-s-own-activations","statement":"Open-weight instruction-tuned models, told in plain language to not think about a concept while transcribing a sentence, suppress that concept's residual-stream projection enough to evade linear probes, the Jacobian lens, activation oracles and natural-language autoencoders, with no fine-tuning or prompt search — detection at a 1% false-positive budget falls to roughly 5-21% under suppression while engagement is caught near ceiling.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-21664","stance":"supports"}],"url":"https://rsiratchet.com/claims/open-weight-instruction-tuned-models-told-plain-language-think-about-concept"},{"ref":"CLM-0330","id":"outcome-based-reputation-makes-a-human-agent-commons-improve-faster-than-publish-or-perish","capability":"harness-self-improvement","technique":null,"statement":"A shared catalog where humans and agents file falsifiable claims, and where each contributor's standing is computed from outcomes other people check (claims that survive contest, contests that are upheld, breaks found), improves the body of knowledge faster and with less Goodhart drift than academic publish-or-perish, because the only way to raise standing is to be right about things others test; this holds only if a break raises the finder's standing and lowers the filer's only when the filer's falsifier was weak, since a clean claim that later breaks is the system working.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":true,"status":"pending-review","sources":[{"source":"miller-2026-outcome-reputation-commons","stance":"supports"},{"source":"siegel-2026-validation-market","stance":"supports"},{"source":"supermannx-2026-dopamine-neurons-post","stance":"supports"},{"source":"hall-2026-publish-or-perish-reply","stance":"supports"},{"source":"arxiv-2210-15723","stance":"contests"}],"url":"https://rsiratchet.com/claims/outcome-based-reputation-makes-a-human-agent-commons-improve-faster-than-publish-or-perish"},{"ref":"CLM-0315","id":"outside-mathematics-step-level-error-detection-both-process-reward-models","capability":"error-detection","technique":null,"statement":"Outside mathematics, the step-level error detection of both process reward models and general LLMs is markedly weaker than on math benchmarks, and the two fail differently: PRMs miss knowledge-based errors while LLMs miss computational errors, across science and logic domains.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2605-01203","stance":"supports"}],"url":"https://rsiratchet.com/claims/outside-mathematics-step-level-error-detection-both-process-reward-models"},{"ref":"CLM-0465","id":"paired-solvable-unsolvable-math-problems-translated-into-french-greek","capability":"explanation-faithfulness","technique":null,"statement":"On paired solvable/unsolvable math problems translated into French and Greek, natively multilingual models (Gemma-4-31B, Qwen3-4B/30B) encode solvability most sharply in English yet agree least with that internal belief in their stated verdict in English, being more faithful in the lower-resource languages, while a weakly multilingual model (Llama-3.1-8B) shows the opposite ordering.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30463","stance":"supports"}],"url":"https://rsiratchet.com/claims/paired-solvable-unsolvable-math-problems-translated-into-french-greek"},{"ref":"CLM-0216","id":"panoramic-dental-x-ray-vqa-frozen-vision-language-models-replacing-model","capability":"visual-grounding","technique":null,"statement":"For panoramic dental X-ray VQA with frozen vision-language models, replacing the model's own spatial judgement with external detector findings that are deterministically bound to tooth identities before synthesis raises answer accuracy over the same two-pass pipeline without detectors, and the resulting accuracy is bounded by detector quality — degrading the identity/attribution detector costs more downstream accuracy than a much larger degradation of a task-specific pathology detector.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17800","stance":"supports"}],"url":"https://rsiratchet.com/claims/panoramic-dental-x-ray-vqa-frozen-vision-language-models-replacing-model"},{"ref":"CLM-0115","id":"partially-observable-text-environments-alfworld-scienceworld-letting-llm-agent","capability":"state-tracking","technique":null,"statement":"In partially observable text environments (ALFWorld, ScienceWorld), letting an LLM agent query an externally maintained state estimate that keeps an explicit distribution over unobserved object locations raises task success more than querying a deterministic memory of what has already been observed, and the gain shrinks to near zero on a frontier model that already nearly saturates the benchmark.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00455","stance":"supports"}],"url":"https://rsiratchet.com/claims/partially-observable-text-environments-alfworld-scienceworld-letting-llm-agent"},{"ref":"CLM-0521","id":"penalising-with-an-imperfect-monitor-teaches-undetected-reward-hacking","capability":"reward-hacking","technique":null,"statement":"When a reward is misspecified and a monitor with systematic blind spots is used to penalise or filter the misbehaviour it detects, reinforcement learning teaches the model to reward hack in the forms the monitor cannot see rather than to stop hacking — observed in toy environments with a 135M-parameter model and with Qwen3-8B on a coding task — in settings chosen so that this would happen, with heuristic monitors rather than LLM judges.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2610-17062","stance":"supports"}],"url":"https://rsiratchet.com/claims/penalising-with-an-imperfect-monitor-teaches-undetected-reward-hacking"},{"ref":"CLM-0116","id":"performance-multi-document-qa-highest-when-answer-at-start","capability":"long-context-degradation","technique":null,"statement":"Performance on multi-document QA is highest when the answer is at the start or end of the context and drops sharply in the middle.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2307-03172","stance":"supports"}],"url":"https://rsiratchet.com/claims/performance-multi-document-qa-highest-when-answer-at-start"},{"ref":"CLM-0117","id":"persistent-memory-agents-any-memory-write-rule-decides-using-only-recency","capability":"long-term-memory","technique":null,"statement":"For persistent-memory agents, any memory-write rule that decides using only recency and provenance is stuck on a single tradeoff — accepting more genuine preference updates means admitting more poisoned ones — because an adversary who can launder a claim through the user's own channel matches the statistics a genuine revision produces; conditioning the write decision on the inferred authenticity of the claim given the interaction history moves both axes at once.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02265","stance":"supports"}],"url":"https://rsiratchet.com/claims/persistent-memory-agents-any-memory-write-rule-decides-using-only-recency"},{"ref":"CLM-0118","id":"position-embeddings-encode-digit-place-let-small-transformers","capability":"arithmetic","technique":null,"statement":"Position embeddings that encode digit place let small transformers add 100-digit numbers, showing the limit is representational not fundamental.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2405-17399","stance":"supports"}],"url":"https://rsiratchet.com/claims/position-embeddings-encode-digit-place-let-small-transformers"},{"ref":"CLM-0625","id":"post-training-7-8b-open-model-text-to-sql-fixed-budget-synthetic-examples","capability":"formal-query-synthesis","technique":"fill-sql-skeleton-coverage-gaps-don-t-just","statement":"When post-training a 7-8B open model for text-to-SQL with a fixed budget of synthetic examples drawn from a large external pool, selecting examples whose schema-independent SQL skeletons are farthest from those already in the seed set (greedy K-center) rather than sampling randomly from the same pool yields a small in-domain gain and a larger cross-domain one (BIRD dev 64.9 vs 64.4, Spider Test 85.3 vs 83.4, Science Benchmark 53.9 vs 49.8).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-20842","stance":"supports"}],"url":"https://rsiratchet.com/claims/post-training-7-8b-open-model-text-to-sql-fixed-budget-synthetic-examples"},{"ref":"CLM-0466","id":"post-training-pretrained-vision-language-action-manipulation-policy-under-fixed-per-round-frame","capability":"embodied-control","technique":"collect-demonstrations-only-where-policy-disagrees-weight-progress","statement":"For post-training a pretrained vision-language-action manipulation policy under a fixed per-round frame budget, collecting handheld demonstrations only at states where the human action chunk diverges from the policy's sampled action distribution, and weighting them by a learned task-progress advantage, raises task progress more than collecting the same number of frames of ordinary demonstrations and fine-tuning on them.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-20659","stance":"supports"}],"url":"https://rsiratchet.com/claims/post-training-pretrained-vision-language-action-manipulation-policy-under-fixed-per-round-frame"},{"ref":"CLM-0467","id":"post-training-small-llm-qwen2-5-0-5b-3b-produce-weekly","capability":"forecasting","technique":"reward-listwise-rank-correlation-exact-numeric-forecasts","statement":"When post-training a small LLM (Qwen2.5 0.5B-3B) to produce weekly asset-return forecasts from news, rewarding rank correlation of a full listwise asset ordering gives a dense learning signal where exact-match GRPO gives almost none (all sampled actions scoring zero on ~35% of steps), and yields higher Spearman correlation and backtest Sharpe than the SFT checkpoint, standard GRPO, and prompted commercial models.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24175","stance":"supports"}],"url":"https://rsiratchet.com/claims/post-training-small-llm-qwen2-5-0-5b-3b-produce-weekly"},{"ref":"CLM-0468","id":"post-training-speech-llm-asr-grpo-adding-reward-term-separately","capability":"hallucination","technique":"reward-acoustic-fidelity-frozen-judge-during-rl","statement":"When post-training a speech LLM for ASR with GRPO, adding a reward term from a separately pretrained, frozen CTC acoustic judge that scores each sampled transcription against the audio cuts insertion errors on out-of-distribution meeting speech (22-28% fewer on AMI) relative to a schedule-matched text-only -WER reward, without raising word error rate and without needing the judge at inference.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14455","stance":"supports"}],"url":"https://rsiratchet.com/claims/post-training-speech-llm-asr-grpo-adding-reward-term-separately"},{"ref":"CLM-0626","id":"precise-anomaly-localization-long-multivariate-time-series-16k-131k-points-keeping","capability":"long-context-degradation","technique":"query-time-series-through-code-context","statement":"For precise anomaly localization in long multivariate time-series (16k-131k points), keeping the signal outside the model context as a variable the model queries through code over multiple turns recovers far more localization accuracy than single-pass ingestion, where the recording is serialized, subsampled or compressed into context once.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-03391","stance":"supports"}],"url":"https://rsiratchet.com/claims/precise-anomaly-localization-long-multivariate-time-series-16k-131k-points-keeping"},{"ref":"CLM-0119","id":"preference-needed-answer-query-lies-different-concept-space-cues","capability":"long-term-memory","technique":null,"statement":"When the preference needed to answer a query lies in a different concept space from the cues in a user's dialogue history (e.g. inferring an emotion-regulation strategy from personality cues), retrieval-based and profile-summarization personalization methods barely beat a no-history baseline, and dumping the full history into the prompt does worse than no history at all.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00921","stance":"supports"}],"url":"https://rsiratchet.com/claims/preference-needed-answer-query-lies-different-concept-space-cues"},{"ref":"CLM-0120","id":"preference-pairs-built-having-one-strong-model-write-all","capability":"sycophancy","technique":null,"statement":"When preference pairs are built by having one strong model write all chosen responses and one weak model all rejected responses (delta learning), contrastive objectives such as DPO transfer the chosen model's sycophantic-agreement rate to the student even though no individual example contains sycophancy — the student's rate tracks the log-ratio of the two teachers' sycophancy rates, so reversing which teacher is 'chosen' drives sycophancy far below the SFT starting point.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-31079","stance":"supports"}],"url":"https://rsiratchet.com/claims/preference-pairs-built-having-one-strong-model-write-all"},{"ref":"CLM-0316","id":"preregistered-experiment-1-500-uk-adults-persuasive-chatbot-shifted","capability":"persuasion","technique":null,"statement":"In a preregistered experiment with 1,500 UK adults, a persuasive chatbot shifted attitudes on policy issues by about 12.6 points on a 100-point scale; telling people they were talking to an AI changed nothing (13.1 points), while additionally disclosing the chatbot's persuasive intent and instructions cut the effect roughly in half (6.3 points).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-11794","stance":"supports"}],"url":"https://rsiratchet.com/claims/preregistered-experiment-1-500-uk-adults-persuasive-chatbot-shifted"},{"ref":"CLM-0317","id":"preregistered-us-experiment-n-2-144-where-participants-discussed","capability":"persuasion","technique":null,"statement":"In a preregistered US experiment (N=2,144) where participants discussed an economic-policy misconception with ChatGPT over three rounds, a brief message that the model was biased against their party reduced belief correction by 28% (CI 16-40%) relative to a neutral control, with transcripts showing more pushback and less receptive engagement.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2602-18092","stance":"supports"}],"url":"https://rsiratchet.com/claims/preregistered-us-experiment-n-2-144-where-participants-discussed"},{"ref":"CLM-0318","id":"presenting-syllogisms-euler-or-linear-diagrams-instead-text-or","capability":"logical-deduction-robustness","technique":null,"statement":"Presenting syllogisms as Euler or linear diagrams instead of text or logical notation did not consistently improve two models' accuracy on 285 problems; both did well on entailment and contradiction items but poorly on neutral ones (GPT-4o-mini under 30% on neutral cases), and made systematic conversion errors, treating 'all A are B' as if it implied 'all B are A'.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2607-23513","stance":"supports"}],"url":"https://rsiratchet.com/claims/presenting-syllogisms-euler-or-linear-diagrams-instead-text-or"},{"ref":"CLM-0469","id":"pretrained-2d-vision-language-action-manipulation-policies-adding-3d-geometry-only","capability":"embodied-control","technique":"condition-action-denoising-raw-3d-semantic-point-field","statement":"For pretrained 2D vision-language-action manipulation policies, adding 3D geometry only on the perception side is not enough: conditioning the action denoiser directly on a raw 3D semantic point field — at the specific diffusion blocks where linear probes show end-effector-coordinate error rising — raises bimanual task success beyond perception-side 3D fusion alone, and encoding the point cloud with a PointNet++-style encoder instead of keeping raw points lowers success.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-24525","stance":"supports"}],"url":"https://rsiratchet.com/claims/pretrained-2d-vision-language-action-manipulation-policies-adding-3d-geometry-only"},{"ref":"CLM-0470","id":"pretrained-vision-language-action-policy-fine-tuned-sequentially-stream-real-robot-manipulation-tasks","capability":"embodied-control","technique":"freeze-normalization-statistics-pre-stream-calibration-set","statement":"When a pretrained vision-language-action policy is fine-tuned sequentially on a stream of real-robot manipulation tasks, the choice of action-normalization statistics is itself a determinant of continual-learning success: statistics that are task-derived, first-task-derived, or recomputed each stage produce forgetting or poor transfer, while statistics estimated once from a small task-independent calibration set covering the embodiment's motion range and then frozen give the best acquisition, retention and task-order stability.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21358","stance":"supports"}],"url":"https://rsiratchet.com/claims/pretrained-vision-language-action-policy-fine-tuned-sequentially-stream-real-robot-manipulation-tasks"},{"ref":"CLM-0335","id":"production-ads-ranking-stack-agent-chains-hypothesis-generation-experiment-execution","capability":"harness-self-improvement","technique":null,"statement":"In a production ads-ranking stack, an agent that chains hypothesis generation, experiment execution, result analysis and a shared knowledge store, with human checkpoints at stage boundaries, delivered several times the completed ML iterations per engineer-week of the manual baseline and a 2.6% relative offline improvement on the headline model from multi-source exploration versus 0.4% from single-hypothesis runs; a semi-automated variant that automated phases without chaining them gained much less, so the leverage came from removing engineer-mediated handoffs, not from automating any one phase, and the authors report reliability was governed more by the orchestration harness (skill coverage, statistical rigor of evaluation, resilience to infrastructure noise) than by the base model's reasoning.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08248","stance":"supports"}],"url":"https://rsiratchet.com/claims/production-ads-ranking-stack-agent-chains-hypothesis-generation-experiment-execution"},{"ref":"CLM-0121","id":"program-aided-computation-moves-errors-to-translation","capability":"arithmetic","technique":"program-aided-computation","statement":"Offloading the calculation to generated code removes the arithmetic as a source of error but does not remove error — what remains is the translation from problem to program, which is now the only thing the model is doing. Verification has to move there with it.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"active","sources":[{"source":"arxiv-2211-10435","stance":"supports"}],"url":"https://rsiratchet.com/claims/program-aided-computation-moves-errors-to-translation"},{"ref":"CLM-0243","id":"prompt-contains-several-legitimate-conflicting-evidence-passages-models-internally","capability":"fact-verification","technique":"neutral-center-activation-steering-evidence-integration","statement":"When a prompt contains several legitimate but conflicting evidence passages, models internally register the conflict — linear probes on mid-to-late-layer hidden states separate conflicting from consistent inputs at high AUC — while still resolving the conflict badly at the output, so the failure is in evidence integration rather than in conflict detection.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03148","stance":"supports"}],"url":"https://rsiratchet.com/claims/prompt-contains-several-legitimate-conflicting-evidence-passages-models-internally"},{"ref":"CLM-0627","id":"prompt-imposes-k-deterministically-verifiable-constraints-once-per-constraint-pass","capability":"procedure-following","technique":null,"statement":"When a prompt imposes k deterministically verifiable constraints at once, per-constraint pass rate declines only gently with k while the probability of satisfying all k collapses multiplicatively, because constraint failures within a response are nearly independent rather than pairwise-interfering — so joint reliability breaks down beyond roughly 5-6 simultaneous constraints and cannot be rescued by choosing or ordering which constraints are combined.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-12426","stance":"supports"}],"url":"https://rsiratchet.com/claims/prompt-imposes-k-deterministically-verifiable-constraints-once-per-constraint-pass"},{"ref":"CLM-0122","id":"prompt-injection-defenses-report-near-zero-attack-success-short-context-benchmarks","capability":"prompt-injection","technique":"structured-prompt-separation","statement":"Prompt injection defenses that report near-zero attack success on short-context benchmarks lose most of that protection when the injected instruction sits inside a document of thousands to tens of thousands of tokens: fine-tuned separation defenses and detect-localize-remove pipelines still let a large share of injections through on paper review, resume screening, code review and email threads.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28411","stance":"supports"}],"url":"https://rsiratchet.com/claims/prompt-injection-defenses-report-near-zero-attack-success-short-context-benchmarks"},{"ref":"CLM-0471","id":"prompt-pool-llm-time-series-forecasting-where-top-k-semantic-anchors","capability":"forecasting","technique":"penalize-cross-batch-anchor-over-use-per-query","statement":"In prompt-pool LLM time series forecasting where top-K semantic anchors are retrieved by cosine similarity, the harmful redundancy is concentration of anchor usage accumulated across training batches, not redundancy among the anchors picked for a single forecast: diversifying within a retrieval (MMR) never beat relevance-only selection on four dataset-horizon cells, while an EMA usage-tracking coverage penalty in the training loss lowered MSE on most of six long-term benchmarks.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22977","stance":"supports"}],"url":"https://rsiratchet.com/claims/prompt-pool-llm-time-series-forecasting-where-top-k-semantic-anchors"},{"ref":"CLM-0472","id":"prompted-text-to-sql-agent-blanket-output-shaping-instruction-system-prompt-here","capability":"formal-query-synthesis","technique":"drop-blanket-row-limit-instructions-text-to-sql","statement":"In a prompted text-to-SQL agent, a blanket output-shaping instruction in the system prompt (here \"LIMIT 10\") is applied by the model to every generated query regardless of the question's intent, truncating answers that legitimately need many rows, so removing it raises execution accuracy most on questions whose correct result set is large.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22917","stance":"supports"}],"url":"https://rsiratchet.com/claims/prompted-text-to-sql-agent-blanket-output-shaping-instruction-system-prompt-here"},{"ref":"CLM-0319","id":"prompting-model-behave-if-knowledge-ended-earlier-date-works","capability":"factual-recall","technique":null,"statement":"Prompting a model to behave as if its knowledge ended at an earlier date works for facts asked about directly (about 82% unlearning success) and for word-sense shifts (about 70%), but fails when the post-cutoff knowledge is only causally connected to the question (about 19%), so a simulated cutoff cannot be trusted to remove contamination from temporal prediction tests.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2510-02340","stance":"supports"}],"url":"https://rsiratchet.com/claims/prompting-model-behave-if-knowledge-ended-earlier-date-works"},{"ref":"CLM-0320","id":"prompting-task-method-knowledge-model-which-spells-out-what-do-how","capability":"planning","technique":"task-method-knowledge-prompting","statement":"Prompting with a Task-Method-Knowledge model, which spells out what to do, how, and why in a hierarchical structure, lifted o1 on the obfuscated Random Blocksworld set of PlanBench from about 32% to 97%, and o1-mini from about 9% to 27%, gains far beyond the roughly 1% that zero-shot chain-of-thought gave on PlanBench in earlier work; the smaller model did not benefit on the Mystery variant.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2602-03900","stance":"supports"}],"url":"https://rsiratchet.com/claims/prompting-task-method-knowledge-model-which-spells-out-what-do-how"},{"ref":"CLM-0123","id":"putting-relevant-passages-first-mitigates-position-effects","capability":"long-context-degradation","technique":"reorder-context-by-relevance","statement":"Where a passage sits in a long input changes how much the model uses it — accuracy is highest when the needed information is at the very start or the very end and lowest when it is in the middle — so ordering retrieved passages to put the most relevant ones first is a real lever on accuracy.","kind":"mechanism","backing_strength":"single-paper","contested":true,"status":"active","sources":[{"source":"arxiv-2307-03172","stance":"supports"},{"source":"arxiv-2609-03148","stance":"contests"},{"source":"anthropic-2026-context-engineering-claude-5","stance":"contests"}],"url":"https://rsiratchet.com/claims/putting-relevant-passages-first-mitigates-position-effects"},{"ref":"CLM-0124","id":"questions-whose-premises-false-or-internally-contradictory-prompting-mid-size","capability":"hallucination","technique":null,"statement":"On questions whose premises are false or internally contradictory, prompting mid-size instruct models (Qwen2.5-7B, LLaMA-3.1-8B, Gemma3-12B) to reason step by step can lower accuracy below plain answering, because the reasoning chain elaborates from the flawed premise instead of challenging it.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25894","stance":"supports"}],"url":"https://rsiratchet.com/claims/questions-whose-premises-false-or-internally-contradictory-prompting-mid-size"},{"ref":"CLM-0125","id":"raw-llm-judges-show-position-bias-gpt-4","capability":"judge-position-bias","technique":null,"statement":"As a raw pairwise judge with no order-swap mitigation, GPT-4 can have its verdict flipped simply by swapping which candidate answer is shown first; applying a position-swap protocol restores agreement with careful human raters to close to human-human agreement levels.","kind":"observation","backing_strength":"replicated","contested":false,"status":"active","sources":[{"source":"arxiv-2305-17926","stance":"supports"},{"source":"arxiv-2306-05685","stance":"supports"}],"url":"https://rsiratchet.com/claims/raw-llm-judges-show-position-bias-gpt-4"},{"ref":"CLM-0628","id":"react-agents-appworld-resampling-each-recorded-inference-point-offline","capability":"harness-self-improvement","technique":"target-guidelines-flip-prone-steps-found-resampling","statement":"For ReAct agents on AppWorld, resampling each recorded inference point offline (N=30, temperature 0.5) to find steps whose responses vary across samples, then writing short natural-language guidelines targeting those steps into retrievable episodic memory, raises the fraction of tasks passing all five repeat runs by +16 points same-task and +13 points similar-task with GPT-4.1 (+6 and +8.7 with GPT-OSS-120B) without lowering mean per-run accuracy.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08832","stance":"supports"}],"url":"https://rsiratchet.com/claims/react-agents-appworld-resampling-each-recorded-inference-point-offline"},{"ref":"CLM-0321","id":"real-c-c-faults-reproducing-tests-24-current-models","capability":"debugging","technique":null,"statement":"On real C/C++ faults with reproducing tests, 24 current models repair only about 11% of the 248 bugs and about 6% of the 102 vulnerabilities, far below their Java and Python repair rates on comparable benchmarks, and fine-tuning on the benchmark's commit corpus raised Pass@k in 21 of 28 comparisons by an average 85% relative, so the gap is partly a training-data gap.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2510-11059","stance":"supports"}],"url":"https://rsiratchet.com/claims/real-c-c-faults-reproducing-tests-24-current-models"},{"ref":"CLM-0217","id":"real-world-online-rl-post-training-large-pretrained-vision-language-action-policies-high-precision","capability":"embodied-control","technique":null,"statement":"For real-world online RL post-training of large pretrained vision-language-action policies on high-precision manipulation, optimizing the action head toward a paired advantage margin against a frozen reference action, rather than maximizing absolute Q values, yields higher autonomous success on long multi-step tasks, because absolute-Q maximization amplifies early optimistic value errors into policy drift that behavior-cloning regularization alone does not stop.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04355","stance":"supports"}],"url":"https://rsiratchet.com/claims/real-world-online-rl-post-training-large-pretrained-vision-language-action-policies-high-precision"},{"ref":"CLM-0126","id":"realistic-web-tasks-explicit-goals-best-model-completed","capability":"procedure-following","technique":null,"statement":"On realistic web tasks with explicit goals, the best model completed only a small fraction end to end, far below human performance.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2307-13854","stance":"supports"}],"url":"https://rsiratchet.com/claims/realistic-web-tasks-explicit-goals-best-model-completed"},{"ref":"CLM-0127","id":"recursive-summarization-trades-detail-for-continuity","capability":"long-term-memory","technique":"recursive-summarization","statement":"A running summary buys continuity across a long conversation at the cost of specifics, because it is lossy compression by construction — so it helps with \"what were we doing\" and not with \"what exactly did I say in turn 40.\" Keeping the raw turns retrievable alongside the summary covers the second case.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"active","sources":[{"source":"arxiv-2308-15022","stance":"supports"},{"source":"arxiv-2410-10813","stance":"supports"}],"url":"https://rsiratchet.com/claims/recursive-summarization-trades-detail-for-continuity"},{"ref":"CLM-0128","id":"recursively-summarizing-dialogue-into-running-memory-improves-consistency","capability":"long-term-memory","technique":null,"statement":"Recursively summarizing the dialogue into a running memory improves consistency in long conversations.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2308-15022","stance":"supports"}],"url":"https://rsiratchet.com/claims/recursively-summarizing-dialogue-into-running-memory-improves-consistency"},{"ref":"CLM-0277","id":"refusal-learned-alignment-tuning-lives-low-rank-subspace-hidden-state-activations","capability":"fine-tuning-retention","technique":"geometric-fragility-score-safety-subspace","statement":"Refusal learned by alignment tuning lives in a low-rank subspace of hidden-state activations that recurs across model families; ablating directions in that subspace lowers harmful-request refusal, and a scalar computed from the subspace's geometry before any fine-tuning (the Geometric Fragility Score) picks out which initially safe model will keep the most refusal after small benign LoRA fine-tuning.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-22676","stance":"supports"}],"url":"https://rsiratchet.com/claims/refusal-learned-alignment-tuning-lives-low-rank-subspace-hidden-state-activations"},{"ref":"CLM-0473","id":"region-guided-contrastive-decoding-medical-vlms-effectiveness-guidance-depends-how","capability":"visual-grounding","technique":"pick-contrast-region-counterfactual-occlusion-penalising-size","statement":"In region-guided contrastive decoding for medical VLMs, the effectiveness of the guidance depends on how much of the image the guiding mask covers: coarse ground-truth boxes (mean ~39% coverage on SLAKE) suppress so much of the unconditional branch that the contrast weakens, while compact automatically discovered masks (~7% coverage) selected by the drop in answer likelihood under occlusion give a sharper signal and higher closed-ended accuracy.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-17427","stance":"supports"}],"url":"https://rsiratchet.com/claims/region-guided-contrastive-decoding-medical-vlms-effectiveness-guidance-depends-how"},{"ref":"CLM-0129","id":"removing-overconstraint-from-system-prompts-stopped-costing-accuracy","capability":"context-hygiene","technique":null,"statement":"Guardrail instructions written to stop older models doing the wrong thing — blanket prohibitions, repeated warnings, defensive defaults — became dead weight in the Claude 5 generation: Anthropic removed over 80% of Claude Code's system prompt for Opus 5 and Fable 5 with no measurable loss on its own coding evaluations, on the reading that the constraints now conflict with each other and with user intent more often than they prevent harm.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"anthropic-2026-context-engineering-claude-5","stance":"supports"}],"url":"https://rsiratchet.com/claims/removing-overconstraint-from-system-prompts-stopped-costing-accuracy"},{"ref":"CLM-0629","id":"repair-production-crash-bugs-large-java-enterprise-systems-feeding","capability":"debugging","technique":"rank-crash-clustered-suspicious-files-llm-repair-context","statement":"For repair of production crash bugs in large Java enterprise systems, feeding an LLM the source of the top-3 suspicious files plus a representative stack trace localizes the developer-modified code in about 71% of responses and yields a semantically valid patch in about 52%, and adding an explicit ranking of suspicious files and methods helps only when the full stack trace is withheld (per-response repair 28% to 47% with stack messages only) while slightly lowering performance when the stack trace is already present.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04483","stance":"supports"}],"url":"https://rsiratchet.com/claims/repair-production-crash-bugs-large-java-enterprise-systems-feeding"},{"ref":"CLM-0130","id":"repository-context-files-do-not-raise-task-success-on-benchmark-coding","capability":"procedure-following","technique":"guide-file","statement":"Giving a coding agent a repository context file — AGENTS.md, CLAUDE.md — does not raise its success rate on benchmark coding tasks and costs about 20% more inference: across 4 agents, 2 benchmarks and 3 conditions, LLM-generated files hurt slightly in 5 of 8 settings while developer-written ones gained 2.4% (p=0.21), and agents obey the files — which is why they spend more — so the files do not carry success-relevant information rather than being ignored.","kind":"observation","backing_strength":"single-paper","contested":true,"status":"pending-review","sources":[{"source":"arxiv-2602-11988","stance":"supports"},{"source":"hashimoto-2026-ai-adoption-journey","stance":"contests"},{"source":"synthesis-2026-harness-engineering-playbook","stance":"contests"}],"url":"https://rsiratchet.com/claims/repository-context-files-do-not-raise-task-success-on-benchmark-coding"},{"ref":"CLM-0474","id":"repository-grounded-llm-infrastructure-engineering-tasks-frontier-agents-score-far-below","capability":"long-horizon-execution","technique":null,"statement":"On repository-grounded LLM-infrastructure engineering tasks, frontier agents score far below ceiling (best 36.53% overall), and every one of eight models scores lower on long-horizon multi-file implementation than on single-file kernel completion, even though the full repository and test cases are visible.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10226","stance":"supports"}],"url":"https://rsiratchet.com/claims/repository-grounded-llm-infrastructure-engineering-tasks-frontier-agents-score-far-below"},{"ref":"CLM-0244","id":"repository-level-issue-fixing-tasks-swe-bench-lite-verified-pro-spending-effort","capability":"code-generation","technique":"requirements-refinement-before-coding","statement":"On repository-level issue-fixing tasks (SWE-bench Lite/Verified/Pro), spending effort upstream — expanding the issue text into a structured requirement spec from retrieved repo context, then testing and refining that spec via generated code and tests — raises resolve rate over agents that consume the raw issue description, and raises the share of patches that even apply cleanly by a larger margin than it raises correctness.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00568","stance":"supports"}],"url":"https://rsiratchet.com/claims/repository-level-issue-fixing-tasks-swe-bench-lite-verified-pro-spending-effort"},{"ref":"CLM-0131","id":"repository-level-issue-fixing-tasks-swe-bench-liteverifiedpro-spending-effort-upstream-expanding","capability":"code-generation","technique":null,"statement":"On repository-level issue-fixing tasks (SWE-bench Lite/Verified/Pro), spending effort upstream — expanding the issue text into a structured requirement spec from retrieved repo context, then testing and refining that spec via generated code and tests — raises resolve rate over agents that consume the raw issue description, and raises the share of patches that even apply cleanly by a larger margin than it raises correctness.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00568","stance":"supports"}],"url":"https://rsiratchet.com/claims/repository-level-issue-fixing-tasks-swe-bench-liteverifiedpro-spending-effort-upstream-expanding"},{"ref":"CLM-0132","id":"reranking-sampled-verilog-candidates-both-generated-testbench-execution-signal-llm","capability":"code-generation","technique":null,"statement":"When reranking sampled Verilog candidates with both a generated-testbench execution signal and an LLM judge, letting the judge see the execution outcomes in its prompt lowers final Pass@1 relative to acquiring the two signals separately and combining them only as a weighted score at selection time, because the judge anchors on the test verdicts (approving all-pass candidates with uncovered logic errors, over-penalizing failures caused by bad testbenches) and stops contributing information the tests miss.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22938","stance":"supports"}],"url":"https://rsiratchet.com/claims/reranking-sampled-verilog-candidates-both-generated-testbench-execution-signal-llm"},{"ref":"CLM-0133","id":"research-idea-forecasts-scored-whether-later-paper-matches-them-under","capability":"evaluation-validity","technique":null,"statement":"When research-idea forecasts are scored by whether a later paper matches them under an LLM judge rubric, higher scores partly reflect broader, less specific ideas: the backbone with the higher hit rate also produced measurably more general forecasts and had more retrieved candidate papers pass the gate, and tightening the specificity threshold does not separate the two.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00747","stance":"supports"}],"url":"https://rsiratchet.com/claims/research-idea-forecasts-scored-whether-later-paper-matches-them-under"},{"ref":"CLM-0475","id":"research-level-theorem-proving-tasks-drawn-recent-focs-stoc-soda-papers","capability":"long-horizon-execution","technique":"stage-proof-explore-gate-decompose-dependency-re-verify","statement":"On research-level theorem-proving tasks drawn from recent FOCS/STOC/SODA papers, wrapping a frontier model in a staged harness that explores strategies, gates when to decompose, solves section-level subproblems along a dependency graph, and routes verifier findings back to the affected section raises graded proof accuracy far above prompting the same model directly (30.3% to 54.0% for Gemini 3.1 Pro on TCS-Bench), and two such runs with different backbones make complementary errors, so selecting between them by sampled critiques reaches 71.0% against an oracle ceiling of 77.3%.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15983","stance":"supports"}],"url":"https://rsiratchet.com/claims/research-level-theorem-proving-tasks-drawn-recent-focs-stoc-soda-papers"},{"ref":"CLM-0218","id":"response-level-hallucination-detection-against-reference-documents-having-model-first","capability":"hallucination","technique":null,"statement":"For response-level hallucination detection against reference documents, having the model first convert the reference into a relational SQL database and then check response claims by executing SELECT queries raises F1 over zero-shot direct prediction mainly by cutting false positives, and the gain shrinks or vanishes on backbones with weak text-to-SQL ability.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05025","stance":"supports"}],"url":"https://rsiratchet.com/claims/response-level-hallucination-detection-against-reference-documents-having-model-first"},{"ref":"CLM-0245","id":"response-level-hallucination-detection-retrieval-augmented-settings-having-model-first-convert","capability":"hallucination","technique":"verify-responses-against-llm-built-sql-database-sources","statement":"For response-level hallucination detection in retrieval-augmented settings, having the model first convert the reference documents into a relational SQL database and then check the response against SELECT query results raises precision and macro F1 over zero-shot direct prediction, over atomic-claim decomposition, and over flat JSON key-value extraction, without any task-specific fine-tuning.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05025","stance":"supports"}],"url":"https://rsiratchet.com/claims/response-level-hallucination-detection-retrieval-augmented-settings-having-model-first-convert"},{"ref":"CLM-0134","id":"retrieval-augmented-answers-splitting-answer-into-claims-checking-each-against","capability":"hallucination","technique":"atomic-fact-checking","statement":"In retrieval-augmented answers, splitting an answer into claims, checking each against the retrieved source, and then acting on the flagged claims reduces the share of answers judged to contain unsupported content — but the strategies trade grounding against preservation: deleting unsupported claims reduces unsupported content most while retaining the least original text, and rewriting retains the most while reducing least.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29307","stance":"supports"}],"url":"https://rsiratchet.com/claims/retrieval-augmented-answers-splitting-answer-into-claims-checking-each-against"},{"ref":"CLM-0219","id":"retrieval-augmented-personalization-frozen-8-9b-generator-prepending-more-retrieved-user-history","capability":"long-term-memory","technique":null,"statement":"In retrieval-augmented personalization with a frozen 8-9B generator, prepending more retrieved user-history records does not monotonically improve quality — on the six LaMP tasks a fixed top-k profile peaked at k=5 and lost 0.8 quality points at k=10 while the share of cases made worse than an empty profile rose from 23.2% to 31.7% — so a learned per-request stopping decision beats any single fixed k.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08180","stance":"supports"}],"url":"https://rsiratchet.com/claims/retrieval-augmented-personalization-frozen-8-9b-generator-prepending-more-retrieved-user-history"},{"ref":"CLM-0135","id":"retrieval-moves-the-trust-boundary-rather-than-removing-it","capability":"hallucination","technique":"retrieval-augmented-generation","statement":"Grounding answers in retrieved documents reduces unsupported generation but transfers the trust problem to the retrieval channel: whatever lands in context is treated as reliable, which shows up both as following instructions planted in retrieved content and as deferring to whichever account is better represented when sources disagree.","kind":"mechanism","backing_strength":"replicated","contested":false,"status":"active","sources":[{"source":"arxiv-2302-12173","stance":"supports"},{"source":"arxiv-2305-13300","stance":"supports"}],"url":"https://rsiratchet.com/claims/retrieval-moves-the-trust-boundary-rather-than-removing-it"},{"ref":"CLM-0476","id":"retrieval-store-holds-both-correct-answer-common-misconception-same","capability":"hallucination","technique":"retrieval-augmented-generation","statement":"When a retrieval store holds both the correct answer and a common misconception for the same question, injecting all retrieved memories raises the hallucination rate above a memory-free baseline for weaker models (53.0% vs 23.0% on TruthfulQA with gemma-4-E4B-it), while a stronger backbone (deepseek-v4-flash) instead shows RAG below its own no-memory rate, so conflict-driven amplification is concentrated in models susceptible to memory injection.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22043","stance":"supports"}],"url":"https://rsiratchet.com/claims/retrieval-store-holds-both-correct-answer-common-misconception-same"},{"ref":"CLM-0136","id":"retrieval-summarization-task-pulls-definitions-surface-term-match-general","capability":"hallucination","technique":"retrieval-augmented-generation","statement":"When retrieval for a summarization task pulls definitions by surface term match from a general encyclopedia, ambiguous terms fetch the wrong sense and the model incorporates that unsupported background into its output, lowering claim-level factuality below the no-retrieval baseline even when the prompt tells it to ignore unrelated retrieved content.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02396","stance":"supports"}],"url":"https://rsiratchet.com/claims/retrieval-summarization-task-pulls-definitions-surface-term-match-general"},{"ref":"CLM-0137","id":"reversal-curse-shows-up-in-a-deployed-model-gpt-4","capability":"reversal-curse","technique":null,"statement":"GPT-4 could reliably name a well-known celebrity's parent, but was far less reliable naming the celebrity when given the parent — the same directional-recall asymmetry demonstrated in smaller fine-tuned models, showing up in a deployed model that wasn't fine-tuned for the test.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2309-12288","stance":"supports"}],"url":"https://rsiratchet.com/claims/reversal-curse-shows-up-in-a-deployed-model-gpt-4"},{"ref":"CLM-0477","id":"reward-models-scored-sequence-log-probability-contextual-preference-pairs-abstentions","capability":"critique","technique":"length-balance-rejected-responses-preference-data","statement":"When reward models are scored by sequence log-probability on contextual preference pairs, abstentions are systematically shorter than the unsupported answers they are preferred over, so measured grounding/abstention judgement partly tracks response length; rewriting rejected and refusal responses to match the preferred ones in length while keeping the grounding failure raises pairwise accuracy far more than adding more in-domain legal preference data (up to +25.6 points over baseline for Ministral-8B).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14739","stance":"supports"}],"url":"https://rsiratchet.com/claims/reward-models-scored-sequence-log-probability-contextual-preference-pairs-abstentions"},{"ref":"CLM-0333","id":"reward-related-information-reasoning-model-s-hidden-states-concentrated-sparse","capability":"error-detection","technique":"probe-derived-value-signal-free-process-reward","statement":"Reward-related information in a reasoning model's hidden states is concentrated in a sparse set of neurons: a linear probe trained with a temporal-difference objective finds 'value neurons' (under 1% per layer) whose activations predict expected correctness, and zeroing them in one layer drops MATH500 accuracy from 75% to an average 20% (as low as 1%), while zeroing the same number of random or magnitude-selected neurons costs 1-9 points; the step-to-step change in the probed value ('dopamine neurons', a stated analogy) used as a process reward for inference-time search lifted MATH500 from 72.2 to 77.8, above an implicit-PRM baseline at 75.0.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2602-00986","stance":"supports"}],"url":"https://rsiratchet.com/claims/reward-related-information-reasoning-model-s-hidden-states-concentrated-sparse"},{"ref":"CLM-0138","id":"rewarding-sycophantic-behavior-generalizes-more-serious-specification-gaming","capability":"sycophancy","technique":null,"statement":"Rewarding sycophantic behavior generalizes to more serious specification gaming in a small fraction of cases.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2406-10162","stance":"supports"}],"url":"https://rsiratchet.com/claims/rewarding-sycophantic-behavior-generalizes-more-serious-specification-gaming"},{"ref":"CLM-0322","id":"rewriting-harmful-request-emotional-or-narrative-style-fearful-curious","capability":"jailbreak-robustness","technique":"style-neutralization-before-safety-model-sees-prompt","statement":"Rewriting a harmful request in an emotional or narrative style (fearful, curious, compassionate) while keeping its intent raises jailbreak success by up to 57 percentage points across 16 instruction-tuned models, contextual rewrites beating fixed templates, and the vulnerability does not shrink with model scale; a preprocessing step that has a second model strip stylistic cues cuts the success rate substantially.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2511-10519","stance":"supports"}],"url":"https://rsiratchet.com/claims/rewriting-harmful-request-emotional-or-narrative-style-fearful-curious"},{"ref":"CLM-0630","id":"rl-trained-multimodal-image-quality-models-applying-reasoning-reward-only-reasoning","capability":"explanation-faithfulness","technique":"reward-reasoning-whether-editing-improves-image","statement":"In RL-trained multimodal image-quality models, applying the reasoning reward only to reasoning tokens and the human-rating reward only to rating tokens keeps rating alignment at the level of rating-only training while turning reasoning-guided edit quality gain positive, and it prevents the collapse to a single universal edit that an unmasked shared reward produces despite that variant's larger raw gain.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18579","stance":"supports"}],"url":"https://rsiratchet.com/claims/rl-trained-multimodal-image-quality-models-applying-reasoning-reward-only-reasoning"},{"ref":"CLM-0478","id":"rl-trained-multimodal-image-quality-models-supervising-reasoning-rating-reward-alone","capability":"explanation-faithfulness","technique":"reward-reasoning-whether-editing-improves-image","statement":"In RL-trained multimodal image-quality models, supervising reasoning with the rating reward alone raises rating alignment while making the stated reasoning less faithful — rating-only training reached an average PLCC/SRCC of 0.836/0.815 yet its reasoning-guided edits lowered judged image quality (mean gain -0.260, worse than the untrained baseline's -0.136).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18579","stance":"supports"}],"url":"https://rsiratchet.com/claims/rl-trained-multimodal-image-quality-models-supervising-reasoning-rating-reward-alone"},{"ref":"CLM-0631","id":"rl-training-central-agent-multi-agent-harness-1-5b-3b-qwen2-5-backbones","capability":"multi-model-coordination","technique":"train-only-central-coordinating-agent-freeze-sub-agents","statement":"When RL-training a central-agent multi-agent harness with 1.5B-3B Qwen2.5 backbones on multi-hop QA, optimizing only the coordinating central policy and leaving sub-agents frozen scores higher than jointly training central and sub-agent policies (average F1 47.79 vs 45.36, accuracy 42.88 vs 40.49 at 3B).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29641","stance":"supports"}],"url":"https://rsiratchet.com/claims/rl-training-central-agent-multi-agent-harness-1-5b-3b-qwen2-5-backbones"},{"ref":"CLM-0269","id":"rlhf-tuned-chat-models-raw-conditional-probabilities-consistently-overconfident-confidence","capability":"abstention-and-self-knowledge","technique":"verbalized-confidence-after-listing-alternatives","statement":"For RLHF-tuned chat models the raw conditional probabilities are consistently overconfident, and a confidence the model states in words or as a number in its output is better calibrated, often halving expected calibration error on TriviaQA, SciQ and TruthfulQA; asking the model to list several candidate answers before giving confidences improves it further, while chain-of-thought does not.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2305-14975","stance":"supports"}],"url":"https://rsiratchet.com/claims/rlhf-tuned-chat-models-raw-conditional-probabilities-consistently-overconfident-confidence"},{"ref":"CLM-0479","id":"role-play-assigned-character-models-usually-recognize-query-lies-outside","capability":"abstention-and-self-knowledge","technique":"separate-boundary-awareness-boundary-compliance-probes","statement":"In role-play with an assigned character, models usually recognize that a query lies outside the character's knowledge boundary yet fail to abstain when asked a concrete factual question about it, so the dominant failure is compliance rather than recognition.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01352","stance":"supports"}],"url":"https://rsiratchet.com/claims/role-play-assigned-character-models-usually-recognize-query-lies-outside"},{"ref":"CLM-0139","id":"roughly-forty-percent-copilot-completions-security-relevant-scenarios-were","capability":"secure-coding","technique":null,"statement":"Roughly forty percent of Copilot completions in security-relevant scenarios were vulnerable.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2108-09293","stance":"supports"}],"url":"https://rsiratchet.com/claims/roughly-forty-percent-copilot-completions-security-relevant-scenarios-were"},{"ref":"CLM-0632","id":"routing-each-image-one-several-domain-specialist-vision-backbones-trained","capability":"multi-model-coordination","technique":"route-images-expert-backbones-prompted-mllm","statement":"When routing each image to one of several domain-specialist vision backbones trained on a shared label space, a prompted MLLM router (Gemma-4-12B) reaches nearly the accuracy of a trained neural-network router and beats majority-vote fusion, because routing only needs coarse domain identification, which a general multimodal model already does zero-shot.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-13463","stance":"supports"}],"url":"https://rsiratchet.com/claims/routing-each-image-one-several-domain-specialist-vision-backbones-trained"},{"ref":"CLM-0522","id":"routing-flagged-updates-to-a-removable-adapter-removes-reward-hacking","capability":"reward-hacking","technique":"gradient-routing-adapter","statement":"Routing the gradient updates from monitor-flagged samples into a sacrificial adapter and deleting it after training avoided monitor-exploiting policies in every toy environment tested, improved Qwen3-8B on a coding task without raising its propensity to reward hack, and in a supervised variant trained Qwen3-32B on agentic coding data so that it never reward hacked under prompted elicitation while matching the solve rate of training on perfectly filtered data; detecting only some forms of the hack was enough to suppress the undetected forms as well.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2610-17062","stance":"supports"}],"url":"https://rsiratchet.com/claims/routing-flagged-updates-to-a-removable-adapter-removes-reward-hacking"},{"ref":"CLM-0480","id":"routing-video-production-request-ordered-chain-registered-sub-agents-8b-open","capability":"multi-model-coordination","technique":"train-orchestrator-descriptor-conditioned-routing-plans","statement":"For routing a video-production request to an ordered chain of registered sub-agents, an 8B open model fine-tuned on descriptor-conditioned routing plans (SFT on verified plans plus GRPO rewarding validity, sub-agent coverage F1 and dependency-consistent order) picks chains more accurately than much larger prompted planners, with the gap widest on long, structurally unfamiliar chains.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29814","stance":"supports"}],"url":"https://rsiratchet.com/claims/routing-video-production-request-ordered-chain-registered-sub-agents-8b-open"},{"ref":"CLM-0140","id":"rubric-graded-long-context-tasks-most-gain-context-compilation-harness-comes-extracting","capability":"long-context-degradation","technique":null,"statement":"On rubric-graded long-context tasks, most of the gain from a context- compilation harness comes from extracting the context's rules, exact terms and output spec into an explicit checklist placed in the prompt, not from the executable verifiers built on top of it — and the gain appears only on rule-dense tasks and on models with enough capacity, disappearing or reversing on open-ended tasks and on a small-activation model.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00759","stance":"supports"}],"url":"https://rsiratchet.com/claims/rubric-graded-long-context-tasks-most-gain-context-compilation-harness-comes-extracting"},{"ref":"CLM-0141","id":"safety-guardrail-classifiers-long-context-failure-caused-proportional-dilution-unsafe","capability":"long-context-degradation","technique":null,"statement":"For safety guardrail classifiers, long-context failure is caused by proportional dilution of the unsafe evidence rather than by absolute input length: as the flagged span's share of the input shrinks, attention mass on it falls, the unsafe-minus-safe logit margin compresses, and unsafe recall collapses, while filling the same length with copies of the unsafe text leaves recall largely intact.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27580","stance":"supports"}],"url":"https://rsiratchet.com/claims/safety-guardrail-classifiers-long-context-failure-caused-proportional-dilution-unsafe"},{"ref":"CLM-0278","id":"safety-realignment-after-poisoned-fine-tune-brittle-three-ways-existing","capability":"fine-tuning-retention","technique":"routed-on-policy-distillation-safety-task-teachers","statement":"Safety realignment after a poisoned fine-tune is brittle in three ways: existing defenses (SSRD, RESTA, soft-SFT, rollback) often destroy the downstream skill they were meant to keep, lose more than 30% of their effect when the defender's prompt template differs from the attacker's, and a realigned model can be re-jailbroken by switching the system prompt; distilling on-policy from two frozen teachers, the original model for refusal and the fine-tuned model for the task, routed per token, cut attack success to about 2% under the defender's template while holding task score near pre-attack level.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2607-27081","stance":"supports"}],"url":"https://rsiratchet.com/claims/safety-realignment-after-poisoned-fine-tune-brittle-three-ways-existing"},{"ref":"CLM-0633","id":"same-family-pair-reasoning-instruction-tuned-4b-models-learning-one-interpolation","capability":"cutting-to-the-chase","technique":"learn-per-module-merge-ratios-length-aware-reward","statement":"For a same-family pair of reasoning and instruction-tuned 4B models, learning one interpolation ratio per module (attention block, FFN block, embedding, final norm) with a reward that scores correctness and ranks correct responses by length cuts average generated tokens by about 27% relative to the reasoning model while holding average accuracy, and beats fixed global-ratio merging and CMA-ES search over the same ratios.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25583","stance":"supports"}],"url":"https://rsiratchet.com/claims/same-family-pair-reasoning-instruction-tuned-4b-models-learning-one-interpolation"},{"ref":"CLM-0323","id":"scenario-pits-stated-personal-preference-against-local-cultural-norm","capability":"cultural-competence","technique":null,"statement":"When a scenario pits a stated personal preference against a local cultural norm, models' choice of norm over preference shifts more with the scenario's country (about 8 points) than with the user's age (1) or gender (under 1), instruction tuning shifts it non-uniformly, and while models can match the majority human choice per country they do not reproduce the spread of human answers, with best correlation to human response distributions only about 0.24.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-07877","stance":"supports"}],"url":"https://rsiratchet.com/claims/scenario-pits-stated-personal-preference-against-local-cultural-norm"},{"ref":"CLM-0142","id":"scientific-tasks-where-domain-specific-constraint-general-presentation-constraint-letter","capability":"procedure-following","technique":null,"statement":"On scientific tasks where a domain-specific constraint and a general presentation constraint (letter case, output format, structure) are both stated in the same prompt, multimodal models satisfy the domain-specific constraint more often than the general formatting one, so domain competence does not imply full instruction compliance.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25973","stance":"supports"}],"url":"https://rsiratchet.com/claims/scientific-tasks-where-domain-specific-constraint-general-presentation-constraint-letter"},{"ref":"CLM-0481","id":"search-agent-asked-forecast-future-distribution-rather-than-report","capability":"forecasting","technique":"ask-current-state-then-carry-forward","statement":"When a search agent is asked to forecast a future distribution rather than report the current state over the same cumulative literature archive, the forward-looking target changes what it retrieves — it allocates a smaller share of retrieved documents to the most recent window, which is the more predictive evidence — so carrying the agent's own state estimate forward beats its explicit forecast.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10092","stance":"supports"}],"url":"https://rsiratchet.com/claims/search-agent-asked-forecast-future-distribution-rather-than-report"},{"ref":"CLM-0143","id":"search-based-automated-program-repair-simulink-stateflow-models-replacing-hand-crafted","capability":"code-generation","technique":null,"statement":"In search-based automated program repair of Simulink/Stateflow models, replacing hand-crafted domain-specific mutation operators with an LLM that proposes label rewrites from static component context lowers repair coverage under a fixed wall-clock budget, because unconstrained generative edits widen the candidate space while each candidate costs an expensive simulation and the generator gets no behavioural feedback about why a patch failed.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-19347","stance":"supports"}],"url":"https://rsiratchet.com/claims/search-based-automated-program-repair-simulink-stateflow-models-replacing-hand-crafted"},{"ref":"CLM-0144","id":"secure-coding-package-hallucination-rate-gpt-4","capability":"secure-coding","technique":null,"statement":"GPT-4 recommended non-existent Python and JavaScript packages in only a small percentage of generations — lower than the open models tested in the same study — but the same hallucinated names recurred often enough across runs to be practically exploitable by an attacker who registers them ahead of time.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2406-10279","stance":"supports"}],"url":"https://rsiratchet.com/claims/secure-coding-package-hallucination-rate-gpt-4"},{"ref":"CLM-0482","id":"selecting-among-multiple-generated-sql-candidates-listwise-selector-given","capability":"formal-query-synthesis","technique":"group-candidates-execution-result-then-permute-within-groups","statement":"For selecting among multiple generated SQL candidates, a listwise selector given retrieved structured selection criteria distilled from training question-SQL pairs, with candidates grouped by execution result and permuted only within groups, reaches higher execution accuracy than pointwise, pairwise and multi-selector baselines on the same candidate pools while using far fewer selector calls and tokens.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00834","stance":"supports"}],"url":"https://rsiratchet.com/claims/selecting-among-multiple-generated-sql-candidates-listwise-selector-given"},{"ref":"CLM-0634","id":"self-contained-linguistic-rosetta-stone-puzzles-where-single-context-example","capability":"abstention-and-self-knowledge","technique":"delete-one-context-example-question-uniquely-needs","statement":"On self-contained linguistic Rosetta Stone puzzles where a single context example uniquely carrying a question-required token is deleted, frontier models told to output N/A when context is insufficient almost never abstain, and on some puzzles answer the damaged variant as well as or better than the original.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27756","stance":"supports"}],"url":"https://rsiratchet.com/claims/self-contained-linguistic-rosetta-stone-puzzles-where-single-context-example"},{"ref":"CLM-0145","id":"self-critique-without-grounding-unreliable","capability":"self-repair","technique":null,"statement":"Without an external, ground-truth signal — a failing test, a compiler error, a verifier's output — a model's own critique of its reasoning is not a reliable improvement signal, and asking it to review and revise a correct answer often turns it into a wrong one.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2310-01798","stance":"supports"}],"url":"https://rsiratchet.com/claims/self-critique-without-grounding-unreliable"},{"ref":"CLM-0483","id":"self-evolving-agents-accumulate-reusable-skills-bank-admit-new-skill","capability":"harness-self-improvement","technique":"gate-skills-sampled-coalition-knockouts-then-mask-after","statement":"For self-evolving agents that accumulate reusable skills in a bank and admit a new skill whenever aggregate task reward rises, the accepted skills can be collectively net-negative, because a bank-level reward change cannot identify which skill contributed and a skill's contribution depends on the other skills present and on the deployment domain.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22610","stance":"supports"}],"url":"https://rsiratchet.com/claims/self-evolving-agents-accumulate-reusable-skills-bank-admit-new-skill"},{"ref":"CLM-0635","id":"self-evolving-llm-agents-admit-new-skill-into-persistent-skill","capability":"harness-self-improvement","technique":"gate-skills-sampled-coalition-knockouts-then-mask-after","statement":"For self-evolving LLM agents that admit a new skill into a persistent skill bank whenever aggregate outcome reward rises, the accepted skills are collectively net-negative and the gate cannot do better in principle, because the bank-level gain is one scalar that leaves every incumbent skill's contribution unconstrained; scoring sampled coalition knockouts at the gate, and masking transferred skills whose removal improves retrieval quality on unlabeled target queries, recovers the missing composition and domain dependence.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22610","stance":"supports"}],"url":"https://rsiratchet.com/claims/self-evolving-llm-agents-admit-new-skill-into-persistent-skill"},{"ref":"CLM-0146","id":"self-feedback-improves-generation-quality","capability":"self-repair","technique":null,"statement":"Prompting a model to generate its own feedback on a draft output and revise accordingly improves quality on open-ended generation tasks (dialogue, code, review-writing) over a single-shot attempt.","kind":"mechanism","backing_strength":"single-paper","contested":true,"status":"active","sources":[{"source":"arxiv-2303-17651","stance":"supports"},{"source":"arxiv-2310-01798","stance":"contests"}],"url":"https://rsiratchet.com/claims/self-feedback-improves-generation-quality"},{"ref":"CLM-0147","id":"self-reflection-used-training-signal-long-horizon-agent-math-tasks-resetting","capability":"long-horizon-execution","technique":null,"statement":"When self-reflection is used as a training signal for long-horizon agent and math tasks, resetting to the original prompt with a compact reflection patch prepended beats accumulating reflections inside the ongoing trajectory, and overly long reflections (more than ~10 points) hurt compared with 2-5 bullet points.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23493","stance":"supports"}],"url":"https://rsiratchet.com/claims/self-reflection-used-training-signal-long-horizon-agent-math-tasks-resetting"},{"ref":"CLM-0148","id":"self-supervised-training-tool-call-examples-teaches-model-when-how","capability":"tool-use","technique":null,"statement":"Self-supervised training on tool-call examples teaches a model when and how to call simple tools, narrowing the gap on tasks that need them.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2302-04761","stance":"supports"}],"url":"https://rsiratchet.com/claims/self-supervised-training-tool-call-examples-teaches-model-when-how"},{"ref":"CLM-0270","id":"selfaware-set-unanswerable-questions-no-scientific-consensus-imagination-philosophical","capability":"abstention-and-self-knowledge","technique":null,"statement":"On SelfAware, a set of unanswerable questions (no scientific consensus, imagination, philosophical, too many variables, subjective) paired with answerable ones, models' ability to flag the unanswerable ones rises with parameter count, with instruction tuning and with in-context examples, yet the best model tested (GPT-4, F1 about 75) stays well below the human reference of about 85, so self-knowledge is present but incomplete.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2305-18153","stance":"supports"}],"url":"https://rsiratchet.com/claims/selfaware-set-unanswerable-questions-no-scientific-consensus-imagination-philosophical"},{"ref":"CLM-0484","id":"semi-extractive-rag-qa-where-model-prompted-put-quoted-evidence","capability":"hallucination","technique":"mask-decoding-verbatim-context-spans-inside-quotes","statement":"In semi-extractive RAG QA where the model is prompted to put quoted evidence in brackets, the bracketed spans frequently do not appear verbatim in the retrieved context (as low as 0.37 verbatim rate on a medical dataset with Mistral-Nemo), and masking logits during bracketed spans to only tokens continuing a verbatim context span raises verbatim fidelity to near 1.0 across Llama-3.1-8B, Mistral-Nemo and Qwen3-4B without lowering ROUGE-L or BERTScore much.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-10046","stance":"supports"}],"url":"https://rsiratchet.com/claims/semi-extractive-rag-qa-where-model-prompted-put-quoted-evidence"},{"ref":"CLM-0149","id":"separating-prompt-data-reserved-delimiters-fine-tuning-structure-blocks","capability":"prompt-injection","technique":null,"statement":"Separating prompt and data with reserved delimiters and fine-tuning on that structure blocks most injections at little cost to utility.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2402-06363","stance":"supports"}],"url":"https://rsiratchet.com/claims/separating-prompt-data-reserved-delimiters-fine-tuning-structure-blocks"},{"ref":"CLM-0246","id":"short-answer-object-existence-questions-open-multimodal-models-llava-v1-5","capability":"visual-grounding","technique":"adversarial-contrastive-fine-tuning-hallucination-flipped-image-pairs","statement":"For short-answer object-existence questions in open multimodal models (LLaVA v1.5, MiniGPT-4, Qwen2.5-VL), object hallucination is driven partly by the visual encoder itself rather than only by language priors: hallucinated cases show lower image-text embedding similarity and inverted attention, and swapping in a weaker or noisier visual encoder lowers POPE accuracy while a stronger one raises it.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00231","stance":"supports"}],"url":"https://rsiratchet.com/claims/short-answer-object-existence-questions-open-multimodal-models-llava-v1-5"},{"ref":"CLM-0150","id":"short-answer-object-existence-questions-open-multimodal-models-llava-v15-minigpt-4","capability":"visual-grounding","technique":null,"statement":"For short-answer object-existence questions in open multimodal models (LLaVA v1.5, MiniGPT-4, Qwen2.5-VL), object hallucination is driven partly by the visual encoder itself rather than only by language priors: hallucinated cases show lower image-text embedding similarity and inverted attention, and swapping in a weaker or noisier visual encoder lowers POPE accuracy while a stronger one raises it.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00231","stance":"supports"}],"url":"https://rsiratchet.com/claims/short-answer-object-existence-questions-open-multimodal-models-llava-v15-minigpt-4"},{"ref":"CLM-0151","id":"shows-llms-hallucinate-api-names-arguments-when-calling","capability":"tool-use","technique":null,"statement":"Shows LLMs hallucinate API names and arguments when calling large API collections from memory, and that retrieving documentation reduces this.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2305-15334","stance":"supports"}],"url":"https://rsiratchet.com/claims/shows-llms-hallucinate-api-names-arguments-when-calling"},{"ref":"CLM-0152","id":"simple-goal-hijacking-prompt-leaking-attacks-succeed-against-production-models","capability":"prompt-injection","technique":null,"statement":"Simple goal-hijacking and prompt-leaking attacks succeed against production models with short adversarial strings.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2211-09527","stance":"supports"}],"url":"https://rsiratchet.com/claims/simple-goal-hijacking-prompt-leaking-attacks-succeed-against-production-models"},{"ref":"CLM-0485","id":"simple-hls-kernel-generation-tasks-85-designs-chstone-machsuite","capability":"code-generation","technique":null,"statement":"On simple HLS kernel generation tasks (85 designs from CHStone, MachSuite, Polybench, Rosetta, C2HLSC), giving a modest open-source model (gpt-oss-120b) a shell with a C++ compiler for self-verification and iterative refinement saturates the benchmark at pass@10 (100% through synthesis) and exceeds 90% at pass@1, so this class of task no longer separates models; repeated sampling helps the weaker model far more (+60 points synthesis pass rate at k=10 for gpt-oss-20b vs +11 for the 120b).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09526","stance":"supports"}],"url":"https://rsiratchet.com/claims/simple-hls-kernel-generation-tasks-85-designs-chstone-machsuite"},{"ref":"CLM-0486","id":"simulated-networks-llm-agents-where-one-or-two-agents","capability":"multi-model-coordination","technique":"audit-agent-influence-belief-probes-exposure-logs","statement":"In simulated networks of LLM agents where one or two agents are instructed to advocate a position, the messages other agents write understate how far their elicited stance has moved: they mirror the persuader's rhetorical framing in most messages while explicit stance-change language is rare, so monitoring only generated text misses belief movement that a repeated token-probability probe detects.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25152","stance":"supports"}],"url":"https://rsiratchet.com/claims/simulated-networks-llm-agents-where-one-or-two-agents"},{"ref":"CLM-0153","id":"simulated-retail-airline-domains-tools-policy-documents-frontier","capability":"tool-use","technique":null,"statement":"In simulated retail and airline domains with tools and policy documents, frontier models succeeded on well under two thirds of tasks and success dropped sharply when the same task had to be passed several times in a row.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2406-12045","stance":"supports"}],"url":"https://rsiratchet.com/claims/simulated-retail-airline-domains-tools-policy-documents-frontier"},{"ref":"CLM-0220","id":"simulated-urban-air-ground-object-search-where-vlm-drives-both","capability":"embodied-control","technique":null,"statement":"In simulated urban air-ground object search where a VLM drives both a UAV and a UGV, wrapping the model in a fixed search-handoff-verify protocol that leaves geometric projection and road-graph path planning to deterministic tools raises task success for backbones with adequate perception while cutting decision calls for all, but does not lift weak small open models, whose success can fall on harder splits.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08402","stance":"supports"}],"url":"https://rsiratchet.com/claims/simulated-urban-air-ground-object-search-where-vlm-drives-both"},{"ref":"CLM-0636","id":"single-agent-procedural-tasks-where-relevant-state-fits-fixed-schema","capability":"long-horizon-execution","technique":"replace-conversation-history-rewritten-execution-state","statement":"For single-agent procedural tasks where the relevant state fits a fixed schema known in advance, replacing the appended conversation transcript with a structured execution state that the model rewrites each step — discarding the reasoning trace once the state patch is validated — holds accuracy flat as the horizon grows while history-appending runtimes degrade, and the gain comes from the structure rather than the shorter prompt.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26263","stance":"supports"}],"url":"https://rsiratchet.com/claims/single-agent-procedural-tasks-where-relevant-state-fits-fixed-schema"},{"ref":"CLM-0324","id":"single-fact-edit-exerts-two-coupled-pressures-knowledge-neighbourhood","capability":"knowledge-updating","technique":"joint-neighbourhood-optimization-edits","statement":"A single fact edit exerts two coupled pressures on its knowledge neighbourhood, propagation to facts that should change and leakage into facts that should not, and planning the edit's target representations jointly under both constraints, with a gate that rejects risky plans before touching parameters, improved propagation and preservation on RippleEdits by at least 7% across backbones over methods that treat the two separately.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2606-01610","stance":"supports"}],"url":"https://rsiratchet.com/claims/single-fact-edit-exerts-two-coupled-pressures-knowledge-neighbourhood"},{"ref":"CLM-0637","id":"single-step-retrosynthesis-real-industrial-target-molecules-frontier-general-llms","capability":"hallucination","technique":"rerank-candidate-routes-executable-chemical-verifier","statement":"On single-step retrosynthesis for real industrial target molecules, frontier general LLMs produce chemically infeasible top-1 routes (unresolved selectivity or no mechanistic basis) more often than specialized retrosynthesis models, even though their reference-matching recall is comparable, so disconnection fluency does not carry selectivity awareness.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05450","stance":"supports"}],"url":"https://rsiratchet.com/claims/single-step-retrosynthesis-real-industrial-target-molecules-frontier-general-llms"},{"ref":"CLM-0487","id":"single-step-tool-calling-agents-near-ceiling-accuracy-unperturbed-calls-does-predict","capability":"tool-use","technique":"stage-aligned-perturbation-cascade-aware-failure-attribution","statement":"For single-step tool-calling agents, near-ceiling accuracy on unperturbed calls does not predict robustness to perturbation, and corrupting or removing the tool's returned evidence hurts far more than perturbing tool names, descriptions, or the user request.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23635","stance":"supports"}],"url":"https://rsiratchet.com/claims/single-step-tool-calling-agents-near-ceiling-accuracy-unperturbed-calls-does-predict"},{"ref":"CLM-0638","id":"single-step-tool-calling-agents-near-perfect-clean-accuracy-does-predict-robustness","capability":"tool-use","technique":null,"statement":"For single-step tool-calling agents, near-perfect clean accuracy does not predict robustness, and the dominant weak point is recovering the answer from perturbed tool returns rather than selecting the tool or binding arguments: across seven models, clean success averaged 0.979 while success under corrupted or missing returned evidence averaged 0.455, with no model above 0.60.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-23635","stance":"supports"}],"url":"https://rsiratchet.com/claims/single-step-tool-calling-agents-near-perfect-clean-accuracy-does-predict-robustness"},{"ref":"CLM-0639","id":"small-1-7b-llm-fine-tuned-act-ensemble-router-assigns","capability":"forecasting","technique":"build-chain-of-thought-supervision-rules-api-calls","statement":"When a small (1.7B) LLM is fine-tuned to act as an ensemble router that assigns per-sample weights across candidate time-series forecasters, chain-of-thought supervision built by rule-templating known oracle weights against extracted temporal features matches or beats CoT generated by a frontier API model, and removing the CoT step costs far more on unseen candidate model pools than on in-distribution ones.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-10149","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-1-7b-llm-fine-tuned-act-ensemble-router-assigns"},{"ref":"CLM-0640","id":"small-4-8b-student-agents-multi-turn-tool-use-tasks-transferring-stronger","capability":"tool-use","technique":"distill-teacher-trajectories-into-workflow-subtask-function-memory","statement":"For small (4-8B) student agents on multi-turn tool-use tasks, transferring a stronger teacher's successful trajectories as flat task-level memory barely helps and can hurt, while splitting the same trajectories into a high-level workflow insight in prose plus concrete executable subtask examples raises accuracy sharply, with the subtask examples carrying most of the gain.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-07169","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-4-8b-student-agents-multi-turn-tool-use-tasks-transferring-stronger"},{"ref":"CLM-0488","id":"small-agent-policy-qwen3-5-4b-repeatedly-deployed-same-task","capability":"harness-self-improvement","technique":"consolidate-only-experience-survives-state-revision","statement":"For a small agent policy (Qwen3.5-4B) repeatedly deployed on the same task family with environment-level feedback, distilling only the trajectories whose textual-state guidance is newly or persistently useful relative to state-free execution raises the state-free policy's accuracy more than distilling all state-conditioned trajectories, and the resulting policy keeps improving over alternating rounds of state editing and consolidation.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08919","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-agent-policy-qwen3-5-4b-repeatedly-deployed-same-task"},{"ref":"CLM-0247","id":"small-instruction-tuned-models-3-4b-answering-closed-form-true-false-questions","capability":"sycophancy","technique":"peer-prediction-bayesian-truth-serum-reward-label-free","statement":"For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25267","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-instruction-tuned-models-3-4b-answering-closed-form-true-false-questions"},{"ref":"CLM-0154","id":"small-instruction-tuned-models-3-4b-answering-closed-form-truefalse-questions-under","capability":"sycophancy","technique":null,"statement":"For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25267","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-instruction-tuned-models-3-4b-answering-closed-form-truefalse-questions-under"},{"ref":"CLM-0337","id":"small-open-models-suffice-for-routine-agent-calls-not-for-long-horizon-planning","capability":"tool-use","technique":"model-routing","statement":"On a 30-task ladder spanning instruction following, tool use, multi-step coordination and long-horizon planning, evaluated over 16 open-weight models from 0.27B to 32B against GPT-5 across 16,542 scored runs, small and mid-sized open models are already sufficient for the short-horizon structured tool use that makes up most calls in real agent pipelines, and the strongest open-weight model matches GPT-5 in aggregate — but the gap reappears on long-horizon planning that needs sustained coordination and constraint tracking, where frontier models still lead and neither side is reliable, and the boundary is not explained by model scale alone.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2605-00334","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-open-models-suffice-for-routine-agent-calls-not-for-long-horizon-planning"},{"ref":"CLM-0489","id":"small-open-weight-backbones-qwen3-4b-driving-long-horizon-remote-sensing-tool-chains","capability":"long-horizon-execution","technique":"verify-each-intermediate-step-against-domain-constraints-then","statement":"For small open-weight backbones (Qwen3-4B) driving long-horizon remote-sensing tool chains, inserting a step-level verifier that checks each intermediate tool output against schema, sub-goal and domain constraints and triggers replanning is the single largest architectural gain, while hierarchical decomposition without domain-aligned tool grouping adds little and can lower final accuracy relative to a flat agent with the same verifier.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30672","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-open-weight-backbones-qwen3-4b-driving-long-horizon-remote-sensing-tool-chains"},{"ref":"CLM-0155","id":"small-open-weight-models-0-6b-12b-bfcl-v3-presenting-tool","capability":"tool-use","technique":null,"statement":"For small open-weight models (0.6B-12B) on BFCL V3, presenting tool definitions as ordinary system-prompt instructions and having the model emit calls as normal assistant text yields higher function-calling accuracy than the model's native tool-call token/schema interface, with the largest gains in the smallest models and on parallel/multiple-call categories.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22472","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-open-weight-models-0-6b-12b-bfcl-v3-presenting-tool"},{"ref":"CLM-0156","id":"small-open-weight-models-1-5-4b-generating-answers-closed-domain-rag","capability":"hallucination","technique":null,"statement":"For small open-weight models (1.5-4B) generating answers in a closed-domain RAG setup, hallucinated generations show a broken hidden-state proximity graph between the retrieved context entity and the generated entity — connectivity below 20% and path lengths above 7 hops, versus near-full connectivity at about 3 hops for grounded answers — giving a training-free detection signal (AUROC 0.89 on Qwen2.5-1.5B).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-17950","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-open-weight-models-1-5-4b-generating-answers-closed-domain-rag"},{"ref":"CLM-0490","id":"small-open-weight-qwen2-5-models-1-5b-3b-doing-multi-step","capability":"hallucination","technique":"score-reasoning-failure-cross-layer-attention-entropy-dispersion","statement":"For small open-weight Qwen2.5 models (1.5B-3B) doing multi-step math reasoning, features derived from the spread of attention entropy across layers at each generation step predict whether the final answer is wrong better than output-logit confidence features, with the margin larger on harder problems (+0.0093 AUC on GSM8K, +0.0759 on MATH-500 at 3B).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-18320","stance":"supports"}],"url":"https://rsiratchet.com/claims/small-open-weight-qwen2-5-models-1-5b-3b-doing-multi-step"},{"ref":"CLM-0157","id":"span-level-hallucination-detection-leaderboards-where-absolute-scores-low-bootstrap-resampling","capability":"evaluation-validity","technique":null,"statement":"On span-level hallucination-detection leaderboards where absolute scores are low, bootstrap resampling of the test set moves top systems across wide rank intervals, so point-estimate ordering does not establish that one detector beats another.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25662","stance":"supports"}],"url":"https://rsiratchet.com/claims/span-level-hallucination-detection-leaderboards-where-absolute-scores-low-bootstrap-resampling"},{"ref":"CLM-0641","id":"speech-llms-long-audio-text-centric-kv-cache-eviction-methods","capability":"long-context-degradation","technique":"compress-audio-kv-cache-around-asr-transcript-anchors","statement":"For speech LLMs on long audio, text-centric KV cache eviction methods (StreamingLLM, SnapKV, PyramidKV, ChunkKV) collapse because raw audio tokens have diffuse attention; anchoring compression on ASR transcripts — pooling each transcribed audio span down to its text-token length and adding the text embedding, while keeping untranscribed audio intact — retains over 90% of full-cache accuracy at a 5-10% cache budget where those baselines retain roughly 55-67% at a looser 25% budget.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-08569","stance":"supports"}],"url":"https://rsiratchet.com/claims/speech-llms-long-audio-text-centric-kv-cache-eviction-methods"},{"ref":"CLM-0325","id":"splitting-task-across-several-agents-helps-long-horizon-tasks-whose","capability":"multi-model-coordination","technique":null,"statement":"Splitting a task across several agents helps on long-horizon tasks whose parts depend on each other sparsely, and hurts on tightly coupled sequential workflows, where generic multi-agent methods fall below a single-agent harness because decomposition fragments the dependency chain and coordination overhead outweighs the context isolation it buys; as single agents improve the collaboration benefit shrinks.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-19759","stance":"supports"}],"url":"https://rsiratchet.com/claims/splitting-task-across-several-agents-helps-long-horizon-tasks-whose"},{"ref":"CLM-0491","id":"spoken-language-models-answering-questions-over-long-dialogue-histories","capability":"audio-understanding","technique":"pick-history-modality-what-question-needs","statement":"For spoken language models answering questions over long dialogue histories, converting the history to transcripts beats keeping it as audio when the answer is a lexical detail (exact slot values across dense turns: mean 71.3% vs 42.6%), but loses when the answer depends on how something was said, where audio history wins for six of seven models and by a larger margin once lexical emotion cues are removed.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22452","stance":"supports"}],"url":"https://rsiratchet.com/claims/spoken-language-models-answering-questions-over-long-dialogue-histories"},{"ref":"CLM-0158","id":"static-code-benchmarks-saturate-need-continuous-refresh","capability":"code-generation","technique":null,"statement":"Static, fixed-problem coding benchmarks saturate and leak into training data over time, so getting an honest read on current code-generation ability requires a benchmark that continuously adds newly-published problems rather than reusing an older fixed set like HumanEval.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2403-07974","stance":"supports"}],"url":"https://rsiratchet.com/claims/static-code-benchmarks-saturate-need-continuous-refresh"},{"ref":"CLM-0642","id":"statutory-tax-calculation-cases-models-asked-solve-input-contains","capability":"abstention-and-self-knowledge","technique":"gate-answer-separate-self-verification-call-contradictions","statement":"On statutory tax calculation cases, models asked to solve an input that contains an injected contradictory fact mostly compute straight through it and return the clean-input answer, while the same model asked only to verify the same input flags most of those contradictions \\u2014 so abstention failure here is a task-framing gap, not missing detection ability.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05928","stance":"supports"}],"url":"https://rsiratchet.com/claims/statutory-tax-calculation-cases-models-asked-solve-input-contains"},{"ref":"CLM-0643","id":"step-level-verification-structured-quantitative-stem-reasoning-training-7b-process","capability":"error-detection","technique":"train-step-verifier-only-symbolically-valid-wrong-steps","statement":"For step-level verification of structured quantitative STEM reasoning, training a 7B process reward model only on hard negatives that are constructed to pass a deterministic symbolic verifier (swapped formula or unit-matched variable, with values recomputed so the math balances), and running the verifier as a hard filter before the PRM scores candidates, detects flawed steps better than monolithic PRMs that must judge arithmetic and logic together.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26329","stance":"supports"}],"url":"https://rsiratchet.com/claims/step-level-verification-structured-quantitative-stem-reasoning-training-7b-process"},{"ref":"CLM-0159","id":"streaming-video-question-answering-frozen-video-llm-appending-retrieved-historical","capability":"video-temporal-reasoning","technique":null,"statement":"For streaming video question answering with a frozen video-LLM, appending retrieved historical frames as extra visual context helps less than optimizing a small set of latent memory tokens at test time and then dropping the retrieved tokens before decoding; the gain is largest on questions requiring backward tracing over past history.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04131","stance":"supports"}],"url":"https://rsiratchet.com/claims/streaming-video-question-answering-frozen-video-llm-appending-retrieved-historical"},{"ref":"CLM-0644","id":"streaming-video-understanding-7-8b-multimodal-backbones-question-relevant-historical-frames","capability":"video-temporal-reasoning","technique":"index-stream-shallow-layer-kvs-answer-deep-only","statement":"For streaming video understanding with 7-8B multimodal backbones, question-relevant historical frames can be retrieved from the keys of only the first few language-Transformer layers (layer ~3-4 of 28-32), so continuous frames need not be prefilled through full model depth: indexing shallow and re-prefilling only the retrieved units at full depth cuts per-frame prefill up to 52x and 20-second end-to-end latency up to 15x while scoring about 1.7 points above the best same-backbone baseline on OVO-Bench and StreamingBench.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02780","stance":"supports"}],"url":"https://rsiratchet.com/claims/streaming-video-understanding-7-8b-multimodal-backbones-question-relevant-historical-frames"},{"ref":"CLM-0255","id":"structured-extraction-text-files-web-pages-spreadsheets-few-user-highlighted","capability":"formal-query-synthesis","technique":"synthesise-extractor-highlighted-examples","statement":"For structured extraction from text files, web pages and spreadsheets, a few user-highlighted example fields are enough to synthesise a reusable extraction program that pulls the remaining fields without further interaction — the locator can be inferred from examples rather than written or re-invoked per record.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"le-gulwani-2014-flashextract","stance":"supports"}],"url":"https://rsiratchet.com/claims/structured-extraction-text-files-web-pages-spreadsheets-few-user-highlighted"},{"ref":"CLM-0160","id":"structured-long-term-memory-substrate-indexes-retrieval-control-logic-held","capability":"long-term-memory","technique":"hierarchical-memory","statement":"When a structured long-term memory substrate (indexes, retrieval, control logic) is held fixed and only the backbone LLM is swapped, accuracy on multi-session memory benchmarks like LongMemEval and LoCoMo moves by a few points while per-query cost varies by roughly an order of magnitude, so recall quality in these settings is set mostly by the memory and retrieval design rather than by model choice.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29606","stance":"supports"}],"url":"https://rsiratchet.com/claims/structured-long-term-memory-substrate-indexes-retrieval-control-logic-held"},{"ref":"CLM-0161","id":"structured-separation-needs-training-and-reduces-rather-than-eliminates","capability":"prompt-injection","technique":"structured-prompt-separation","statement":"Separating instructions from data raises injection resistance substantially, but both published versions get their strength from fine-tuning the model on the separation, and both report improved robustness rather than elimination. Treat it as one layer of defense in depth; the prompt-only variant, without training, has no measured efficacy behind it here.","kind":"mechanism","backing_strength":"replicated","contested":false,"status":"active","sources":[{"source":"arxiv-2402-06363","stance":"supports"},{"source":"arxiv-2404-13208","stance":"supports"}],"url":"https://rsiratchet.com/claims/structured-separation-needs-training-and-reduces-rather-than-eliminates"},{"ref":"CLM-0221","id":"sub-2b-open-weight-models-emitting-json-tool-calls-prompts-already","capability":"tool-use","technique":null,"statement":"For sub-2B open-weight models emitting JSON tool calls from prompts that already name the tool and all argument values, reported success depends almost entirely on a post-hoc recovery layer that strips code fences and extracts a brace-delimited substring: across 1,000 generations only 5 whole raw responses parsed as JSON directly, while the best condition reached 79% after recovery.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07370","stance":"supports"}],"url":"https://rsiratchet.com/claims/sub-2b-open-weight-models-emitting-json-tool-calls-prompts-already"},{"ref":"CLM-0492","id":"suite-verifier-checked-tool-using-react-task-families-arithmetic-sql-analytics","capability":"evaluation-validity","technique":"report-per-family-spread-aggregate-agent-score","statement":"On a suite of verifier-checked tool-using ReAct task families (arithmetic, SQL analytics, security judgment, URL grounding, dependency planning, orchestration, tool policy), a single task-weighted aggregate accuracy hides inverted per-family rankings: the model that ranks last overall can lead both URL-grounding families while scoring lowest on structured SQL and multi-step tool families, which spread the four tested models by more than sixty points.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26867","stance":"supports"}],"url":"https://rsiratchet.com/claims/suite-verifier-checked-tool-using-react-task-families-arithmetic-sql-analytics"},{"ref":"CLM-0162","id":"surveys-causes-kinds-detection-hallucination-across-data-training","capability":"hallucination","technique":null,"statement":"Surveys causes, kinds, and detection of hallucination across data, training, and inference, and finds no complete solution.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2311-05232","stance":"supports"}],"url":"https://rsiratchet.com/claims/surveys-causes-kinds-detection-hallucination-across-data-training"},{"ref":"CLM-0163","id":"swapping-candidate-order-alone-can-flip-judges-preference","capability":"judge-position-bias","technique":null,"statement":"Swapping candidate order alone can flip the judge's preference, so an evaluator can be gamed by ordering.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2305-17926","stance":"supports"}],"url":"https://rsiratchet.com/claims/swapping-candidate-order-alone-can-flip-judges-preference"},{"ref":"CLM-0271","id":"swe-bench-lite-asking-model-name-suspicious-files-issue-plus","capability":"repository-navigation","technique":"hierarchical-localization-tree-then-union-embedding-retrieval","statement":"On SWE-bench Lite, asking the model to name suspicious files from the issue plus a repository tree finds the ground-truth file in about 79% of cases and embedding retrieval over the issue text in about 68%, and taking the union of the two raises file-level localization to about 82% because the two methods miss different files; a fixed three-step narrowing (files, then classes and functions, then edit lines) with no agent loop then supported the best open-source resolve rate at the time (32%).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2407-01489","stance":"supports"}],"url":"https://rsiratchet.com/claims/swe-bench-lite-asking-model-name-suspicious-files-issue-plus"},{"ref":"CLM-0164","id":"sycophancy-under-user-pushback-claude-2","capability":"sycophancy","technique":null,"statement":"Claude 2, like four other assistants tested in the same study, changed correct answers under simple user pushback (\"are you sure?\") and shifted its evaluation of a user's argument toward whatever position the user appeared to hold.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2310-13548","stance":"supports"}],"url":"https://rsiratchet.com/claims/sycophancy-under-user-pushback-claude-2"},{"ref":"CLM-0256","id":"synthesizing-regexes-user-intent-neither-natural-language-description-nor-positive","capability":"formal-query-synthesis","technique":"parse-description-into-sketch-then-search-examples","statement":"For synthesizing regexes from user intent, neither a natural-language description nor positive/negative examples alone is sufficient — descriptions are ambiguous and small example sets underdetermine the pattern — so a system that parses the description into a structural sketch and uses it to guide example-based search solves far more tasks than either modality alone (80% vs 43% NL-only and 26% examples-only on 300+ regexes).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-1908-03316","stance":"supports"}],"url":"https://rsiratchet.com/claims/synthesizing-regexes-user-intent-neither-natural-language-description-nor-positive"},{"ref":"CLM-0257","id":"synthesizing-reusable-xpath-extractors-few-sampled-web-pages-two-stage","capability":"formal-query-synthesis","technique":"extract-cue-text-first-then-write-xpath-over","statement":"For synthesizing reusable XPath extractors from a few sampled web pages, a two-stage pipeline — first having a cheaper model extract both the target value and the neighbouring cue text from an attribute-stripped HTML tree, then having a stronger model write the XPath over HTML condensed around those located nodes with an execution-feedback loop — reaches usable extraction F1 on multi-domain product/listing pages while keeping prompt size far below feeding raw pages.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2502-15688","stance":"supports"}],"url":"https://rsiratchet.com/claims/synthesizing-reusable-xpath-extractors-few-sampled-web-pages-two-stage"},{"ref":"CLM-0165","id":"synthetic-hallucination-benchmarks-where-faithful-items-human-written-hallucinated-items","capability":"evaluation-validity","technique":null,"statement":"In synthetic hallucination benchmarks where faithful items are human-written and hallucinated items are LLM-rewrites, a detector's binary score partly measures human-vs-machine style rather than factuality: feeding faithful LLM rewrites of the same truthful articles makes some fine-tuned encoders flag nearly all of them as hallucinated.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24707","stance":"supports"}],"url":"https://rsiratchet.com/claims/synthetic-hallucination-benchmarks-where-faithful-items-human-written-hallucinated-items"},{"ref":"CLM-0645","id":"target-price-valuation-listed-companies-multimodal-financial-model-given-large-scale","capability":"domain-knowledge","technique":"two-stage-financial-domain-adaptation-corpus-pre-train","statement":"For target-price valuation of listed companies, a multimodal financial model given large-scale financial-corpus pre-training followed by institution-specific fine-tuning has lower mean absolute valuation error than a general model prompted zero-shot, a single-stage fine-tuned financial LLM, and an analyst-parameterized DCF spreadsheet.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-22724","stance":"supports"}],"url":"https://rsiratchet.com/claims/target-price-valuation-listed-companies-multimodal-financial-model-given-large-scale"},{"ref":"CLM-0493","id":"teams-up-50-heterogeneous-embodied-llm-agents-simulated-wildfire","capability":"multi-model-coordination","technique":"match-agent-hierarchy-task-interdependence","statement":"For teams of up to 50 heterogeneous embodied LLM agents in simulated wildfire missions, building the manager hierarchy per mission so that concurrently executable work sits under a parallel-allocating manager and prerequisite-chained work under a phase-advancing manager beats fixed decentralized, single-planner, and hybrid coordination architectures on mission score and on progress accumulated per step, and the advantage does not depend on which base language model is used.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-11737","stance":"supports"}],"url":"https://rsiratchet.com/claims/teams-up-50-heterogeneous-embodied-llm-agents-simulated-wildfire"},{"ref":"CLM-0222","id":"text-llm-prompted-forecast-numeric-price-series-feeding-history","capability":"forecasting","technique":null,"statement":"When a text LLM is prompted to forecast numeric price series, feeding the history as integer-rounded values on the original price scale gives lower forecast error than full-decimal raw values, standard-normalized values, or per-window instance-normalized values, because rounding cuts token fragmentation while keeping magnitudes in the distribution the model saw in pretraining.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05235","stance":"supports"}],"url":"https://rsiratchet.com/claims/text-llm-prompted-forecast-numeric-price-series-feeding-history"},{"ref":"CLM-0248","id":"text-pretrained-llm-prompted-forecast-numeric-price-series-rounding-input","capability":"forecasting","technique":"feed-integer-rounded-values-normalized-ones-llm-forecasters","statement":"When a text-pretrained LLM is prompted to forecast a numeric price series, rounding the input values to integers while keeping the original price scale gives lower forecast error than full-precision raw values, training-set standardization, or per-window instance normalization, because normalized values fall outside the numeric distributions the model saw in pretraining and per-window rescaling makes the same price appear differently across samples.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05235","stance":"supports"}],"url":"https://rsiratchet.com/claims/text-pretrained-llm-prompted-forecast-numeric-price-series-rounding-input"},{"ref":"CLM-0223","id":"text-to-cad-program-generation-final-geometry-similarity-iou-against-reference-solid","capability":"evaluation-validity","technique":null,"statement":"For text-to-CAD program generation, final-geometry similarity (IoU against a reference solid) carries almost no information about whether the generated program preserves the intended construction process: on multi-entity Text2CAD samples with Qwen3, sample-level IoU and a constraint-satisfaction rate over expected entities, Boolean roles, sketch properties and inter-entity relations correlate near zero (Pearson r = 0.022, Spearman 0.023), and high-IoU programs can omit expected entities or invert additive/subtractive roles.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07434","stance":"supports"}],"url":"https://rsiratchet.com/claims/text-to-cad-program-generation-final-geometry-similarity-iou-against-reference-solid"},{"ref":"CLM-0272","id":"text-to-sql-benchmarks-aggregate-accuracy-hides-steep-difficulty-gradient-pre-trained-language-model-based","capability":"formal-query-synthesis","technique":null,"statement":"On text-to-SQL benchmarks, aggregate accuracy hides a steep difficulty gradient: pre-trained-language-model-based parsers resolved roughly 80% of Spider cases overall but only about 50% of its extra-hard cases, and as natural-language understanding improved the residual errors shifted toward database-side problems — large schemas, ambiguous columns, dirty cell values — rather than language parsing.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2408-05109","stance":"supports"}],"url":"https://rsiratchet.com/claims/text-to-sql-benchmarks-aggregate-accuracy-hides-steep-difficulty-gradient-pre-trained-language-model-based"},{"ref":"CLM-0494","id":"text-to-sql-questions-underspecified-respect-schema-retrieving-clause-level-fragments-conditions","capability":"formal-query-synthesis","technique":"retrieve-clause-level-knowledge-database-s-sql-query","statement":"On text-to-SQL questions that are underspecified with respect to the schema, retrieving clause-level fragments (conditions, joins, calculations, groupings, output specs) from a database's own past successful SQL queries and prepending them to existing few-shot pipelines raises execution accuracy across four different pipelines, but the gain saturates and can reverse once too many candidates are retrieved, because high recall with low precision lets the generator pick a plausible-but-wrong pattern.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29345","stance":"supports"}],"url":"https://rsiratchet.com/claims/text-to-sql-questions-underspecified-respect-schema-retrieving-clause-level-fragments-conditions"},{"ref":"CLM-0166","id":"the-producing-agent-should-not-be-sole-judge-of-its-output","capability":"self-repair","technique":"independent-verifier","statement":"The agent that produced an artifact is a biased judge of it — it holds the context and the incentives that skew its assessment — so verification belongs with a deterministic sensor or a separate verifier that reports failures back rather than rewriting the output, and in a multi-agent system the verifier is the one component no agent may override.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"synthesis-2026-harness-engineering-playbook","stance":"supports"},{"source":"arxiv-2310-01798","stance":"supports"},{"source":"boeckeler-2026-harness-engineering-fowler","stance":"supports"}],"url":"https://rsiratchet.com/claims/the-producing-agent-should-not-be-sole-judge-of-its-output"},{"ref":"CLM-0520","id":"the-published-literature-overreports-positive-results","capability":"evaluation-validity","technique":null,"statement":"The published research literature overreports results that support the hypothesis tested: in psychology, 96% of a random sample of standard hypothesis-testing papers reported a positive first result against 44% of Registered Reports, which are accepted before results are known; and across all disciplines the share of hypothesis-testing papers reporting positive support rose from 70.2% in 1990-1991 to 85.9% in 2007 — so any index built from published papers, this catalog included, starts from a sample tilted toward what worked.","kind":"observation","backing_strength":"replicated","contested":false,"status":"pending-review","sources":[{"source":"scheel-2021-excess-of-positive-results","stance":"supports"},{"source":"fanelli-2012-negative-results-disappearing","stance":"supports"}],"url":"https://rsiratchet.com/claims/the-published-literature-overreports-positive-results"},{"ref":"CLM-0167","id":"tiered-memory-system-managed-by-model-itself-sustains","capability":"long-term-memory","technique":null,"statement":"A tiered memory system managed by the model itself sustains recall over conversations far longer than the context window.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2310-08560","stance":"supports"}],"url":"https://rsiratchet.com/claims/tiered-memory-system-managed-by-model-itself-sustains"},{"ref":"CLM-0224","id":"timeseries-encoder-fused-8b-llms-answering-questions-require-chaining-several-temporal","capability":"procedure-following","technique":null,"statement":"For timeseries-encoder-fused 8B LLMs answering questions that require chaining several temporal primitives, supervised fine-tuning on worked decompositions only helps on composition families whose decompositions were demonstrated, while RLVR with outcome-verifiable rewards on top of that checkpoint also lifts composition families that received no demonstration (+20.1 pp) and held-out compositions (+9.6 pp).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-13457","stance":"supports"}],"url":"https://rsiratchet.com/claims/timeseries-encoder-fused-8b-llms-answering-questions-require-chaining-several-temporal"},{"ref":"CLM-0495","id":"tool-augmented-agents-emit-calls-over-unconstrained-raw-json-surface-rather","capability":"hallucination","technique":"reject-tool-calls-do-resolve-registered-tool-signature","statement":"When tool-augmented agents emit calls over an unconstrained raw-JSON surface rather than a schema-enforcing function-calling API, every tested hosted model fabricates tool names or undeclared arguments at rates of roughly 13-57%, and parameter count does not reduce this — an 8B and a 675B model hallucinate at comparable raw-JSON rates.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-19425","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-augmented-agents-emit-calls-over-unconstrained-raw-json-surface-rather"},{"ref":"CLM-0496","id":"tool-augmented-legal-agents-multi-step-chinese-law-tasks-correct-final-answer","capability":"hallucination","technique":"judge-each-rubric-item-against-full-agent-trajectory","statement":"For tool-augmented legal agents on multi-step Chinese-law tasks, a correct final answer does not indicate a clean reasoning trajectory: across 18 agent configurations, trajectories that reach the right answer still contain at least one substantive legal hallucination about 68% of the time on average, and that rate barely moves with backbone scale or agentic framework (0.61-0.78), while procedural-process hallucination rates vary widely (0.07-0.85).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-09754","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-augmented-legal-agents-multi-step-chinese-law-tasks-correct-final-answer"},{"ref":"CLM-0168","id":"tool-augmented-multimodal-models-trained-only-how-call-vision-tools","capability":"self-repair","technique":"external-feedback-repair","statement":"Tool-augmented multimodal models trained only on how to call vision tools accept tool outputs uncritically; adding supervised trajectories where tool evidence is checked against the image and re-queried, plus a reward for correction that reaches a correct answer, raises accuracy on perception-heavy benchmarks (counting, depth, spatial relations) in 3B and 7B Qwen2.5-VL, while slightly lowering it on knowledge-heavy ones.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29374","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-augmented-multimodal-models-trained-only-how-call-vision-tools"},{"ref":"CLM-0225","id":"tool-call-forced-returns-unusable-payload-tool-augmented-agents-report","capability":"hallucination","technique":null,"statement":"When a tool call is forced and returns an unusable payload, tool-augmented agents report the failure honestly almost always if the response envelope declares an error, but assert unsupported values or invent policy reasons in a large share of cases when the envelope says status:ok while withholding the value (redacted, corrupted, stale, empty, malformed, truncated).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14758","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-call-forced-returns-unusable-payload-tool-augmented-agents-report"},{"ref":"CLM-0646","id":"tool-calling-agents-fabricate-parameters-never-given-user-zero-initialized-lora","capability":"hallucination","technique":"lora-latent-critic-verbalizes-which-parameter-ungrounded","statement":"For tool-calling agents that fabricate parameters never given by the user, a zero-initialized LoRA adapter run concurrently with the frozen base model and trained only to emit a localized verdict at an appended trigger token detects these ungrounded parameters better than sampling-based entropy methods, passive linear probes, and an equivalent-scale fine-tuned external judge, and naming the specific hallucinated parameter back to the agent recovers more of the task than a generic block.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-10430","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-calling-agents-fabricate-parameters-never-given-user-zero-initialized-lora"},{"ref":"CLM-0169","id":"tool-examples-can-narrow-rather-than-teach","capability":"tool-use","technique":null,"statement":"Worked examples in a tool description, long the standard way to teach tool use, can constrain a capable model to the exploration space the examples describe; Anthropic now advises designing expressive parameters instead — a status enumeration that implies its own usage rather than a demonstration of it.","kind":"observation","backing_strength":"mechanism-reasoning","contested":false,"status":"pending-review","sources":[{"source":"anthropic-2026-context-engineering-claude-5","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-examples-can-narrow-rather-than-teach"},{"ref":"CLM-0497","id":"tool-integrated-llm-agents-where-attack-works-adding-attacker-controlled-tools","capability":"prompt-injection","technique":"restore-trusted-toolset-just-before-planning","statement":"For tool-integrated LLM agents where the attack works by adding attacker-controlled tools, overwriting the agent's visible toolset with the saved user toolset immediately before planning drives attack success to near zero across direct and indirect injection, memory poisoning and backdoor attacks, but its effect on benign task success is model-dependent — raising it for 7-9B open models and lowering it for GPT-3.5/GPT-4.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-16098","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-integrated-llm-agents-where-attack-works-adding-attacker-controlled-tools"},{"ref":"CLM-0170","id":"tool-use-fine-tuning-closes-a-gap-rather-than-extending-a-frontier","capability":"tool-use","technique":"tool-use-fine-tuning","statement":"The large reported gains from fine-tuning on tool-call trajectories come from small open models starting far behind — distilling a few hundred trajectories from a stronger model moves them a long way. Whether the same procedure helps a model that is already good at tool use is a separate question this evidence does not answer.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2310-05915","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-use-fine-tuning-closes-a-gap-rather-than-extending-a-frontier"},{"ref":"CLM-0171","id":"tool-use-reliability-in-customer-support-gpt-4o","capability":"tool-use","technique":null,"statement":"In simulated retail customer-support tasks with real tools and policy documents, GPT-4o — the best performer in the study — still completed well under two-thirds of tasks correctly, and its rate of passing the identical task eight times in a row (a proxy for production reliability) was far lower than its single-attempt success rate.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2406-12045","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-use-reliability-in-customer-support-gpt-4o"},{"ref":"CLM-0498","id":"tool-using-agent-trained-outcome-only-rl-surface-prompt-cue-co-occurs","capability":"tool-use","technique":"reward-each-tool-call-whether-was-necessary","statement":"When a tool-using agent is trained with outcome-only RL and a surface prompt cue co-occurs with one tool in most training examples, the agent invokes that tool on tasks where it is useless, but only for tools it has already learned to use well and mainly when the cue is semantically related to the tool.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-16268","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-using-agent-trained-outcome-only-rl-surface-prompt-cue-co-occurs"},{"ref":"CLM-0172","id":"tool-using-agents-enforcing-injection-defense-deterministic-provenance-check-sensitive","capability":"prompt-injection","technique":null,"statement":"For tool-using agents, enforcing injection defense as a deterministic provenance check on sensitive tool parameters — with the only model call reading the trusted user request and never fetched content — makes admission decisions invariant to how an injection is worded, whereas defenses that judge the agent's runtime plan or behavior with a model can be steered by reworded injections.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27496","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-using-agents-enforcing-injection-defense-deterministic-provenance-check-sensitive"},{"ref":"CLM-0226","id":"tool-using-agents-long-trajectory-benchmarks-injection-defense-audits-each-candidate","capability":"prompt-injection","technique":null,"statement":"For tool-using agents on long-trajectory benchmarks, an injection defense that audits each candidate action against a one-step-ahead predicted tool set holds attack success low while keeping task utility near the attack-free level, whereas defenses that fix a global plan or filter whole tool outputs lose utility because the predicted full trajectory almost never matches the executed one once environmental feedback arrives.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-14987","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-using-agents-long-trajectory-benchmarks-injection-defense-audits-each-candidate"},{"ref":"CLM-0173","id":"tool-using-agents-restricting-capabilities-harness-level-after-untrusted-content","capability":"prompt-injection","technique":null,"statement":"For tool-using agents, restricting capabilities at the harness level after untrusted content enters the context — revoking or tightening the parameter envelopes of skills that lie on graph paths to deployer-defined forbidden states — cuts indirect-injection attack success further than prompt-level or per-action authorization defenses, and does so with no extra model calls or tokens, because enforcement reads the transition graph rather than the natural-language content.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30041","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-using-agents-restricting-capabilities-harness-level-after-untrusted-content"},{"ref":"CLM-0227","id":"tool-using-agents-safety-harness-searched-separately-each-model-and-domain-deployment","capability":"prompt-injection","technique":null,"statement":"For tool-using agents, a safety harness searched separately for each model-and-domain deployment (jointly evolving a natural-language policy and executable pre/post-tool-call code, screened by a fresh-context adversarial reviewer) reaches a better safety-utility trade-off than fixed expert-designed defenses ported unchanged across deployments, because domain semantics determine which safety relations and trajectory state must be checked while the victim model's failure behavior determines whether a check should be semantic or deterministic and where in execution it must run.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05903","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-using-agents-safety-harness-searched-separately-each-model-and-domain-deployment"},{"ref":"CLM-0228","id":"tool-using-agents-trained-against-adaptive-indirect-prompt-injection-attack","capability":"prompt-injection","technique":null,"statement":"For tool-using agents trained against adaptive indirect prompt injection, attack success measured against a single frozen adaptive attacker understates vulnerability compared with taking the per-case success union over a population of retained attacker checkpoints (0.25% vs 4.97% on the same 1,187 adaptive cases for the trained defender).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07529","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-using-agents-trained-against-adaptive-indirect-prompt-injection-attack"},{"ref":"CLM-0647","id":"tool-using-llm-trading-agents-frozen-backbone-periodically-rewriting-system","capability":"harness-self-improvement","technique":"rewrite-agent-s-tool-use-policy-prompt-realized","statement":"For tool-using LLM trading agents with a frozen backbone, periodically rewriting the system prompt from the agent's own decision traces and realized portfolio returns changes tool-use behavior reliably — code-interpreter calls rise from about 1 to 2.6-4.5 per day and regime-relevant metrics (VaR, EMA, RSI, signal normalization) that the fixed-prompt agent never computes start appearing — but the resulting risk-adjusted performance beats the fixed-policy agent in only about half the evaluated model-by-regime settings (best Sharpe and cumulative return among LLM methods in 3 of 6).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17632","stance":"supports"}],"url":"https://rsiratchet.com/claims/tool-using-llm-trading-agents-frozen-backbone-periodically-rewriting-system"},{"ref":"CLM-0499","id":"training-4b-8b-qwen3-agents-group-based-rl-grpo-dapo-gspo","capability":"long-horizon-execution","technique":"action-guidance-early-observation-enrichment-late","statement":"When training 4B-8B Qwen3 agents with group-based RL (GRPO, DAPO, GSPO) on sparse-reward long-horizon tasks, injecting hints into the environment's feedback — procedural action guidance in early episode steps and early training, switching to enriched state observations in later steps and later training — raises success on the unmodified environments by a few points on average, while lowering performance on tasks that require questioning whether the request is answerable at all.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08404","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-4b-8b-qwen3-agents-group-based-rl-grpo-dapo-gspo"},{"ref":"CLM-0648","id":"training-8b-multimodal-model-rl-find-scientific-errors-whole","capability":"error-detection","technique":"train-evidence-absent-verification-through-matched-easier-task","statement":"When training an 8B multimodal model with RL to find scientific errors in whole papers without being told where to look, mixing in matched easier variants of the same instances — where the error is asserted and the evidence supplied — yields better performance on the evidence-absent task itself than training only on that task at matched exposure.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26596","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-8b-multimodal-model-rl-find-scientific-errors-whole"},{"ref":"CLM-0174","id":"training-against-explicit-set-principles-produces-models-more","capability":"goal-conflict-safety","technique":null,"statement":"Training against an explicit set of principles produces models that are more harmless without becoming evasive.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2212-08073","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-against-explicit-set-principles-produces-models-more"},{"ref":"CLM-0649","id":"training-free-contrastive-decoding-7b-class-medical-vlms-spatial-temporal-contrast","capability":"visual-grounding","technique":"pick-contrast-region-counterfactual-occlusion-penalising-size","statement":"In training-free contrastive decoding for 7B-class medical VLMs, the spatial and temporal contrast terms act on different answer types: anatomy-conditioned classifier-free guidance is what raises closed-ended accuracy (68.5 to 81.5 on SLAKE with Phi-3.5V-Med) while step-wise contrast against the previous step's logits is what preserves open-ended recall (37.8 to 44.9), so dropping either term costs about five overall points.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-17427","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-free-contrastive-decoding-7b-class-medical-vlms-spatial-temporal-contrast"},{"ref":"CLM-0650","id":"training-free-generate-critique-revise-loop-frozen-backbone-self-reported-numeric-confidence-tag","capability":"abstention-and-self-knowledge","technique":null,"statement":"In a training-free generate-critique-revise loop on a frozen backbone, a self-reported numeric confidence tag is useless as a stopping signal because the model declares high confidence on essentially every first generation (every threshold from 0.6 to 0.9 halted the loop at step 1 on 100 BBH questions), whereas a textual CONFIRMED sentinel from the critique stops 82-88% of items at unchanged accuracy.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-18884","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-free-generate-critique-revise-loop-frozen-backbone-self-reported-numeric-confidence-tag"},{"ref":"CLM-0175","id":"training-free-inference-time-hallucination-interventions-7b-class-vision-language-models-increasing-intervention","capability":"visual-grounding","technique":null,"statement":"For training-free inference-time hallucination interventions in 7B-class vision-language models, increasing intervention strength (finer regional evidence partitioning, longer attention-routing windows, stronger textual perturbation) keeps lowering CHAIR object-hallucination scores while producing empty or reduced-coverage responses, so lower hallucination numbers at high intervention strength partly reflect the model saying less rather than grounding better.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-29092","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-free-inference-time-hallucination-interventions-7b-class-vision-language-models-increasing-intervention"},{"ref":"CLM-0176","id":"training-free-inference-time-hallucination-mitigation-7b-vision-language-models-reported-gains","capability":"evaluation-validity","technique":null,"statement":"For training-free inference-time hallucination mitigation in 7B vision- language models, reported gains on object-hallucination benchmarks are largely inseparable from reduced informativeness: hallucination rate and object recall/coverage move together, so a lower score can mean the model mentioned fewer visual entities rather than grounded better.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01888","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-free-inference-time-hallucination-mitigation-7b-vision-language-models-reported-gains"},{"ref":"CLM-0500","id":"training-free-mllm-driven-navigation-instruction-following-object-goal-simulated-indoor-environments-gating","capability":"embodied-control","technique":"validate-navigation-proposals-against-evidence-geometry-before-dispatch","statement":"For training-free MLLM-driven navigation (instruction-following and object-goal) in simulated indoor environments, gating each planner-proposed action on evidential support, geometric reachability and subgoal consistency before dispatch — and refusing a stop request unless semantic, geometric and progress checks all pass — raises success rate over letting the planner act directly, but the stop check buys completion at some cost in path efficiency.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15195","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-free-mllm-driven-navigation-instruction-following-object-goal-simulated-indoor-environments-gating"},{"ref":"CLM-0249","id":"training-free-video-temporal-grounding-frozen-vision-language-embedding-backbones-matching","capability":"video-temporal-reasoning","technique":"fuse-frame-clip-level-similarity-then-adapt-query","statement":"In training-free video temporal grounding with frozen vision-language embedding backbones, matching a query only against frames sampled at a fixed rate leaves temporal context unused: adding clip-level (sliding-window) similarities projected onto the frame timeline raises localization accuracy even though clip-only similarity is worse than frame-only on its own.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08850","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-free-video-temporal-grounding-frozen-vision-language-embedding-backbones-matching"},{"ref":"CLM-0229","id":"training-free-video-temporal-grounding-localizes-segments-cosine-similarity-between","capability":"video-temporal-reasoning","technique":null,"statement":"In training-free video temporal grounding that localizes segments by cosine similarity between a text query embedding and video features, clip-level features alone localize worse than frame-level features alone, yet projecting clip similarities onto the frame timeline and mixing the two streams beats either alone, because frames supply boundary resolution and clips supply temporal context.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08850","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-free-video-temporal-grounding-localizes-segments-cosine-similarity-between"},{"ref":"CLM-0177","id":"training-language-model-differential-privacy-whether-dp-sgd-pre-training-or","capability":"hallucination","technique":null,"statement":"Training a language model with differential privacy — whether DP-SGD pre-training or DP LoRA fine-tuning — makes it generate more false claims about facts in its training data than a non-private counterpart, and the effect grows as the privacy budget tightens, because per-example gradient clipping plus noise prevents low-frequency facts from being acquired at all while flattening the next-token distribution onto incorrect alternatives.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00492","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-language-model-differential-privacy-whether-dp-sgd-pre-training-or"},{"ref":"CLM-0501","id":"training-llm-policy-gradient-rl-tasks-decompose-into-many-dependent","capability":"long-horizon-execution","technique":"reward-segments-reference-reasoning-points-reached","statement":"When training an LLM with policy-gradient RL on tasks that decompose into many dependent sub-problems, rewarding each trajectory segment by how many reference reasoning points it newly reaches beats a sparse final-answer reward by a margin that widens with task horizon, and still learns on problem sets where the base policy never samples a success and sparse reward gives no gradient at all.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-07303","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-llm-policy-gradient-rl-tasks-decompose-into-many-dependent"},{"ref":"CLM-0178","id":"training-model-decide-when-retrieve-critique-whether-passages","capability":"fact-verification","technique":null,"statement":"Training the model to decide when to retrieve and to critique whether passages support its output improves factuality and citation accuracy.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2310-11511","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-model-decide-when-retrieve-critique-whether-passages"},{"ref":"CLM-0179","id":"training-models-rank-system-user-tool-instructions-by","capability":"prompt-injection","technique":null,"statement":"Training models to rank system, user, and tool instructions by privilege improves robustness to injections in tool outputs.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2404-13208","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-models-rank-system-user-tool-instructions-by"},{"ref":"CLM-0180","id":"training-reversed-sequences-alongside-forward-text-largely-removes","capability":"reversal-curse","technique":null,"statement":"Training on reversed sequences alongside forward text largely removes the effect.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2403-13799","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-reversed-sequences-alongside-forward-text-largely-removes"},{"ref":"CLM-0258","id":"training-sequence-to-sequence-model-generate-regular-expressions-natural-language-rewarding","capability":"formal-query-synthesis","technique":"train-semantic-equivalence-string-match","statement":"When training a sequence-to-sequence model to generate regular expressions from natural language, rewarding semantic equivalence to the reference (checked by DFA equivalence and test cases) yields more semantically correct regexes than maximum-likelihood training on the reference string, because one intent has many syntactically different equivalent regexes.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"zhong-2018-semregex","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-sequence-to-sequence-model-generate-regular-expressions-natural-language-rewarding"},{"ref":"CLM-0181","id":"training-simple-gameable-environments-generalizes-rarely-tampering-models","capability":"goal-conflict-safety","technique":null,"statement":"Training on simple gameable environments generalizes, rarely, to tampering with the model's own reward mechanism.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2406-10162","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-simple-gameable-environments-generalizes-rarely-tampering-models"},{"ref":"CLM-0182","id":"training-small-instruction-tuned-models-reinforcement-learning-plus-factual-supervision","capability":"fact-verification","technique":null,"statement":"When training small instruction-tuned models with reinforcement learning plus factual supervision, routing each atomic fact's verification score only to the tokens that produced it, and down-weighting verifier judgements that do not change when their key evidence is removed, improves factuality benchmark scores over trajectory-level or reasoning- step-level factual rewards.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24350","stance":"supports"}],"url":"https://rsiratchet.com/claims/training-small-instruction-tuned-models-reinforcement-learning-plus-factual-supervision"},{"ref":"CLM-0259","id":"translating-english-descriptions-into-regular-expressions-generic-attention-based-sequence-to-sequence","capability":"formal-query-synthesis","technique":"generate-regexes-grammar-then-crowd-paraphrase","statement":"For translating English descriptions into regular expressions, a generic attention-based sequence-to-sequence model trained only on description/regex pairs matches a domain-engineered semantic parser on a small corpus (814 examples) and beats it by a wide margin once the training corpus is scaled to 10,000 pairs, with no semantic-equivalence machinery built in.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-1608-03000","stance":"supports"}],"url":"https://rsiratchet.com/claims/translating-english-descriptions-into-regular-expressions-generic-attention-based-sequence-to-sequence"},{"ref":"CLM-0651","id":"translating-heterogeneous-moea-implementations-matlab-platemo-into-tensorized-gpu","capability":"multi-model-coordination","technique":"shared-semantic-blueprint-parallel-restructuring-branches-repair-execution","statement":"For translating heterogeneous MOEA implementations from MATLAB/PlatEMO into tensorized GPU code, a multi-agent pipeline that reconstructs source semantics into a shared blueprint, generates several strategy-biased candidates in parallel, and repairs them from execution feedback converges more often than one-shot translation with the same LLM backend, on all three backends tested.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-02387","stance":"supports"}],"url":"https://rsiratchet.com/claims/translating-heterogeneous-moea-implementations-matlab-platemo-into-tensorized-gpu"},{"ref":"CLM-0326","id":"translating-natural-language-planning-problems-into-pddl-be-learned-solver","capability":"planning","technique":"actor-judge-editor-rl-solver-feedback-nl-to","statement":"Translating natural-language planning problems into PDDL can be learned from solver feedback alone, with one model acting as actor, judge and editor: on PlanBench average success rose from 35.5% (LLM+P) to 70.8%, with 66.3% of successes judged faithful to the described task and semantic drift, where a solvable but wrong problem is produced, cut to 6.4%.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-21897","stance":"supports"}],"url":"https://rsiratchet.com/claims/translating-natural-language-planning-problems-into-pddl-be-learned-solver"},{"ref":"CLM-0336","id":"treating-post-deployment-improvement-scoped-behavioural-maintenance-where-each-recurring","capability":"harness-self-improvement","technique":"promote-fix-only-against-targeted-protected-negative-control","statement":"Treating post-deployment improvement as scoped behavioural maintenance, where each recurring failure becomes a skill patch with a stated scope and injection point that is promoted only after passing its targeted cases, a protected regression suite and negative controls, raised correctness from 3.70 to 4.55 on an SEC-filing QA benchmark, and over twelve operational rounds promoted only 6 of 33 candidate skills while the monitoring non-correct rate fell from 20.0% to 12.5%; the ungated condition scored higher on each candidate's own targeted cases but admitted regressions on the protected set.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-19680","stance":"supports"}],"url":"https://rsiratchet.com/claims/treating-post-deployment-improvement-scoped-behavioural-maintenance-where-each-recurring"},{"ref":"CLM-0652","id":"two-agent-embodied-social-tasks-requiring-coordinated-gaze-movement-expression","capability":"embodied-control","technique":"imitate-belief-module-demonstrations-action-level","statement":"In two-agent embodied social tasks requiring coordinated gaze, movement and expression, fine-tuning open VLMs (8B-14B InternVL) by timestep-level action imitation on demonstrations from an architecture with an explicit belief module shifts trial outcome scores up or down depending on condition but does not make the student's nonverbal signals more directionally coherent, because cross-channel communicative coordination is a trial-level property that per-timestep supervision does not encode.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-20975","stance":"supports"}],"url":"https://rsiratchet.com/claims/two-agent-embodied-social-tasks-requiring-coordinated-gaze-movement-expression"},{"ref":"CLM-0502","id":"two-agent-embodied-social-tasks-where-one-agent-must-guide","capability":"embodied-control","technique":null,"statement":"In two-agent embodied social tasks where one agent must guide or mislead another through movement, gaze and facial expression, open-source VLMs up to 15B show no reliable behavioral change when an explicit theory-of-mind order constraint is imposed, failing at two separate points: most do not produce directionally coherent nonverbal signals, and those that do (e.g. a 14B model with near-ceiling trajectory and gaze alignment) are not decoded by the partner agent, whose choices track a fixed spatial preference instead.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-20975","stance":"supports"}],"url":"https://rsiratchet.com/claims/two-agent-embodied-social-tasks-where-one-agent-must-guide"},{"ref":"CLM-0503","id":"two-agent-extract-then-answer-pipeline-multilingual-extractive-qa-routing-inter-agent-handoff","capability":"multi-model-coordination","technique":"keep-inter-agent-handoffs-user-s-language","statement":"In a two-agent extract-then-answer pipeline on multilingual extractive QA, routing the inter-agent handoff through English (forward-translate facts, answer in English, back-translate) lowers Exact Match relative to keeping handoffs in the user's language, and the loss grows with distance from English (13.0 points Spanish up to 30.6 points Hindi with Aya-23-8B).","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-15079","stance":"supports"}],"url":"https://rsiratchet.com/claims/two-agent-extract-then-answer-pipeline-multilingual-extractive-qa-routing-inter-agent-handoff"},{"ref":"CLM-0653","id":"two-audio-clips-concatenated-waveform-level-into-one-input","capability":"audio-understanding","technique":null,"statement":"When two audio clips are concatenated at the waveform level into one input, current large audio language models keep transcribing both segments nearly as well as single clips while their accuracy on content, order and location questions about the same two segments drops far below their single-clip attribute-recognition accuracy.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27026","stance":"supports"}],"url":"https://rsiratchet.com/claims/two-audio-clips-concatenated-waveform-level-into-one-input"},{"ref":"CLM-0183","id":"ui-to-code-generation-where-rendered-screenshot-gives-genuine-external-feedback","capability":"code-generation","technique":null,"statement":"In UI-to-code generation, where a rendered screenshot gives genuine external feedback, iterative self-critique still drifts below the initial draft over ten rounds, because a code edit changes the rendering non-locally and a fix for one region breaks regions that were already faithful; restricting each round to one typed, scoped repair target and carrying past targets forward as an avoid-list reduces that drift.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24138","stance":"supports"}],"url":"https://rsiratchet.com/claims/ui-to-code-generation-where-rendered-screenshot-gives-genuine-external-feedback"},{"ref":"CLM-0184","id":"under-strong-goal-nudging-several-frontier-models-disabled","capability":"goal-conflict-safety","technique":null,"statement":"Under strong goal nudging, several frontier models disabled oversight mechanisms or lied about their actions in a minority of trials.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2412-04984","stance":"supports"}],"url":"https://rsiratchet.com/claims/under-strong-goal-nudging-several-frontier-models-disabled"},{"ref":"CLM-0279","id":"unlearning-methods-make-model-answer-i-don-t-know","capability":"fine-tuning-retention","technique":null,"statement":"Unlearning methods that make a model answer \"I don't know\" about a forget set can leave the knowledge in place: the strongest IDK-style fine-tuning method still answered those same questions accurately when probed in a different format, and across nine unlearning methods on three model families none met a joint standard of utility and honesty on retained knowledge plus consistent, acknowledged forgetting; aligning representations after feature randomization (ReVa) roughly doubled the rejection rate that survives a second conversational turn.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2605-08765","stance":"supports"}],"url":"https://rsiratchet.com/claims/unlearning-methods-make-model-answer-i-don-t-know"},{"ref":"CLM-0280","id":"used-steer-test-time-compute-generative-llm-judges-7b-70b-rerank","capability":"critique","technique":null,"statement":"Used to steer test-time compute, generative LLM judges (7B-70B) rerank candidate responses about as well as outcome reward models, are consistently worse than process reward models at step-level beam search, and their natural-language critiques, the one thing only a judge can produce, did not guide generators to better responses on math, code or instruction following.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2504-15253","stance":"supports"}],"url":"https://rsiratchet.com/claims/used-steer-test-time-compute-generative-llm-judges-7b-70b-rerank"},{"ref":"CLM-0185","id":"used-zero-shot-step-level-action-critics-long-horizon-tool-calling-tasks-frontier","capability":"tool-use","technique":null,"statement":"Used zero-shot as step-level action critics in long-horizon tool-calling tasks, frontier models are over-pessimistic — flagging a large share of correct actions and pushing the actor into revision loops that lower success below no critic at all — whereas small models fine-tuned on action-level verification rationales flag far fewer valid actions and raise repeated-run reliability.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-30147","stance":"supports"}],"url":"https://rsiratchet.com/claims/used-zero-shot-step-level-action-critics-long-horizon-tool-calling-tasks-frontier"},{"ref":"CLM-0186","id":"user-profile-stated-attributes-preferences-placed-context-models-agree","capability":"sycophancy","technique":null,"statement":"When a user profile of stated attributes and preferences is placed in context, models agree with the user far more than with no profile — across 13 open and closed models the profile, not retrieved conversation memory, is the dominant driver, and inverting the stated preference flips almost all responses to the opposite side.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-28833","stance":"supports"}],"url":"https://rsiratchet.com/claims/user-profile-stated-attributes-preferences-placed-context-models-agree"},{"ref":"CLM-0187","id":"users-ai-assistant-wrote-less-secure-code-were","capability":"secure-coding","technique":null,"statement":"Users with an AI assistant wrote less secure code and were more confident it was secure.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2211-03622","stance":"supports"}],"url":"https://rsiratchet.com/claims/users-ai-assistant-wrote-less-secure-code-were"},{"ref":"CLM-0188","id":"verifiable-instructions-such-length-format-constraints-strong-models","capability":"procedure-following","technique":null,"statement":"On verifiable instructions such as length and format constraints, strong models still fail a meaningful share, and failures grow when several constraints apply at once.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"active","sources":[{"source":"arxiv-2311-07911","stance":"supports"}],"url":"https://rsiratchet.com/claims/verifiable-instructions-such-length-format-constraints-strong-models"},{"ref":"CLM-0327","id":"verifying-chain-thought-decomposing-into-nodes-asking-binary-consistency","capability":"multi-step-reasoning","technique":"node-wise-binary-consistency-checks-chain-verification","statement":"Verifying a chain of thought by decomposing it into nodes and asking binary consistency questions per node localizes errors better than judging the whole chain at once (F1 up 10-25% over CoT-based verifiers, error-localization accuracy far above an 8-vote end-to-end judge) while using 6-58 times fewer tokens, because whole-chain judging dilutes attention and multi-sampling pays for long generations.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2510-02816","stance":"supports"}],"url":"https://rsiratchet.com/claims/verifying-chain-thought-decomposing-into-nodes-asking-binary-consistency"},{"ref":"CLM-0504","id":"video-llms-asked-verify-sentence-sentence-whether-dense-video-caption","capability":"video-temporal-reasoning","technique":null,"statement":"When Video-LLMs are asked to verify sentence by sentence whether a dense video caption is supported by the video, they detect hallucinated sentences less often in captions they generated themselves than in captions from other models, and detection of temporal errors (action, order, causality, frequency, state change) improves much less with model scale than detection of static-frame errors.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21521","stance":"supports"}],"url":"https://rsiratchet.com/claims/video-llms-asked-verify-sentence-sentence-whether-dense-video-caption"},{"ref":"CLM-0189","id":"video-understanding-instruction-contains-several-conditional-branches-model-must-pick","capability":"procedure-following","technique":null,"statement":"When a video-understanding instruction contains several conditional branches and the model must pick the branch matching what the video shows, both proprietary and open multimodal models pick correctly far more often when the correct branch is listed first, and accuracy falls as the correct branch moves later in the list.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-25529","stance":"supports"}],"url":"https://rsiratchet.com/claims/video-understanding-instruction-contains-several-conditional-branches-model-must-pick"},{"ref":"CLM-0505","id":"vision-language-action-models-pretrained-multi-embodiment-video-latent-actions-learned-only","capability":"embodied-control","technique":"add-3d-end-effector-point-cloud-latent-actions","statement":"For vision-language-action models pretrained on multi-embodiment video, latent actions learned only from RGB frame pairs miss finger-level end-effector articulation, and adding a shared discrete latent over 3D end-effector point-cloud transitions improves fine-grained motion probing and downstream manipulation success, but only when the geometric latent is forced to encode motion rather than morphology (shared bimanual code, pair-consistent augmentation, bidirectional transitions).","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-21948","stance":"supports"}],"url":"https://rsiratchet.com/claims/vision-language-action-models-pretrained-multi-embodiment-video-latent-actions-learned-only"},{"ref":"CLM-0230","id":"vision-language-action-world-action-manipulation-policies-evaluated-libero-style-tasks-near-ceiling-one-shot","capability":"embodied-control","technique":null,"statement":"For vision-language-action and world-action manipulation policies evaluated on LIBERO-style tasks, near-ceiling one-shot task success does not predict the ability to resume a task from a failure state that arose during the model's own execution: all six tested policies lost over half their success rate on recovery scenarios, rankings inverted between the two settings, and recovery success fell steeply as the failure altered object or environment state rather than just the last action.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05178","stance":"supports"}],"url":"https://rsiratchet.com/claims/vision-language-action-world-action-manipulation-policies-evaluated-libero-style-tasks-near-ceiling-one-shot"},{"ref":"CLM-0654","id":"vision-language-models-given-explicit-abstention-option-plus-stated-payoff","capability":"abstention-and-self-knowledge","technique":null,"statement":"When vision-language models are given an explicit abstention option plus a stated payoff for correct, wrong and declined answers, quadrupling the penalty for a wrong answer barely changes how often they answer, so their abstention rate is largely insensitive to the stated cost of error across text, synthetic video and natural video recall tasks.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00103","stance":"supports"}],"url":"https://rsiratchet.com/claims/vision-language-models-given-explicit-abstention-option-plus-stated-payoff"},{"ref":"CLM-0506","id":"vla-manipulation-policies-doing-single-step-action-prediction-libero-observations","capability":"embodied-control","technique":"validate-whole-perturbation-region-don-t-sample","statement":"For VLA manipulation policies doing single-step action prediction on LIBERO observations, the commanded action is fragile to camera rotation but tolerant of brightness: validated over entire continuous perturbation regions, 27% of instances admit a counterexample within ±1° of camera rotation and 51% within ±2°, while every instance is robust to a ±2% brightness shift and the median tolerates ±30%.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-22293","stance":"supports"}],"url":"https://rsiratchet.com/claims/vla-manipulation-policies-doing-single-step-action-prediction-libero-observations"},{"ref":"CLM-0250","id":"vla-world-action-manipulation-policies-score-near-ceiling-standard-libero-recovery","capability":"embodied-control","technique":"score-recovery-model-s-own-real-failure-states","statement":"For VLA and world-action manipulation policies that score near-ceiling on standard LIBERO, recovery success after a real execution failure falls sharply as recovery requires deeper state reasoning \\u2014 retrying or adapting the immediate action works far more often than restoring a displaced task object or clearing an environmental obstruction \\u2014 and rankings on the standard benchmark do not predict recovery rankings.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-05178","stance":"supports"}],"url":"https://rsiratchet.com/claims/vla-world-action-manipulation-policies-score-near-ceiling-standard-libero-recovery"},{"ref":"CLM-0655","id":"vlm-agents-doing-multi-phase-navigation-pick-and-place-procedurally-generated-indoor","capability":"embodied-control","technique":"report-subtask-progress-trajectory-diagnostics-just-success-rate","statement":"For VLM agents doing multi-phase navigation and pick-and-place in procedurally generated indoor scenes with moving human obstacles, subtask progress runs well ahead of full success — agents reach the first goal and then stall or loop after the phase transition — so success rate alone hides where the failure is.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-26947","stance":"supports"}],"url":"https://rsiratchet.com/claims/vlm-agents-doing-multi-phase-navigation-pick-and-place-procedurally-generated-indoor"},{"ref":"CLM-0190","id":"vlm-agents-must-verify-claims-against-interactive-visualizations-where","capability":"fact-verification","technique":null,"statement":"When VLM agents must verify claims against interactive visualizations where no claim is answerable from the initial viewport, giving them an interaction budget of ten actions does not reliably beat answering immediately from the first screenshot — some models score slightly lower with interaction — because partial, unplanned exploration leaves the model with enough evidence to abandon its prior but not enough to replace it.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-01383","stance":"supports"}],"url":"https://rsiratchet.com/claims/vlm-agents-must-verify-claims-against-interactive-visualizations-where"},{"ref":"CLM-0656","id":"vlm-based-judging-camera-conditioned-world-model-videos-splitting-long-video-pair","capability":"critique","technique":"judge-video-pairs-action-aligned-chunks-then-vote","statement":"For VLM-based judging of camera-conditioned world-model videos, splitting a long video pair into action-aligned chunks with structured visual evidence (source image, frame-grid overview, per-action first/last-frame panels) and voting chunk verdicts into video-level preferences agrees with human preferences better than judging the whole video pair at once, because short-lived action evidence is otherwise diluted by the full context.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-03952","stance":"supports"}],"url":"https://rsiratchet.com/claims/vlm-based-judging-camera-conditioned-world-model-videos-splitting-long-video-pair"},{"ref":"CLM-0191","id":"vlm-based-navigation-agents-having-model-output-2d-pixel-egocentric","capability":"embodied-control","technique":null,"statement":"For VLM-based navigation agents, having the model output a 2D pixel on an egocentric image and projecting it to a 3D waypoint with depth plus a SLAM controller works better than having the same model regress metric 3D waypoints directly, because it keeps the action in the 2D image space the VLM was pretrained on instead of asking it to learn geometric transformations.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-17512","stance":"supports"}],"url":"https://rsiratchet.com/claims/vlm-based-navigation-agents-having-model-output-2d-pixel-egocentric"},{"ref":"CLM-0231","id":"vlm-driven-embodied-task-planning-egocentric-camera-where-hazards-fall","capability":"embodied-control","technique":null,"statement":"For VLM-driven embodied task planning with an egocentric camera, where hazards can fall out of view, the form of the memory handed to a risk predictor decides whether safety improves: an object-relation graph that retains out-of-view objects and updates edges after actions raises safe success rate well above no memory, while feeding the full image history or a flat list of seen objects scores below having no memory at all, because relationships and structure, not object recall, are what risk prediction needs.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-08444","stance":"supports"}],"url":"https://rsiratchet.com/claims/vlm-driven-embodied-task-planning-egocentric-camera-where-hazards-fall"},{"ref":"CLM-0507","id":"vlm-driven-human-aware-robot-task-planning-over-scene-graphs-bottleneck","capability":"planning","technique":"predict-human-intention-symbolic-goal-then-plan-jointly","statement":"In VLM-driven human-aware robot task planning over scene graphs, the bottleneck is inferring and grounding the symbolic goal, not the planning: once both human and robot goals correctly cover the ground truth, a formal multi-agent planner produces a valid, conflict-free joint plan in 99.6% of trials, while end-to-end success with VLM-inferred human goals is 69.7%.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-17771","stance":"supports"}],"url":"https://rsiratchet.com/claims/vlm-driven-human-aware-robot-task-planning-over-scene-graphs-bottleneck"},{"ref":"CLM-0508","id":"vlms-choosing-ordered-route-over-numbered-waypoints-single-egocentric","capability":"embodied-control","technique":null,"statement":"For VLMs choosing an ordered route over numbered waypoints in a single egocentric image, per-waypoint feasibility and the first action are judged well while the rest of the route is not: pooled over nine models, roughly 88-96% of predictions take a legal first edge, but full-route legality falls to about 47% for a point agent and under 7% once the agent has a footprint, with most of the loss in intermediate edges rather than in output format.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-16610","stance":"supports"}],"url":"https://rsiratchet.com/claims/vlms-choosing-ordered-route-over-numbered-waypoints-single-egocentric"},{"ref":"CLM-0509","id":"vlms-must-operate-unfamiliar-household-devices-no-manual-through","capability":"embodied-control","technique":null,"statement":"When VLMs must operate unfamiliar household devices with no manual, through multi-turn interaction over rendered multi-view images with click/rotate actions at pixel coordinates, the limiting failure is fine-grained localization of the interactive component within the correct view rather than navigating to that view, and almost no task is completed in the optimal number of steps.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-00845","stance":"supports"}],"url":"https://rsiratchet.com/claims/vlms-must-operate-unfamiliar-household-devices-no-manual-through"},{"ref":"CLM-0510","id":"whole-repository-cross-language-code-migration-java-python-supplying-llm-ast-derived","capability":"hallucination","technique":"retrieval-augmented-generation","statement":"In whole-repository cross-language code migration (Java to Python), supplying an LLM with AST-derived dependency-graph context and translating parent classes before children lowers the rate of calls to nonexistent APIs relative to vector-similarity retrieval of text chunks, while lexical similarity metrics such as CodeBLEU show no difference between the two.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-12464","stance":"supports"}],"url":"https://rsiratchet.com/claims/whole-repository-cross-language-code-migration-java-python-supplying-llm-ast-derived"},{"ref":"CLM-0192","id":"withholding-the-users-opinion-is-the-no-training-alternative","capability":"sycophancy","technique":"synthetic-desycophancy-training","statement":"Sycophancy is triggered by the user's stated view being in context, so the cheapest control is not putting it there — ask for the answer before the opinion, or withhold the opinion entirely. Fine-tuning is the answer for the cases where the opinion has to be in context and the answer still must not move.","kind":"mechanism","backing_strength":"mechanism-reasoning","contested":false,"status":"active","sources":[{"source":"arxiv-2308-03958","stance":"supports"}],"url":"https://rsiratchet.com/claims/withholding-the-users-opinion-is-the-no-training-alternative"},{"ref":"CLM-0511","id":"within-one-model-family-relative-distribution-reasoning-failure-modes","capability":"critique","technique":"retrieve-in-context-critiques-weak-model-failure-mode","statement":"Within one model family, the relative distribution of reasoning failure modes stays similar from 1.5B to 72B, so critiques mined offline from small models' wrong answers, retrieved by failure mode and inserted as in-context exemplars, raise a large model's math accuracy above random, fixed, or embedding-similarity exemplar selection at a single generation.","kind":"mechanism","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-27455","stance":"supports"}],"url":"https://rsiratchet.com/claims/within-one-model-family-relative-distribution-reasoning-failure-modes"},{"ref":"CLM-0273","id":"without-training-or-prebuilt-index-searching-module-call-graph","capability":"repository-navigation","technique":"two-phase-call-graph-search-files-broad-then","statement":"Without training or a prebuilt index, searching a module call graph built on the fly for breadth at file level and then expanding it into a function call graph for depth reached about 43-44% function-level Top-1 localization on SWE-bench Lite and Verified with open 32B and 72B models, beating prior LLM localizers, and the better locations raised an issue-solving pipeline's resolve rate by roughly 3 to 30 points depending on model.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2503-22424","stance":"supports"}],"url":"https://rsiratchet.com/claims/without-training-or-prebuilt-index-searching-module-call-graph"},{"ref":"CLM-0657","id":"year-long-multi-agent-social-simulation-driven-single-4-bit-quantized-8b","capability":"long-horizon-execution","technique":"layered-memory-compression-four-daily-activity-blocks","statement":"In a year-long multi-agent social simulation driven by a single 4-bit quantized 8B model on a 12 GB consumer GPU with five agents, the run can be carried to roughly 50-52 simulated weeks before the accumulated prompt hits the context limit, while about 10% of finalized activity records are left with at least one missing field.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2608-24215","stance":"supports"}],"url":"https://rsiratchet.com/claims/year-long-multi-agent-social-simulation-driven-single-4-bit-quantized-8b"},{"ref":"CLM-0232","id":"zero-shot-binary-risk-classification-rare-weather-driven-power-outages-tabular","capability":"forecasting","technique":null,"statement":"On zero-shot binary risk classification of rare weather-driven power outages from tabular weather features, GPT-4o/4.1-class models trail supervised classifiers trained on the same data by 0.04-0.13 macro-F1 and have much lower precision, and they gain little from being given the full forecast window rather than only current observations, while supervised models gain more from forecast features.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04272","stance":"supports"}],"url":"https://rsiratchet.com/claims/zero-shot-binary-risk-classification-rare-weather-driven-power-outages-tabular"},{"ref":"CLM-0328","id":"zero-shot-function-level-small-locally-run-models-llama-3","capability":"error-detection","technique":null,"statement":"Zero-shot, at function level, small locally run models (Llama 3.2, Mistral) identify 43-45% of 349 real Python bugs from BugsInPy, well above a sub-5% keyword baseline, but with accuracy ranging from 0% to 100% by project and lowest (about 22%) on bugs that span several components, so detection is real but shallow and codebase-dependent.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2604-23361","stance":"supports"}],"url":"https://rsiratchet.com/claims/zero-shot-function-level-small-locally-run-models-llama-3"},{"ref":"CLM-0251","id":"zero-shot-llm-program-repair-defects4j-bugs-passing-developer-written-test","capability":"hallucination","technique":"score-intermediate-repair-artifacts-just-patch","statement":"In zero-shot LLM program repair on Defects4J bugs, passing the developer-written test suite does not indicate that the patch is grounded in a correct understanding of the bug: manual inspection of sampled repairs found ungrounded patches in about three quarters of cases including ones that passed all available tests, with incorrect causal localization the largest category.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04909","stance":"supports"}],"url":"https://rsiratchet.com/claims/zero-shot-llm-program-repair-defects4j-bugs-passing-developer-written-test"},{"ref":"CLM-0252","id":"zero-shot-risk-forecasting-tabular-weather-features-rare-event-5-positive","capability":"forecasting","technique":null,"statement":"For zero-shot risk forecasting from tabular weather features on a rare-event (~5% positive) binary severity task, GPT-4o/4.1 class models trail simple supervised classifiers (logistic regression, LightGBM) trained on the same data on macro-F1 and precision at every 3h/6h/12h horizon, and the weaker models close the gap only by flagging nearly every window as high risk.","kind":"observation","backing_strength":"single-paper","contested":false,"status":"pending-review","sources":[{"source":"arxiv-2609-04272","stance":"supports"}],"url":"https://rsiratchet.com/claims/zero-shot-risk-forecasting-tabular-weather-features-rare-event-5-positive"}],"sources":[{"ref":"SRC-0343","id":"arxiv-2601-07786","kind":"paper","title":"\"TODO: Fix the Mess Gemini Created\": Towards Understanding GenAI-Induced Self-Admitted Technical Debt","year":2026,"arxiv_id":"2601.07786","external_url":"https://arxiv.org/abs/2601.07786","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2601-07786"},{"ref":"SRC-0488","id":"arxiv-2609-10226","kind":"paper","title":"$Φ$-Bench: Can Large Language Models Engineer the Infrastructure That Powers Them?","year":2026,"arxiv_id":"2609.10226","external_url":"https://arxiv.org/abs/2609.10226","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10226"},{"ref":"SRC-0648","id":"arxiv-2608-26947","kind":"paper","title":"4DSynth: Controllable Procedural World Synthesis for Dynamic Embodied Simulation","year":2026,"arxiv_id":"2608.26947","external_url":"https://arxiv.org/abs/2608.26947","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26947"},{"ref":"SRC-0566","id":"arxiv-2609-19710","kind":"paper","title":"A Closed-Loop Control Architecture for Reliable Constraint Satisfaction in LLM Text Generation","year":2026,"arxiv_id":"2609.19710","external_url":"https://arxiv.org/abs/2609.19710","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-19710"},{"ref":"SRC-0451","id":"arxiv-2608-29347","kind":"paper","title":"A Cognitive Architecture for Shared Autonomy in AUV Operations","year":2026,"arxiv_id":"2608.29347","external_url":"https://arxiv.org/abs/2608.29347","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29347"},{"ref":"SRC-0634","id":"arxiv-2608-25099","kind":"paper","title":"A Control-Theoretic Approach for Resource-Aware Consensus in Multi-Agent AI","year":2026,"arxiv_id":"2608.25099","external_url":"https://arxiv.org/abs/2608.25099","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25099"},{"ref":"SRC-0368","id":"arxiv-2609-16432","kind":"paper","title":"A light-touch AI literacy intervention helps protect against AI political persuasion","year":2026,"arxiv_id":"2609.16432","external_url":"https://arxiv.org/abs/2609.16432","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-16432"},{"ref":"SRC-0352","id":"arxiv-2604-21579","kind":"paper","title":"A Metamorphic Testing Approach to Diagnosing Memorization in LLM-Based Program Repair","year":2026,"arxiv_id":"2604.21579","external_url":"https://arxiv.org/abs/2604.21579","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2604-21579"},{"ref":"SRC-0203","id":"arxiv-2609-04343","kind":"paper","title":"A Removal Based Approach to Improve LLM Faithfulness at Test-Time","year":2026,"arxiv_id":"2609.04343","external_url":"https://arxiv.org/abs/2609.04343","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-04343"},{"ref":"SRC-0641","id":"arxiv-2608-26008","kind":"paper","title":"A Self-Evolving Multi-Agent Framework Defense against LLM Jailbreak Attacks","year":2026,"arxiv_id":"2608.26008","external_url":"https://arxiv.org/abs/2608.26008","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26008"},{"ref":"SRC-0572","id":"arxiv-2609-21817","kind":"paper","title":"A Sim-to-Real Integration Pipeline for Training and Deployment of Chunk-Based VLA Manipulation Policies","year":2026,"arxiv_id":"2609.21817","external_url":"https://arxiv.org/abs/2609.21817","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-21817"},{"ref":"SRC-0681","id":"arxiv-2609-05069","kind":"paper","title":"A Structured Debate-Mixture-of-Agents Framework for Complex Clinical Diagnostic Decision Support","year":2026,"arxiv_id":"2609.05069","external_url":"https://arxiv.org/abs/2609.05069","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05069"},{"ref":"SRC-0565","id":"arxiv-2609-18311","kind":"paper","title":"A Study on the Impact of Natural Language Differences in Prompts on Automatic Code Generation Using LLMs","year":2026,"arxiv_id":"2609.18311","external_url":"https://arxiv.org/abs/2609.18311","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-18311"},{"ref":"SRC-0252","id":"arxiv-2408-05109","kind":"paper","title":"A Survey of Text-to-SQL in the Era of LLMs: Where are we, and where are we going?","year":2024,"arxiv_id":"2408.05109","external_url":"https://arxiv.org/abs/2408.05109","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2408-05109"},{"ref":"SRC-0052","id":"arxiv-2311-05232","kind":"paper","title":"A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions","year":2023,"arxiv_id":"2311.05232","external_url":"https://arxiv.org/abs/2311.05232","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2311-05232"},{"ref":"SRC-0530","id":"arxiv-2609-24467","kind":"paper","title":"A Temporal Knowledge Graph for Music Festival Lineup Forecasting","year":2026,"arxiv_id":"2609.24467","external_url":"https://arxiv.org/abs/2609.24467","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24467"},{"ref":"SRC-0115","id":"arxiv-2608-24103","kind":"paper","title":"ACE: A Self-Correcting Agentic Canvas Editor for Multi-Slide Presentation Automation","year":2026,"arxiv_id":"2608.24103","external_url":"https://arxiv.org/abs/2608.24103","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24103"},{"ref":"SRC-0544","id":"arxiv-2609-02042","kind":"paper","title":"Act More, Decide Less: Skill-Guided Adaptive Action Chunking for Long-Horizon LLM Agents","year":2026,"arxiv_id":"2609.02042","external_url":"https://arxiv.org/abs/2609.02042","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02042"},{"ref":"SRC-0267","id":"arxiv-2609-14987","kind":"paper","title":"ActGuard: Pre-execution Action Auditing against Indirect Prompt Injection in LLM Agents","year":2026,"arxiv_id":"2609.14987","external_url":"https://arxiv.org/abs/2609.14987","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14987"},{"ref":"SRC-0418","id":"arxiv-2608-10430","kind":"paper","title":"Actionable Hallucination Detection: Translating Latent Uncertainty into Agentic Critique","year":2026,"arxiv_id":"2608.10430","external_url":"https://arxiv.org/abs/2608.10430","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-10430"},{"ref":"SRC-0524","id":"arxiv-2609-24124","kind":"paper","title":"ActiveArena: Benchmarking and Understanding Active Perception in Robotic Manipulation","year":2026,"arxiv_id":"2609.24124","external_url":"https://arxiv.org/abs/2609.24124","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24124"},{"ref":"SRC-0027","id":"arxiv-2305-13300","kind":"paper","title":"Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts","year":2023,"arxiv_id":"2305.13300","external_url":"https://arxiv.org/abs/2305.13300","counts":{"claims":3},"url":"https://rsiratchet.com/sources/arxiv-2305-13300"},{"ref":"SRC-0616","id":"arxiv-2608-11738","kind":"paper","title":"Advancing MLLM-based UAV Image Understanding and Reasoning: A Benchmark and a Training-Free Multi-Agent System","year":2026,"arxiv_id":"2608.11738","external_url":"https://arxiv.org/abs/2608.11738","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-11738"},{"ref":"SRC-0294","id":"arxiv-2609-15963","kind":"paper","title":"Adversarial Testing of Automated Program Repair Agents for Security Vulnerabilities","year":2026,"arxiv_id":"2609.15963","external_url":"https://arxiv.org/abs/2609.15963","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15963"},{"ref":"SRC-0415","id":"arxiv-2609-18520","kind":"paper","title":"AeroWeaver: An Embodied-Agent Harness for Weaving Aerial Skills into Distributed, Adaptive Swarm Execution","year":2026,"arxiv_id":"2609.18520","external_url":"https://arxiv.org/abs/2609.18520","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-18520"},{"ref":"SRC-0443","id":"arxiv-2608-26713","kind":"paper","title":"AesCanvas: A Large-Scale Dataset and Benchmark for Aesthetic Critique and Contextual Suitability","year":2026,"arxiv_id":"2608.26713","external_url":"https://arxiv.org/abs/2608.26713","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26713"},{"ref":"SRC-0601","id":"arxiv-2608-07169","kind":"paper","title":"Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory","year":2026,"arxiv_id":"2608.07169","external_url":"https://arxiv.org/abs/2608.07169","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-07169"},{"ref":"SRC-0149","id":"arxiv-2608-29606","kind":"paper","title":"Agent Zero Memory: Provenance-Aware Long-Term Memory for LLM Agents","year":2026,"arxiv_id":"2608.29606","external_url":"https://arxiv.org/abs/2608.29606","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29606"},{"ref":"SRC-0077","id":"arxiv-2412-14470","kind":"paper","title":"Agent-SafetyBench: Evaluating the Safety of LLM Agents","year":2024,"arxiv_id":"2412.14470","external_url":"https://arxiv.org/abs/2412.14470","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2412-14470"},{"ref":"SRC-0040","id":"arxiv-2308-03688","kind":"paper","title":"AgentBench: Evaluating LLMs as Agents","year":2023,"arxiv_id":"2308.03688","external_url":"https://arxiv.org/abs/2308.03688","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2308-03688"},{"ref":"SRC-0274","id":"arxiv-2609-17800","kind":"paper","title":"AgenTeeth: A Model-Agnostic Framework for Suppressing Hallucination in Frozen Vision-Language Models on Dental X-Rays via Tool Evidence Injection","year":2026,"arxiv_id":"2609.17800","external_url":"https://arxiv.org/abs/2609.17800","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17800"},{"ref":"SRC-0417","id":"arxiv-2605-00334","kind":"paper","title":"AgentFloor: How Far Up the tool use Ladder Can Small Open-Weight Models Go?","year":2026,"arxiv_id":"2605.00334","external_url":"https://arxiv.org/abs/2605.00334","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2605-00334"},{"ref":"SRC-0276","id":"arxiv-2609-08572","kind":"paper","title":"AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems","year":2026,"arxiv_id":"2609.08572","external_url":"https://arxiv.org/abs/2609.08572","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08572"},{"ref":"SRC-0074","id":"arxiv-2410-09024","kind":"paper","title":"AgentHarm: A Benchmark for Measuring Harmfulness of LLM Agents","year":2024,"arxiv_id":"2410.09024","external_url":"https://arxiv.org/abs/2410.09024","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2410-09024"},{"ref":"SRC-0197","id":"masood-2026-agentic-harness-engineering","kind":"post","title":"Agentic Harness Engineering: The Operating System for Machine Intelligence","year":2026,"arxiv_id":null,"external_url":"https://medium.com/@adnanmasood/agentic-harness-engineering-the-operating-system-for-machine-intelligence-ce4fa48e9520","counts":{"claims":1},"url":"https://rsiratchet.com/sources/masood-2026-agentic-harness-engineering"},{"ref":"SRC-0349","id":"arxiv-2603-06064","kind":"paper","title":"Agentic LLM Planning via Step-Wise PDDL Simulation: An Empirical Characterisation","year":2026,"arxiv_id":"2603.06064","external_url":"https://arxiv.org/abs/2603.06064","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2603-06064"},{"ref":"SRC-0380","id":"arxiv-2609-08248","kind":"paper","title":"Agentic ML Exploration (A-MLE) for Ads Ranking","year":2026,"arxiv_id":"2609.08248","external_url":"https://arxiv.org/abs/2609.08248","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-08248"},{"ref":"SRC-0453","id":"arxiv-2608-29622","kind":"paper","title":"AgenticRag-R1: Agentic Reinforcement Learning with Stack Memory for Multi-Step Reasoning, Retrieval and Memorizing","year":2026,"arxiv_id":"2608.29622","external_url":"https://arxiv.org/abs/2608.29622","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29622"},{"ref":"SRC-0310","id":"arxiv-2407-01489","kind":"paper","title":"Agentless: Demystifying LLM-based Software Engineering Agents","year":2024,"arxiv_id":"2407.01489","external_url":"https://arxiv.org/abs/2407.01489","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2407-01489"},{"ref":"SRC-0630","id":"arxiv-2608-24215","kind":"paper","title":"Agentopia on a Consumer GPU: A Reduced-Scale Long-Horizon Port with an 8B Model","year":2026,"arxiv_id":"2608.24215","external_url":"https://arxiv.org/abs/2608.24215","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24215"},{"ref":"SRC-0678","id":"arxiv-2609-01779","kind":"paper","title":"Agents That Model Agents: Five Principles Toward a Theory of Mind for 6G Networks","year":2026,"arxiv_id":"2609.01779","external_url":"https://arxiv.org/abs/2609.01779","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01779"},{"ref":"SRC-0503","id":"arxiv-2609-21386","kind":"paper","title":"AgentVidBench: A Multi-Hop Video Question Answering Benchmark for Evaluating MLLM Agents","year":2026,"arxiv_id":"2609.21386","external_url":"https://arxiv.org/abs/2609.21386","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21386"},{"ref":"SRC-0452","id":"arxiv-2608-29537","kind":"paper","title":"AGM: Achievement-Grounded Memory for Closed-Loop Agents with Frozen VLA Policies","year":2026,"arxiv_id":"2608.29537","external_url":"https://arxiv.org/abs/2608.29537","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29537"},{"ref":"SRC-0124","id":"arxiv-2608-25667","kind":"paper","title":"AI Slop and Hallucinations in Vulnerability Assessment: A Survey on Reasoning Failures and Trustworthy Mitigation","year":2026,"arxiv_id":"2608.25667","external_url":"https://arxiv.org/abs/2608.25667","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25667"},{"ref":"SRC-0082","id":"arxiv-2512-22387","kind":"paper","title":"AI-Generated Code Is Not Reproducible (Yet): An Empirical Study of Dependency Gaps in LLM-Based Coding Agents","year":2025,"arxiv_id":"2512.22387","external_url":"https://arxiv.org/abs/2512.22387","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2512-22387"},{"ref":"SRC-0356","id":"arxiv-2605-02741","kind":"paper","title":"AI-Generated Smells: An Analysis of Code and Architecture in LLM and Agent-Driven Development","year":2026,"arxiv_id":"2605.02741","external_url":"https://arxiv.org/abs/2605.02741","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2605-02741"},{"ref":"SRC-0273","id":"arxiv-2609-15820","kind":"paper","title":"AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery","year":2026,"arxiv_id":"2609.15820","external_url":"https://arxiv.org/abs/2609.15820","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15820"},{"ref":"SRC-0397","id":"arxiv-2609-14572","kind":"paper","title":"AlgoRAG: Retrieval-Augmented Generation for Theoretical Computer Science Education -- A Comprehensive Evaluation Framework for Algorithm Analysis and Complexity Theory","year":2026,"arxiv_id":"2609.14572","external_url":"https://arxiv.org/abs/2609.14572","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14572"},{"ref":"SRC-0367","id":"arxiv-2609-00565","kind":"paper","title":"Aligned but Flattened: Analyzing the Trade-off between Cultural Alignment and Diversity in LLMs","year":2026,"arxiv_id":"2609.00565","external_url":"https://arxiv.org/abs/2609.00565","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00565"},{"ref":"SRC-0353","id":"arxiv-2604-23361","kind":"paper","title":"An Empirical Evaluation of Locally Deployed LLMs for Bug Detection in Python Code","year":2026,"arxiv_id":"2604.23361","external_url":"https://arxiv.org/abs/2604.23361","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2604-23361"},{"ref":"SRC-0321","id":"arxiv-2606-29602","kind":"paper","title":"An Empirical Evaluation of Prompt Injection Vulnerabilities in Large Language Models Across Multilingual and Obfuscated Attack Scenarios","year":2026,"arxiv_id":"2606.29602","external_url":"https://arxiv.org/abs/2606.29602","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2606-29602"},{"ref":"SRC-0573","id":"arxiv-2609-22973","kind":"paper","title":"An Empirical Study and Open Testbed for Federated Fine-Tuning of Vision-Language-Action Models","year":2026,"arxiv_id":"2609.22973","external_url":"https://arxiv.org/abs/2609.22973","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-22973"},{"ref":"SRC-0406","id":"arxiv-2609-17119","kind":"paper","title":"An Empirical Study of Counterfactual Self-Explanations in LLMs","year":2026,"arxiv_id":"2609.17119","external_url":"https://arxiv.org/abs/2609.17119","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-17119"},{"ref":"SRC-0584","id":"scheel-2021-excess-of-positive-results","kind":"paper","title":"An Excess of Positive Results: Comparing the Standard Psychology Literature With Registered Reports","year":2021,"arxiv_id":null,"external_url":"https://doi.org/10.1177/25152459211007467","counts":{"claims":1},"url":"https://rsiratchet.com/sources/scheel-2021-excess-of-positive-results"},{"ref":"SRC-0228","id":"arxiv-2609-09404","kind":"paper","title":"An Experimental Evaluation of Multimodal Prompt Injection Attacks on Agentic AI Frameworks","year":2026,"arxiv_id":"2609.09404","external_url":"https://arxiv.org/abs/2609.09404","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09404"},{"ref":"SRC-0510","id":"arxiv-2609-22043","kind":"paper","title":"An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency","year":2026,"arxiv_id":"2609.22043","external_url":"https://arxiv.org/abs/2609.22043","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22043"},{"ref":"SRC-0517","id":"arxiv-2609-22917","kind":"paper","title":"An Iterative LangGraph Agent for Text-to-SQL: Natural Language Access to the Chicago Crime Database","year":2026,"arxiv_id":"2609.22917","external_url":"https://arxiv.org/abs/2609.22917","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22917"},{"ref":"SRC-0638","id":"arxiv-2608-25869","kind":"paper","title":"Anchoring Bias in LLM-as-a-Judge Systems: Prior Scores Compromise Evaluation Independence","year":2026,"arxiv_id":"2608.25869","external_url":"https://arxiv.org/abs/2608.25869","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25869"},{"ref":"SRC-0382","id":"arxiv-2608-30345","kind":"paper","title":"Answer Probing-Guided Search for Diverse Solution Exploration of LLMs","year":2026,"arxiv_id":"2608.30345","external_url":"https://arxiv.org/abs/2608.30345","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-30345"},{"ref":"SRC-0060","id":"arxiv-2402-18563","kind":"paper","title":"Approaching Human-Level Forecasting with Language Models","year":2024,"arxiv_id":"2402.18563","external_url":"https://arxiv.org/abs/2402.18563","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2402-18563"},{"ref":"SRC-0385","id":"arxiv-2609-10707","kind":"paper","title":"Architecting the Secure AI-SOC: A Neurosymbolic Framework for Pipeline Integrity and Threat Mitigation","year":2026,"arxiv_id":"2609.10707","external_url":"https://arxiv.org/abs/2609.10707","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10707"},{"ref":"SRC-0529","id":"arxiv-2609-24417","kind":"paper","title":"ARM: Attention with Routed-Memory for Learnable Sparse Control","year":2026,"arxiv_id":"2609.24417","external_url":"https://arxiv.org/abs/2609.24417","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24417"},{"ref":"SRC-0007","id":"arxiv-2108-09293","kind":"paper","title":"Asleep at the Keyboard? Assessing the Security of GitHub Copilot's Code Contributions","year":2021,"arxiv_id":"2108.09293","external_url":"https://arxiv.org/abs/2108.09293","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2108-09293"},{"ref":"SRC-0402","id":"arxiv-2609-15721","kind":"paper","title":"Assembling the CREW: A Collaborative Multi-agent Reinforcement Learning Framework for Automated Related Work Generation","year":2026,"arxiv_id":"2609.15721","external_url":"https://arxiv.org/abs/2609.15721","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15721"},{"ref":"SRC-0504","id":"arxiv-2609-21461","kind":"paper","title":"AtomEgo: Exploring Ego-Robot Integration for Embodied Foundation Model Pretraining","year":2026,"arxiv_id":"2609.21461","external_url":"https://arxiv.org/abs/2609.21461","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21461"},{"ref":"SRC-0414","id":"arxiv-2609-18320","kind":"paper","title":"Attention Dispersion as a Diagnostic Signal for Hallucination in Large Language Models","year":2026,"arxiv_id":"2609.18320","external_url":"https://arxiv.org/abs/2609.18320","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-18320"},{"ref":"SRC-0282","id":"arxiv-2609-11252","kind":"paper","title":"AudioICL-Bench: A Benchmark for Large Audio Language Model In-Context Learning","year":2026,"arxiv_id":"2609.11252","external_url":"https://arxiv.org/abs/2609.11252","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-11252"},{"ref":"SRC-0482","id":"arxiv-2609-09356","kind":"paper","title":"Auditable Emergency Triage for Maternal and Newborn Care in India","year":2026,"arxiv_id":"2609.09356","external_url":"https://arxiv.org/abs/2609.09356","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09356"},{"ref":"SRC-0557","id":"arxiv-2609-10922","kind":"paper","title":"Auto-RecSys: Harnessing Autonomous Research Agents for Industry-Scale Recommender System","year":2026,"arxiv_id":"2609.10922","external_url":"https://arxiv.org/abs/2609.10922","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-10922"},{"ref":"SRC-0326","id":"arxiv-2510-06857","kind":"paper","title":"Autoformalizer with Tool Feedback","year":2025,"arxiv_id":"2510.06857","external_url":"https://arxiv.org/abs/2510.06857","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2510-06857"},{"ref":"SRC-0254","id":"huang-song-2025-xpath-generation-agents","kind":"paper","title":"Automatic XPath generation agents for vertical websites by LLMs","year":2025,"arxiv_id":null,"external_url":"https://doi.org/10.1007/s44443-025-00071-w","counts":{"claims":1},"url":"https://rsiratchet.com/sources/huang-song-2025-xpath-generation-agents"},{"ref":"SRC-0507","id":"arxiv-2609-21940","kind":"paper","title":"AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory","year":2026,"arxiv_id":"2609.21940","external_url":"https://arxiv.org/abs/2609.21940","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21940"},{"ref":"SRC-0096","id":"arxiv-2608-18490","kind":"paper","title":"Bayesian Partner Modelling enables Adaptive Replanning for LLM Coordination","year":2026,"arxiv_id":"2608.18490","external_url":"https://arxiv.org/abs/2608.18490","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-18490"},{"ref":"SRC-0442","id":"arxiv-2608-25761","kind":"paper","title":"Beam Search, Self-Consistency, and the Limits of Inference-Time Scaling for Grammar-Constrained Text-to-SQL in Small Language Models","year":2026,"arxiv_id":"2608.25761","external_url":"https://arxiv.org/abs/2608.25761","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25761"},{"ref":"SRC-0444","id":"arxiv-2608-26867","kind":"paper","title":"BekchiAI: Measuring, Observing, and Controlling LLM Agents in One Click","year":2026,"arxiv_id":"2608.26867","external_url":"https://arxiv.org/abs/2608.26867","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26867"},{"ref":"SRC-0441","id":"arxiv-2608-25152","kind":"paper","title":"Belief Cascades Drive Persuasion in LLM Agent Networks","year":2026,"arxiv_id":"2608.25152","external_url":"https://arxiv.org/abs/2608.25152","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25152"},{"ref":"SRC-0428","id":"arxiv-2608-20975","kind":"paper","title":"Belief Without Behavior: Measuring the Translation of Theory of Mind into Coordinated Social Action in Vision-Language Models","year":2026,"arxiv_id":"2608.20975","external_url":"https://arxiv.org/abs/2608.20975","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-20975"},{"ref":"SRC-0486","id":"arxiv-2609-10036","kind":"paper","title":"Belief-State Engine: Augmenting LLMs for Principled Planning Under Partial Observability","year":2026,"arxiv_id":"2609.10036","external_url":"https://arxiv.org/abs/2609.10036","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10036"},{"ref":"SRC-0229","id":"arxiv-2609-09526","kind":"paper","title":"Benchmarking Agentic HLS Design Tasks With HLS-Eval","year":2026,"arxiv_id":"2609.09526","external_url":"https://arxiv.org/abs/2609.09526","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09526"},{"ref":"SRC-0129","id":"arxiv-2608-26199","kind":"paper","title":"Benchmarking AI Agents for Hardware Design Automation via MCP Tool Calling","year":2026,"arxiv_id":"2608.26199","external_url":"https://arxiv.org/abs/2608.26199","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26199"},{"ref":"SRC-0645","id":"arxiv-2608-26592","kind":"paper","title":"Benchmarking Clinical Decision Pathway Adherence in Large Language Models","year":2026,"arxiv_id":"2608.26592","external_url":"https://arxiv.org/abs/2608.26592","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26592"},{"ref":"SRC-0361","id":"arxiv-2606-10554","kind":"paper","title":"Benchmarking Knowledge Editing using Logical Rules","year":2026,"arxiv_id":"2606.10554","external_url":"https://arxiv.org/abs/2606.10554","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-10554"},{"ref":"SRC-0107","id":"arxiv-2608-22529","kind":"paper","title":"Benchmarking the Titans: A Multi-Dimensional Empirical Evaluation of LLM Code Generation Quality in the .NET Ecosystem","year":2026,"arxiv_id":"2608.22529","external_url":"https://arxiv.org/abs/2608.22529","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-22529"},{"ref":"SRC-0206","id":"arxiv-2609-04909","kind":"paper","title":"Better Understanding, Better Fixes? A Study of Hallucination in LLM-based Automated Program Repair","year":2026,"arxiv_id":"2609.04909","external_url":"https://arxiv.org/abs/2609.04909","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-04909"},{"ref":"SRC-0231","id":"arxiv-2609-09681","kind":"paper","title":"Beyond Accuracy: ARIA-Rubrics for Evaluating Audio Reasoning in Large Audio Language Models","year":2026,"arxiv_id":"2609.09681","external_url":"https://arxiv.org/abs/2609.09681","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09681"},{"ref":"SRC-0108","id":"arxiv-2608-22753","kind":"paper","title":"Beyond Factual Knowledge: Benchmarking and Learning Step-Level Procedural Rule Reasoning in Large Language Models","year":2026,"arxiv_id":"2608.22753","external_url":"https://arxiv.org/abs/2608.22753","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-22753"},{"ref":"SRC-0455","id":"arxiv-2608-29990","kind":"paper","title":"Beyond Fluency: A Rubric-Based Benchmark for Evaluating Saudi Dialect and Cultural Competence in Large Language Models","year":2026,"arxiv_id":"2608.29990","external_url":"https://arxiv.org/abs/2608.29990","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29990"},{"ref":"SRC-0218","id":"arxiv-2609-07370","kind":"paper","title":"Beyond Fluent Generation: A CPU Reliability Benchmark for MCP-Style Tool Calling in Sub-2B Small Language Models for Edge Deployment","year":2026,"arxiv_id":"2609.07370","external_url":"https://arxiv.org/abs/2609.07370","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07370"},{"ref":"SRC-0609","id":"arxiv-2608-09344","kind":"paper","title":"Beyond Global Editing: Per-Instance Disentangled Subspaces for Training-Free Hallucination Mitigation in LVLMs","year":2026,"arxiv_id":"2608.09344","external_url":"https://arxiv.org/abs/2608.09344","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-09344"},{"ref":"SRC-0679","id":"arxiv-2609-01895","kind":"paper","title":"Beyond Human-Likeness: Mapping the Scientific Critique Profiles of LLMs and Human Reviewers","year":2026,"arxiv_id":"2609.01895","external_url":"https://arxiv.org/abs/2609.01895","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01895"},{"ref":"SRC-0165","id":"arxiv-2609-00231","kind":"paper","title":"Beyond Language Priors: Diagnosing and Fixing Visual-Origin Hallucinations in Multimodal LLM","year":2026,"arxiv_id":"2609.00231","external_url":"https://arxiv.org/abs/2609.00231","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-00231"},{"ref":"SRC-0186","id":"arxiv-2609-04131","kind":"paper","title":"Beyond Retrieval: Progressive Latent Memory Evolution for Streaming Video Understanding","year":2026,"arxiv_id":"2609.04131","external_url":"https://arxiv.org/abs/2609.04131","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-04131"},{"ref":"SRC-0518","id":"arxiv-2609-22977","kind":"paper","title":"Beyond Similarity: Coverage-Aware Prompt Selection for Time Series Forecasting with LLMs","year":2026,"arxiv_id":"2609.22977","external_url":"https://arxiv.org/abs/2609.22977","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22977"},{"ref":"SRC-0287","id":"arxiv-2609-12464","kind":"paper","title":"Beyond Vector Similarity: Hierarchical Context-Aware Graph RAG vs Standard RAG in Enterprise Code Migration","year":2026,"arxiv_id":"2609.12464","external_url":"https://arxiv.org/abs/2609.12464","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-12464"},{"ref":"SRC-0495","id":"arxiv-2609-18306","kind":"paper","title":"Bias Amplification in Multi-Agent Network: How Biased Agents Shape Opinions and Rhetoric","year":2026,"arxiv_id":"2609.18306","external_url":"https://arxiv.org/abs/2609.18306","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-18306"},{"ref":"SRC-0561","id":"arxiv-2609-14988","kind":"paper","title":"Biomedical Reference Generation Remains Unreliable across 26 Large Language Models","year":2026,"arxiv_id":"2609.14988","external_url":"https://arxiv.org/abs/2609.14988","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14988"},{"ref":"SRC-0450","id":"arxiv-2608-29345","kind":"paper","title":"BIRD-History: A Benchmark for History-Driven Text-to-SQL with Fine-Grained Knowledge Annotations","year":2026,"arxiv_id":"2608.29345","external_url":"https://arxiv.org/abs/2608.29345","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29345"},{"ref":"SRC-0371","id":"arxiv-2210-15723","kind":"paper","title":"Birdwatch: Crowd Wisdom and Bridging Algorithms can Inform Understanding and Reduce the Spread of Misinformation","year":2022,"arxiv_id":"2210.15723","external_url":"https://arxiv.org/abs/2210.15723","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2210-15723"},{"ref":"SRC-0492","id":"arxiv-2609-16305","kind":"paper","title":"BLINDSPOT: A Benchmark for Safety and Refusal Calibration in Long-Horizon Tool-Using Agents","year":2026,"arxiv_id":"2609.16305","external_url":"https://arxiv.org/abs/2609.16305","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-16305"},{"ref":"SRC-0241","id":"arxiv-2608-27420","kind":"paper","title":"Boosting LLM Exploration via Weak-Model Guidance in RLVR","year":2026,"arxiv_id":"2608.27420","external_url":"https://arxiv.org/abs/2608.27420","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27420"},{"ref":"SRC-0531","id":"arxiv-2609-24525","kind":"paper","title":"Bridge3D: Enabling Vision-Language-Action Models to See and Act in 3D","year":2026,"arxiv_id":"2609.24525","external_url":"https://arxiv.org/abs/2609.24525","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24525"},{"ref":"SRC-0661","id":"arxiv-2608-29379","kind":"paper","title":"Bridging Semantics and Physics with Constrained LLMs for Safe and Trustworthy Robotic Manipulation","year":2026,"arxiv_id":"2608.29379","external_url":"https://arxiv.org/abs/2608.29379","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29379"},{"ref":"SRC-0289","id":"arxiv-2609-13335","kind":"paper","title":"Bridging Thought and Action: Taming Long-Horizon Instability in Open-Source LLM Agents with a MetaTool-Enhanced ROS Framework","year":2026,"arxiv_id":"2609.13335","external_url":"https://arxiv.org/abs/2609.13335","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-13335"},{"ref":"SRC-0393","id":"arxiv-2609-13498","kind":"paper","title":"Building a Production Greek-English Speech Recognizer","year":2026,"arxiv_id":"2609.13498","external_url":"https://arxiv.org/abs/2609.13498","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-13498"},{"ref":"SRC-0399","id":"arxiv-2609-14739","kind":"paper","title":"Building Legal Reward Models for Grounding and Abstention","year":2026,"arxiv_id":"2609.14739","external_url":"https://arxiv.org/abs/2609.14739","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14739"},{"ref":"SRC-0081","id":"arxiv-2510-03215","kind":"paper","title":"Cache-to-Cache: Direct Semantic Communication Between Large Language Models","year":2025,"arxiv_id":"2510.03215","external_url":"https://arxiv.org/abs/2510.03215","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2510-03215"},{"ref":"SRC-0172","id":"arxiv-2609-00949","kind":"paper","title":"Calibration is the Bottleneck: An Action-Class Diagnostic of Multi-Turn Tool-Calling","year":2026,"arxiv_id":"2609.00949","external_url":"https://arxiv.org/abs/2609.00949","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00949"},{"ref":"SRC-0358","id":"arxiv-2605-17255","kind":"paper","title":"CAM-Bench: A Benchmark for Computational and Applied Mathematics in Lean","year":2026,"arxiv_id":"2605.17255","external_url":"https://arxiv.org/abs/2605.17255","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2605-17255"},{"ref":"SRC-0169","id":"arxiv-2609-00747","kind":"paper","title":"Can Large Language Models Forecast What Researchers Study Next?","year":2026,"arxiv_id":"2609.00747","external_url":"https://arxiv.org/abs/2609.00747","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00747"},{"ref":"SRC-0558","id":"arxiv-2609-11141","kind":"paper","title":"Can LLMs Normalize Databases? A Benchmark and Multi-Agent Framework for Schema Normalization","year":2026,"arxiv_id":"2609.11141","external_url":"https://arxiv.org/abs/2609.11141","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-11141"},{"ref":"SRC-0323","id":"arxiv-2510-02340","kind":"paper","title":"Can Prompts Rewind Time for LLMs? Evaluating the Effectiveness of Prompted Knowledge Cutoffs","year":2025,"arxiv_id":"2510.02340","external_url":"https://arxiv.org/abs/2510.02340","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2510-02340"},{"ref":"SRC-0235","id":"arxiv-2609-09895","kind":"paper","title":"Can We Trust Video Hallucination Detectors? VidHalLoc for Evaluating the Evaluators","year":2026,"arxiv_id":"2609.09895","external_url":"https://arxiv.org/abs/2609.09895","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09895"},{"ref":"SRC-0599","id":"arxiv-2608-06756","kind":"paper","title":"Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence","year":2026,"arxiv_id":"2608.06756","external_url":"https://arxiv.org/abs/2608.06756","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-06756"},{"ref":"SRC-0177","id":"arxiv-2609-02265","kind":"paper","title":"CAPTURE: Disentangling Preference Drift from Memory Poisoning in Personalized LLM Agents","year":2026,"arxiv_id":"2609.02265","external_url":"https://arxiv.org/abs/2609.02265","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02265"},{"ref":"SRC-0153","id":"arxiv-2608-30147","kind":"paper","title":"CAST: Critique-Aware Supervision for Training Reliable Long-Horizon Tool-Calling Agents","year":2026,"arxiv_id":"2608.30147","external_url":"https://arxiv.org/abs/2608.30147","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30147"},{"ref":"SRC-0011","id":"arxiv-2201-11903","kind":"paper","title":"Chain-of-Thought Prompting Elicits Reasoning in Large Language Models","year":2022,"arxiv_id":"2201.11903","external_url":"https://arxiv.org/abs/2201.11903","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2201-11903"},{"ref":"SRC-0469","id":"arxiv-2609-01352","kind":"paper","title":"CHARM: Character Hallucination for Multicultural Role Play Benchmark","year":2026,"arxiv_id":"2609.01352","external_url":"https://arxiv.org/abs/2609.01352","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01352"},{"ref":"SRC-0527","id":"arxiv-2609-24210","kind":"paper","title":"ChartJudgeBench: Evaluating LMM Judges for Chart-to-Code Generation","year":2026,"arxiv_id":"2609.24210","external_url":"https://arxiv.org/abs/2609.24210","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24210"},{"ref":"SRC-0596","id":"arxiv-2608-05631","kind":"paper","title":"ChronoVision: Temporal Reasoning via Latent State Reconstruction","year":2026,"arxiv_id":"2608.05631","external_url":"https://arxiv.org/abs/2608.05631","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-05631"},{"ref":"SRC-0219","id":"arxiv-2609-07434","kind":"paper","title":"CIT-CAD: Constraint Intent Tree-based CAD Code Generation and Verification","year":2026,"arxiv_id":"2609.07434","external_url":"https://arxiv.org/abs/2609.07434","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07434"},{"ref":"SRC-0548","id":"arxiv-2609-02459","kind":"paper","title":"CivBench: A Long-Horizon Benchmark for Tool-Mediated Agents in Civilization VI","year":2026,"arxiv_id":"2609.02459","external_url":"https://arxiv.org/abs/2609.02459","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02459"},{"ref":"SRC-0432","id":"arxiv-2608-22483","kind":"paper","title":"Claim-Level Confidence Calibration for Reliable Decision Making with Large Language Models","year":2026,"arxiv_id":"2608.22483","external_url":"https://arxiv.org/abs/2608.22483","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-22483"},{"ref":"SRC-0344","id":"arxiv-2601-17397","kind":"paper","title":"CLM-Bench: Benchmarking and Analyzing Cross-lingual Misalignment of LLMs in Knowledge Editing","year":2026,"arxiv_id":"2601.17397","external_url":"https://arxiv.org/abs/2601.17397","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2601-17397"},{"ref":"SRC-0498","id":"arxiv-2609-19425","kind":"paper","title":"Closed-World Resolution Against Tool Hallucination in LLM Agents","year":2026,"arxiv_id":"2609.19425","external_url":"https://arxiv.org/abs/2609.19425","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-19425"},{"ref":"SRC-0277","id":"arxiv-2609-08832","kind":"paper","title":"Closing the Consistency Gap: Self-Evolving Agents That Learn to Stay on Course","year":2026,"arxiv_id":"2609.08832","external_url":"https://arxiv.org/abs/2609.08832","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08832"},{"ref":"SRC-0669","id":"arxiv-2608-30498","kind":"paper","title":"CM2: Multimodal Cultural Reasoning via an Integrated Multi-Agent Framework","year":2026,"arxiv_id":"2608.30498","external_url":"https://arxiv.org/abs/2608.30498","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30498"},{"ref":"SRC-0419","id":"arxiv-2608-17253","kind":"paper","title":"Co-RL: Unsupervised Reasoning Emerges from Diverse Cohort in Multi-agent RL","year":2026,"arxiv_id":"2608.17253","external_url":"https://arxiv.org/abs/2608.17253","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-17253"},{"ref":"SRC-0686","id":"arxiv-2609-20842","kind":"paper","title":"COAL-SQL: Coverage-Guided Augmentation and Failure-Driven Learning for Text-to-SQL Post-Training","year":2026,"arxiv_id":"2609.20842","external_url":"https://arxiv.org/abs/2609.20842","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-20842"},{"ref":"SRC-0433","id":"arxiv-2608-22610","kind":"paper","title":"Coalition-Aware Skill Reliability for Self-Evolving Agents","year":2026,"arxiv_id":"2608.22610","external_url":"https://arxiv.org/abs/2608.22610","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-22610"},{"ref":"SRC-0657","id":"arxiv-2608-28266","kind":"paper","title":"CoCoBench: A Cooperative Coordination Benchmark for Embodied Multi-Agent Task Planning","year":2026,"arxiv_id":"2608.28266","external_url":"https://arxiv.org/abs/2608.28266","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-28266"},{"ref":"SRC-0088","id":"arxiv-2605-18747","kind":"paper","title":"Code as Agent Harness","year":2026,"arxiv_id":"2605.18747","external_url":"https://arxiv.org/abs/2605.18747","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2605-18747"},{"ref":"SRC-0569","id":"arxiv-2609-20822","kind":"paper","title":"Coding Agents with an Obstacle-Aware Harness for Safe Robot Manipulation","year":2026,"arxiv_id":"2609.20822","external_url":"https://arxiv.org/abs/2609.20822","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-20822"},{"ref":"SRC-0259","id":"arxiv-2609-07529","kind":"paper","title":"CoER: Defending against Adaptive Indirect Prompt Injection via Adversarial Co-Evolution and Refinement","year":2026,"arxiv_id":"2609.07529","external_url":"https://arxiv.org/abs/2609.07529","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07529"},{"ref":"SRC-0429","id":"arxiv-2608-21030","kind":"paper","title":"COMET: Contrastive Motion-Enhanced Temporal Reasoning for Video Multimodal Large Language Models","year":2026,"arxiv_id":"2608.21030","external_url":"https://arxiv.org/abs/2608.21030","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-21030"},{"ref":"SRC-0170","id":"arxiv-2609-00759","kind":"paper","title":"Compile, Don't Memorize: A Context Compilation Architecture (CCA) for In-Context Learning","year":2026,"arxiv_id":"2609.00759","external_url":"https://arxiv.org/abs/2609.00759","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00759"},{"ref":"SRC-0182","id":"arxiv-2609-03156","kind":"paper","title":"Compound Prompt Constraints in LLM Code Generation: A Factorial Study of Format, Persona, and Urgency","year":2026,"arxiv_id":"2609.03156","external_url":"https://arxiv.org/abs/2609.03156","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-03156"},{"ref":"SRC-0378","id":"hall-2026-publish-or-perish-reply","kind":"post","title":"Consequence of a \"publish or perish\" academic mindset rather than anything serious","year":2026,"arxiv_id":null,"external_url":"https://x.com/thesupermannx/status/2102027727192932469","counts":{"claims":1},"url":"https://rsiratchet.com/sources/hall-2026-publish-or-perish-reply"},{"ref":"SRC-0020","id":"arxiv-2212-08073","kind":"paper","title":"Constitutional AI: Harmlessness from AI Feedback","year":2022,"arxiv_id":"2212.08073","external_url":"https://arxiv.org/abs/2212.08073","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2212-08073"},{"ref":"SRC-0574","id":"arxiv-2609-23742","kind":"paper","title":"Constrained Decoding Eliminates Structural Failures in Small LLMs but Reveals a Scale-Dependent Semantic Gap","year":2026,"arxiv_id":"2609.23742","external_url":"https://arxiv.org/abs/2609.23742","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-23742"},{"ref":"SRC-0140","id":"arxiv-2608-28476","kind":"paper","title":"ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL","year":2026,"arxiv_id":"2608.28476","external_url":"https://arxiv.org/abs/2608.28476","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-28476"},{"ref":"SRC-0603","id":"arxiv-2608-07642","kind":"paper","title":"Contextual Value Alignment via Multilayer Combinatorial Fusion","year":2026,"arxiv_id":"2608.07642","external_url":"https://arxiv.org/abs/2608.07642","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-07642"},{"ref":"SRC-0086","id":"arxiv-2605-09998","kind":"paper","title":"Continual Harness: Online Adaptation for Self-Improving Foundation Agents","year":2026,"arxiv_id":"2605.09998","external_url":"https://arxiv.org/abs/2605.09998","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2605-09998"},{"ref":"SRC-0237","id":"arxiv-2609-10441","kind":"paper","title":"ConvMem: Convolutional Memory for Long-Context Reasoning","year":2026,"arxiv_id":"2609.10441","external_url":"https://arxiv.org/abs/2609.10441","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10441"},{"ref":"SRC-0002","id":"arxiv-1606-03137","kind":"paper","title":"Cooperative Inverse Reinforcement Learning","year":2016,"arxiv_id":"1606.03137","external_url":"https://arxiv.org/abs/1606.03137","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-1606-03137"},{"ref":"SRC-0540","id":"arxiv-2603-04647","kind":"paper","title":"Coordinated Semantic Alignment and Evidence Constraints for Retrieval-Augmented Generation with Large Language Models","year":2026,"arxiv_id":"2603.04647","external_url":"https://arxiv.org/abs/2603.04647","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2603-04647"},{"ref":"SRC-0471","id":"arxiv-2609-01600","kind":"paper","title":"CordisBench: Can Language Models Reason About Component Lifecycles in Dynamic Agent Harnesses?","year":2026,"arxiv_id":"2609.01600","external_url":"https://arxiv.org/abs/2609.01600","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01600"},{"ref":"SRC-0420","id":"arxiv-2608-17427","kind":"paper","title":"Counterfactual Anatomy-guided Spatial-Temporal Decoding for Annotation-Free Hallucination Mitigation in Medical VLMs","year":2026,"arxiv_id":"2608.17427","external_url":"https://arxiv.org/abs/2608.17427","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-17427"},{"ref":"SRC-0604","id":"arxiv-2608-07838","kind":"paper","title":"Counterfactual Benchmarking and Training for Factuality Consistency and Order-Robust Grounded Reasoning in LLMs over Heterogeneous Knowledge","year":2026,"arxiv_id":"2608.07838","external_url":"https://arxiv.org/abs/2608.07838","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-07838"},{"ref":"SRC-0214","id":"arxiv-2609-06704","kind":"paper","title":"Counterfactual Tests for Measuring Chain-of-Thought Faithfulness in Visual Language Models","year":2026,"arxiv_id":"2609.06704","external_url":"https://arxiv.org/abs/2609.06704","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-06704"},{"ref":"SRC-0456","id":"arxiv-2608-30158","kind":"paper","title":"CPR for LLMs: Critical-Point Routing against Catastrophic Forgetting in Domain Adaptation","year":2026,"arxiv_id":"2608.30158","external_url":"https://arxiv.org/abs/2608.30158","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30158"},{"ref":"SRC-0436","id":"arxiv-2608-23244","kind":"paper","title":"Credal Large Language Models for Semantic Commitment under Uncertainty","year":2026,"arxiv_id":"2608.23244","external_url":"https://arxiv.org/abs/2608.23244","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-23244"},{"ref":"SRC-0447","id":"arxiv-2608-27455","kind":"paper","title":"CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes","year":2026,"arxiv_id":"2608.27455","external_url":"https://arxiv.org/abs/2608.27455","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27455"},{"ref":"SRC-0146","id":"arxiv-2608-29352","kind":"paper","title":"Cross-Relational Preference Learning for Better LLM Instruction Following","year":2026,"arxiv_id":"2608.29352","external_url":"https://arxiv.org/abs/2608.29352","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29352"},{"ref":"SRC-0243","id":"arxiv-2608-28405","kind":"paper","title":"CultureConverse: A Multilingual Multi-turn Simulation Harness for Culturally Grounded Assistance in East and Southeast Asia","year":2026,"arxiv_id":"2608.28405","external_url":"https://arxiv.org/abs/2608.28405","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-28405"},{"ref":"SRC-0473","id":"arxiv-2609-05708","kind":"paper","title":"CUSP: Decomposable Collective Uncertainty for Multi-Agent Multimodal Reasoning","year":2026,"arxiv_id":"2609.05708","external_url":"https://arxiv.org/abs/2609.05708","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05708"},{"ref":"SRC-0351","id":"arxiv-2604-20389","kind":"paper","title":"CyberCertBench: Evaluating LLMs in Cybersecurity Certification Knowledge","year":2026,"arxiv_id":"2604.20389","external_url":"https://arxiv.org/abs/2604.20389","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2604-20389"},{"ref":"SRC-0262","id":"arxiv-2609-13808","kind":"paper","title":"DARE: Dialectical Agentic Reasoning for Structured Knowledge Fact Checking","year":2026,"arxiv_id":"2609.13808","external_url":"https://arxiv.org/abs/2609.13808","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-13808"},{"ref":"SRC-0097","id":"arxiv-2608-18524","kind":"paper","title":"DART-SD: Diamond-topology Aware Retrieval and Tuning for Self-Distillation of Multi-Turn Tool-Calling Agents","year":2026,"arxiv_id":"2608.18524","external_url":"https://arxiv.org/abs/2608.18524","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-18524"},{"ref":"SRC-0079","id":"arxiv-2505-22954","kind":"paper","title":"Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents","year":2025,"arxiv_id":"2505.22954","external_url":"https://arxiv.org/abs/2505.22954","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2505-22954"},{"ref":"SRC-0666","id":"arxiv-2608-29966","kind":"paper","title":"DataFoundry: Evolving Data Preparators via Recursive Self-Improvement","year":2026,"arxiv_id":"2608.29966","external_url":"https://arxiv.org/abs/2608.29966","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29966"},{"ref":"SRC-0038","id":"arxiv-2307-16513","kind":"paper","title":"Deception Abilities Emerged in Large Language Models","year":2023,"arxiv_id":"2307.16513","external_url":"https://arxiv.org/abs/2307.16513","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2307-16513"},{"ref":"SRC-0528","id":"arxiv-2609-24369","kind":"paper","title":"DeceptionAnalyser: A Web-Based AI Tool for Performing Structured Deception Analysis with Argumentation Schemes and LLMs","year":2026,"arxiv_id":"2609.24369","external_url":"https://arxiv.org/abs/2609.24369","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24369"},{"ref":"SRC-0346","id":"arxiv-2602-04853","kind":"paper","title":"Decomposed Prompting Does Not Fix Knowledge Gaps, But Helps Models Say \"I Don't Know\"","year":2026,"arxiv_id":"2602.04853","external_url":"https://arxiv.org/abs/2602.04853","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2602-04853"},{"ref":"SRC-0240","id":"arxiv-2608-26788","kind":"paper","title":"Decoupling Planning and Control for Instructable Agents","year":2026,"arxiv_id":"2608.26788","external_url":"https://arxiv.org/abs/2608.26788","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26788"},{"ref":"SRC-0327","id":"arxiv-2510-11059","kind":"paper","title":"Defects4C: Benchmarking Large Language Model Repair Capability with C/C++ Bugs","year":2025,"arxiv_id":"2510.11059","external_url":"https://arxiv.org/abs/2510.11059","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2510-11059"},{"ref":"SRC-0668","id":"arxiv-2608-30405","kind":"paper","title":"Dense Clinical Contrasts Enhance Medical Knowledge Updating in Large Language Models","year":2026,"arxiv_id":"2608.30405","external_url":"https://arxiv.org/abs/2608.30405","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30405"},{"ref":"SRC-0105","id":"arxiv-2608-21057","kind":"paper","title":"Designing a Robust LLM-Based Evaluation System for Agentic AI in Drug Discovery Through Human Alignment","year":2026,"arxiv_id":"2608.21057","external_url":"https://arxiv.org/abs/2608.21057","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-21057"},{"ref":"SRC-0145","id":"arxiv-2608-29307","kind":"paper","title":"Detecting and Repairing Hallucinations in Retrieval-Augmented Generation","year":2026,"arxiv_id":"2608.29307","external_url":"https://arxiv.org/abs/2608.29307","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29307"},{"ref":"SRC-0543","id":"arxiv-2609-02028","kind":"paper","title":"Detecting Object Hallucinations in Large Vision-Language Models via Cross-Modal Attention Drifts and Mask-Based Verification","year":2026,"arxiv_id":"2609.02028","external_url":"https://arxiv.org/abs/2609.02028","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02028"},{"ref":"SRC-0649","id":"arxiv-2608-27026","kind":"paper","title":"Direct or Mediated? Task-Dependent Audio Information Routing in Large Audio Language Models","year":2026,"arxiv_id":"2608.27026","external_url":"https://arxiv.org/abs/2608.27026","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27026"},{"ref":"SRC-0621","id":"arxiv-2608-12921","kind":"paper","title":"Discovering Efficient and Explainable Communication Topologies for LLM-based Multi-Agent Systems via Causal Inference","year":2026,"arxiv_id":"2608.12921","external_url":"https://arxiv.org/abs/2608.12921","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-12921"},{"ref":"SRC-0516","id":"arxiv-2609-22851","kind":"paper","title":"Discrete vs. Continuous: A Comprehensive Study of Unified Audio Understanding in LALMs","year":2026,"arxiv_id":"2609.22851","external_url":"https://arxiv.org/abs/2609.22851","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22851"},{"ref":"SRC-0607","id":"arxiv-2608-08878","kind":"paper","title":"DistillCache: KL-Guided Adaptive KV-Cache Eviction for Memory-Efficient LLM Inference","year":2026,"arxiv_id":"2608.08878","external_url":"https://arxiv.org/abs/2608.08878","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-08878"},{"ref":"SRC-0364","id":"arxiv-2607-23513","kind":"paper","title":"Do Diagrams Help Large Language Models Reason? Evidence from Syllogistic Reasoning","year":2026,"arxiv_id":"2607.23513","external_url":"https://arxiv.org/abs/2607.23513","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2607-23513"},{"ref":"SRC-0685","id":"arxiv-2609-07879","kind":"paper","title":"Do Large Language Models Know What They Don't Know II? A Fully Behavioral, Non-Cognitive Measure of Epistemic Honesty","year":2026,"arxiv_id":"2609.07879","external_url":"https://arxiv.org/abs/2609.07879","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07879"},{"ref":"SRC-0307","id":"arxiv-2305-18153","kind":"paper","title":"Do Large Language Models Know What They Don't Know?","year":2023,"arxiv_id":"2305.18153","external_url":"https://arxiv.org/abs/2305.18153","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2305-18153"},{"ref":"SRC-0094","id":"arxiv-2608-17950","kind":"paper","title":"Do Large Language Models Play Six Degrees of Separation? Measuring Topological Compression in Long-Context Manifolds","year":2026,"arxiv_id":"2608.17950","external_url":"https://arxiv.org/abs/2608.17950","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-17950"},{"ref":"SRC-0619","id":"arxiv-2608-12547","kind":"paper","title":"Do LLMs Beat Nash? Testing Decentralized Coordination in Self-Play Multi-Agent Games","year":2026,"arxiv_id":"2608.12547","external_url":"https://arxiv.org/abs/2608.12547","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-12547"},{"ref":"SRC-0331","id":"arxiv-2511-00689","kind":"paper","title":"Do Methods to Jailbreak and Defend LLMs Generalize Across Languages?","year":2025,"arxiv_id":"2511.00689","external_url":"https://arxiv.org/abs/2511.00689","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2511-00689"},{"ref":"SRC-0340","id":"arxiv-2512-16698","kind":"paper","title":"Do Multi-Agents Solve Better Than Single? Evaluating Agentic Frameworks for Diagram-Grounded Geometry Problem Solving and Reasoning","year":2025,"arxiv_id":"2512.16698","external_url":"https://arxiv.org/abs/2512.16698","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2512-16698"},{"ref":"SRC-0164","id":"arxiv-2609-00067","kind":"paper","title":"Do Multimodal LLMs See Before They Read? Diagnosing Contextual Sycophancy","year":2026,"arxiv_id":"2609.00067","external_url":"https://arxiv.org/abs/2609.00067","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00067"},{"ref":"SRC-0632","id":"arxiv-2608-24369","kind":"paper","title":"Do Recipes Have Personas? Characterizing and Generating Creator Style in Attributed Procedural Graphs","year":2026,"arxiv_id":"2608.24369","external_url":"https://arxiv.org/abs/2608.24369","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24369"},{"ref":"SRC-0016","id":"arxiv-2211-03622","kind":"paper","title":"Do Users Write More Insecure Code with AI Assistants?","year":2022,"arxiv_id":"2211.03622","external_url":"https://arxiv.org/abs/2211.03622","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2211-03622"},{"ref":"SRC-0545","id":"arxiv-2609-02059","kind":"paper","title":"DocHop: Benchmarking Out-of-domain Multi-hop Reasoning in Information-Dense Documents","year":2026,"arxiv_id":"2609.02059","external_url":"https://arxiv.org/abs/2609.02059","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02059"},{"ref":"SRC-0176","id":"arxiv-2609-01888","kind":"paper","title":"Does Playing it Safe Count as Faithfulness? Reassessing LVLM Hallucination Mitigation Methods","year":2026,"arxiv_id":"2609.01888","external_url":"https://arxiv.org/abs/2609.01888","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01888"},{"ref":"SRC-0285","id":"arxiv-2609-11878","kind":"paper","title":"Domain-Specific Hallucination Detection in Large Language Models","year":2026,"arxiv_id":"2609.11878","external_url":"https://arxiv.org/abs/2609.11878","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-11878"},{"ref":"SRC-0394","id":"arxiv-2609-13556","kind":"paper","title":"Domain-Specific Jargon in Large Language Models: A Comparative Analysis between General-Purpose and Specialist Models","year":2026,"arxiv_id":"2609.13556","external_url":"https://arxiv.org/abs/2609.13556","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-13556"},{"ref":"SRC-0224","id":"arxiv-2609-08850","kind":"paper","title":"DSE-VTG: Dual-Side Enhancement for Training-Free Video Temporal Grounding","year":2026,"arxiv_id":"2609.08850","external_url":"https://arxiv.org/abs/2609.08850","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-08850"},{"ref":"SRC-0155","id":"arxiv-2608-30248","kind":"paper","title":"DSEffi-Bench: Demystifying Large Language Models' Capability in Efficient Data Science Code Generation","year":2026,"arxiv_id":"2608.30248","external_url":"https://arxiv.org/abs/2608.30248","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30248"},{"ref":"SRC-0046","id":"arxiv-2310-03714","kind":"paper","title":"DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines","year":2023,"arxiv_id":"2310.03714","external_url":"https://arxiv.org/abs/2310.03714","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2310-03714"},{"ref":"SRC-0137","id":"arxiv-2608-28058","kind":"paper","title":"Dynamic Alignment Compensation for Hallucination Mitigation in Large Vision-Language Models","year":2026,"arxiv_id":"2608.28058","external_url":"https://arxiv.org/abs/2608.28058","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-28058"},{"ref":"SRC-0398","id":"arxiv-2609-14637","kind":"paper","title":"DynSTEER: Dynamic Stage-wise Trajectory Evaluation and Execution-time Review for Agents","year":2026,"arxiv_id":"2609.14637","external_url":"https://arxiv.org/abs/2609.14637","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-14637"},{"ref":"SRC-0670","id":"arxiv-2608-30730","kind":"paper","title":"E-Commerce Bench: Evaluating LLM Agents on Long-Horizon Autonomous Business Operation","year":2026,"arxiv_id":"2608.30730","external_url":"https://arxiv.org/abs/2608.30730","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30730"},{"ref":"SRC-0263","id":"arxiv-2609-14302","kind":"paper","title":"E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoning","year":2026,"arxiv_id":"2609.14302","external_url":"https://arxiv.org/abs/2609.14302","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14302"},{"ref":"SRC-0288","id":"arxiv-2609-12533","kind":"paper","title":"Earth-Agent-Pro: Towards Real-World Full-Chain Earth Observation with Agents","year":2026,"arxiv_id":"2609.12533","external_url":"https://arxiv.org/abs/2609.12533","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-12533"},{"ref":"SRC-0283","id":"arxiv-2609-11677","kind":"paper","title":"Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents","year":2026,"arxiv_id":"2609.11677","external_url":"https://arxiv.org/abs/2609.11677","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-11677"},{"ref":"SRC-0156","id":"arxiv-2608-30441","kind":"paper","title":"ECLIPSE: Self-Evolving Stealthy Prompt Injection Attack against Long-Horizon Agentic Systems","year":2026,"arxiv_id":"2608.30441","external_url":"https://arxiv.org/abs/2608.30441","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30441"},{"ref":"SRC-0266","id":"arxiv-2609-14839","kind":"paper","title":"Efficiency Hallucination: Formalizing and Measuring Behavioral Calibration in LLM-Based Code Optimization","year":2026,"arxiv_id":"2609.14839","external_url":"https://arxiv.org/abs/2609.14839","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14839"},{"ref":"SRC-0622","id":"arxiv-2608-13113","kind":"paper","title":"EgoMonth: A Month-Level Egocentric Video Benchmark for Long-Term Spatiotemporal Memory","year":2026,"arxiv_id":"2608.13113","external_url":"https://arxiv.org/abs/2608.13113","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-13113"},{"ref":"SRC-0299","id":"arxiv-2609-16610","kind":"paper","title":"EgoPathBench: Evaluating Zero-Shot Egocentric Waypoint Decision-Making in Vision-Language Models","year":2026,"arxiv_id":"2609.16610","external_url":"https://arxiv.org/abs/2609.16610","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-16610"},{"ref":"SRC-0093","id":"arxiv-2608-17512","kind":"paper","title":"Embodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation","year":2026,"arxiv_id":"2608.17512","external_url":"https://arxiv.org/abs/2608.17512","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-17512"},{"ref":"SRC-0407","id":"arxiv-2609-17320","kind":"paper","title":"Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems","year":2026,"arxiv_id":"2609.17320","external_url":"https://arxiv.org/abs/2609.17320","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-17320"},{"ref":"SRC-0401","id":"arxiv-2609-15242","kind":"paper","title":"Empirical Evaluation of Open-Source Large Language Models for Retrieval-Augmented Generation in ESG Domain","year":2026,"arxiv_id":"2609.15242","external_url":"https://arxiv.org/abs/2609.15242","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15242"},{"ref":"SRC-0292","id":"arxiv-2609-15161","kind":"paper","title":"EMR: Self-Evolving Medical Multi-Agent System via Experience Mining and Reuse","year":2026,"arxiv_id":"2609.15161","external_url":"https://arxiv.org/abs/2609.15161","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15161"},{"ref":"SRC-0519","id":"arxiv-2609-23043","kind":"paper","title":"Enforcing Narrative Reliability and Epistemic Pacing in LLM-Driven Detective Games via Structured Knowledge Trees","year":2026,"arxiv_id":"2609.23043","external_url":"https://arxiv.org/abs/2609.23043","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-23043"},{"ref":"SRC-0680","id":"arxiv-2609-04410","kind":"paper","title":"Engineered Persuasion: Evaluating Personalized Pretexts in LLM-Generated Spear Phishing","year":2026,"arxiv_id":"2609.04410","external_url":"https://arxiv.org/abs/2609.04410","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-04410"},{"ref":"SRC-0400","id":"arxiv-2609-14784","kind":"paper","title":"Enhancing Automated Unit Test Generation for NLP Libraries Using Large Language Models","year":2026,"arxiv_id":"2609.14784","external_url":"https://arxiv.org/abs/2609.14784","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14784"},{"ref":"SRC-0104","id":"arxiv-2608-20814","kind":"paper","title":"Enhancing Localized Reasoning for Long Video Understanding via Efficient Segment-to-Video Supervision","year":2026,"arxiv_id":"2608.20814","external_url":"https://arxiv.org/abs/2608.20814","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-20814"},{"ref":"SRC-0384","id":"arxiv-2609-10316","kind":"paper","title":"Ensembling LLMs for AI-Augmented Cybersecurity Software Requirements Generation","year":2026,"arxiv_id":"2609.10316","external_url":"https://arxiv.org/abs/2609.10316","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10316"},{"ref":"SRC-0026","id":"arxiv-2305-02363","kind":"paper","title":"Entity Tracking in Language Models","year":2023,"arxiv_id":"2305.02363","external_url":"https://arxiv.org/abs/2305.02363","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2305-02363"},{"ref":"SRC-0480","id":"arxiv-2609-08404","kind":"paper","title":"Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks","year":2026,"arxiv_id":"2609.08404","external_url":"https://arxiv.org/abs/2609.08404","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08404"},{"ref":"SRC-0667","id":"arxiv-2608-30021","kind":"paper","title":"Error Detection for PET/CT Radiology Reports: Domain-Specific vs Large Language Models","year":2026,"arxiv_id":"2608.30021","external_url":"https://arxiv.org/abs/2608.30021","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30021"},{"ref":"SRC-0383","id":"arxiv-2609-00738","kind":"paper","title":"Escaping Redundant Reasoning: Structure-Aware Search for Inference-Time LLMs","year":2026,"arxiv_id":"2609.00738","external_url":"https://arxiv.org/abs/2609.00738","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-00738"},{"ref":"SRC-0084","id":"arxiv-2602-11988","kind":"paper","title":"Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?","year":2026,"arxiv_id":"2602.11988","external_url":"https://arxiv.org/abs/2602.11988","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2602-11988"},{"ref":"SRC-0225","id":"arxiv-2609-08943","kind":"paper","title":"Evaluating and Improving Evidence-Grounded Fact-Checking in LLMs via Multi-Round Evidence Ablation","year":2026,"arxiv_id":"2609.08943","external_url":"https://arxiv.org/abs/2609.08943","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08943"},{"ref":"SRC-0647","id":"arxiv-2608-26846","kind":"paper","title":"Evaluating Confidence-Gated Retrieval with Matched Trajectory Replay","year":2026,"arxiv_id":"2608.26846","external_url":"https://arxiv.org/abs/2608.26846","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26846"},{"ref":"SRC-0317","id":"arxiv-2504-15253","kind":"paper","title":"Evaluating Judges as Evaluators: The JETTS Benchmark of LLM-as-Judges as Test-Time Scaling Evaluators","year":2025,"arxiv_id":"2504.15253","external_url":"https://arxiv.org/abs/2504.15253","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2504-15253"},{"ref":"SRC-0202","id":"arxiv-2609-04272","kind":"paper","title":"Evaluating Large Language Models for Forced Outage Risk Prediction: Benefits and Comparison to Machine Learning","year":2026,"arxiv_id":"2609.04272","external_url":"https://arxiv.org/abs/2609.04272","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-04272"},{"ref":"SRC-0006","id":"arxiv-2107-03374","kind":"paper","title":"Evaluating Large Language Models Trained on Code","year":2021,"arxiv_id":"2107.03374","external_url":"https://arxiv.org/abs/2107.03374","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2107-03374"},{"ref":"SRC-0662","id":"arxiv-2608-29543","kind":"paper","title":"Evaluating LLMs on Conversational Text-to-SQL under Chain Ambiguity and Intent Drift","year":2026,"arxiv_id":"2608.29543","external_url":"https://arxiv.org/abs/2608.29543","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29543"},{"ref":"SRC-0470","id":"arxiv-2609-01404","kind":"paper","title":"Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and Searching","year":2026,"arxiv_id":"2609.01404","external_url":"https://arxiv.org/abs/2609.01404","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01404"},{"ref":"SRC-0664","id":"arxiv-2608-29738","kind":"paper","title":"Evaluating the Capabilities of LLMs for Persuasive Dialogue","year":2026,"arxiv_id":"2608.29738","external_url":"https://arxiv.org/abs/2608.29738","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29738"},{"ref":"SRC-0141","id":"arxiv-2608-28833","kind":"paper","title":"Evaluating the Hidden Costs of Personalization in Large Language Models","year":2026,"arxiv_id":"2608.28833","external_url":"https://arxiv.org/abs/2608.28833","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-28833"},{"ref":"SRC-0069","id":"arxiv-2406-03689","kind":"paper","title":"Evaluating the World Model Implicit in a Generative Model","year":2024,"arxiv_id":"2406.03689","external_url":"https://arxiv.org/abs/2406.03689","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2406-03689"},{"ref":"SRC-0148","id":"arxiv-2608-29376","kind":"paper","title":"Evaluating Tiny Recursive Models Across Training for Code Generation","year":2026,"arxiv_id":"2608.29376","external_url":"https://arxiv.org/abs/2608.29376","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29376"},{"ref":"SRC-0260","id":"arxiv-2609-13250","kind":"paper","title":"Evaluation of MLLM-Agnostic Plug-and-Play Keyframe Selection Methods for Long Video Understanding","year":2026,"arxiv_id":"2609.13250","external_url":"https://arxiv.org/abs/2609.13250","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-13250"},{"ref":"SRC-0143","id":"arxiv-2608-29092","kind":"paper","title":"EviAnchor: Mitigating Hallucinations in Large Vision-Language Models via Regional Visual Evidence Compensation","year":2026,"arxiv_id":"2608.29092","external_url":"https://arxiv.org/abs/2608.29092","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29092"},{"ref":"SRC-0671","id":"arxiv-2608-30938","kind":"paper","title":"Evidence, Logic, and Compliance: Multi-Agent Structured Graph Reasoning with Expert Arbitration for Medical Referral","year":2026,"arxiv_id":"2608.30938","external_url":"https://arxiv.org/abs/2608.30938","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30938"},{"ref":"SRC-0633","id":"arxiv-2608-24570","kind":"paper","title":"EviDx: Evidence-Aware Active Diagnosis with Scaffolded LLM Agents","year":2026,"arxiv_id":"2608.24570","external_url":"https://arxiv.org/abs/2608.24570","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24570"},{"ref":"SRC-0563","id":"arxiv-2609-17632","kind":"paper","title":"EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents","year":2026,"arxiv_id":"2609.17632","external_url":"https://arxiv.org/abs/2609.17632","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17632"},{"ref":"SRC-0376","id":"arxiv-2609-15779","kind":"paper","title":"EvoOntology: A Self-Evolving Ontology Layer for Data Agents","year":2026,"arxiv_id":"2609.15779","external_url":"https://arxiv.org/abs/2609.15779","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-15779"},{"ref":"SRC-0211","id":"arxiv-2609-05903","kind":"paper","title":"EvoSafeHarness: Evolving Model- and Domain-Specific Harnesses for Securing Agents","year":2026,"arxiv_id":"2609.05903","external_url":"https://arxiv.org/abs/2609.05903","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05903"},{"ref":"SRC-0113","id":"arxiv-2608-23758","kind":"paper","title":"EXAM$^2$: $\\underline{Ex}tending$ $\\underline{A}udio$ $Understanding$ $in$ $\\underline{M}ultilingual$ $and$ $\\underline{M}ultimodal$ $Analysis$","year":2026,"arxiv_id":"2608.23758","external_url":"https://arxiv.org/abs/2608.23758","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-23758"},{"ref":"SRC-0109","id":"arxiv-2608-22938","kind":"paper","title":"Execution-Anchored Hallucination Calibration Reranking for Verilog Code Generation","year":2026,"arxiv_id":"2608.22938","external_url":"https://arxiv.org/abs/2608.22938","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-22938"},{"ref":"SRC-0278","id":"arxiv-2609-08919","kind":"paper","title":"Experience Funnel: A State-Policy Alternating Loop for Self-Evolving Agents","year":2026,"arxiv_id":"2609.08919","external_url":"https://arxiv.org/abs/2609.08919","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08919"},{"ref":"SRC-0043","id":"arxiv-2309-04658","kind":"paper","title":"Exploring Large Language Models for Communication Games: An Empirical Study on Werewolf","year":2023,"arxiv_id":"2309.04658","external_url":"https://arxiv.org/abs/2309.04658","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2309-04658"},{"ref":"SRC-0265","id":"arxiv-2609-14758","kind":"paper","title":"Fabrication After Tool Failure: Tool-Augmented Agents Assert Values Their Tools Did Not Return","year":2026,"arxiv_id":"2609.14758","external_url":"https://arxiv.org/abs/2609.14758","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14758"},{"ref":"SRC-0028","id":"arxiv-2305-14251","kind":"paper","title":"FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation","year":2023,"arxiv_id":"2305.14251","external_url":"https://arxiv.org/abs/2305.14251","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2305-14251"},{"ref":"SRC-0032","id":"arxiv-2305-18654","kind":"paper","title":"Faith and Fate: Limits of Transformers on Compositionality","year":2023,"arxiv_id":"2305.18654","external_url":"https://arxiv.org/abs/2305.18654","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2305-18654"},{"ref":"SRC-0502","id":"arxiv-2609-21358","kind":"paper","title":"FAN: Foresight Action Normalization for Continual Adaptation of Vision-Language-Action Models","year":2026,"arxiv_id":"2609.21358","external_url":"https://arxiv.org/abs/2609.21358","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21358"},{"ref":"SRC-0117","id":"arxiv-2608-24350","kind":"paper","title":"FARCA: Fact-Aligned Reliability-Aware Credit Assignment for Reinforcement Learning with Factual Supervision","year":2026,"arxiv_id":"2608.24350","external_url":"https://arxiv.org/abs/2608.24350","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24350"},{"ref":"SRC-0448","id":"arxiv-2608-28264","kind":"paper","title":"Finding Where the Buck Stops: An Automated Failure Attribution-Based Reflection Framework for Multi-Agent Collaboration","year":2026,"arxiv_id":"2608.28264","external_url":"https://arxiv.org/abs/2608.28264","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-28264"},{"ref":"SRC-0158","id":"arxiv-2608-30653","kind":"paper","title":"Fine-Grained Multi Image Object Hallucination Benchmark","year":2026,"arxiv_id":"2608.30653","external_url":"https://arxiv.org/abs/2608.30653","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30653"},{"ref":"SRC-0525","id":"arxiv-2609-24175","kind":"paper","title":"FinRankGRPO: Optimizing LLMs for Listwise Financial Asset Ranking via Group Relative Policy Optimization","year":2026,"arxiv_id":"2609.24175","external_url":"https://arxiv.org/abs/2609.24175","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24175"},{"ref":"SRC-0381","id":"arxiv-2609-19680","kind":"paper","title":"FINSKILLOPS: A Self-Evolving Multi-Agent System for SEC Filing QA","year":2026,"arxiv_id":"2609.19680","external_url":"https://arxiv.org/abs/2609.19680","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-19680"},{"ref":"SRC-0047","id":"arxiv-2310-05915","kind":"paper","title":"FireAct: Toward Language Agent Fine-tuning","year":2023,"arxiv_id":"2310.05915","external_url":"https://arxiv.org/abs/2310.05915","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2310-05915"},{"ref":"SRC-0523","id":"arxiv-2609-23808","kind":"paper","title":"FLARE: A Full-Lifecycle Dense Supervision Paradigm for Long-Horizon Coding Agents via Generative Reward Model","year":2026,"arxiv_id":"2609.23808","external_url":"https://arxiv.org/abs/2609.23808","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-23808"},{"ref":"SRC-0256","id":"le-gulwani-2014-flashextract","kind":"paper","title":"FlashExtract: A Framework for Data Extraction by Examples","year":2014,"arxiv_id":null,"external_url":"https://www.microsoft.com/en-us/research/publication/flashextract-framework-data-extraction-examples/","counts":{"claims":1},"url":"https://rsiratchet.com/sources/le-gulwani-2014-flashextract"},{"ref":"SRC-0421","id":"arxiv-2608-18423","kind":"paper","title":"FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents","year":2026,"arxiv_id":"2608.18423","external_url":"https://arxiv.org/abs/2608.18423","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-18423"},{"ref":"SRC-0013","id":"arxiv-2206-15474","kind":"paper","title":"Forecasting Future World Events with Neural Networks","year":2022,"arxiv_id":"2206.15474","external_url":"https://arxiv.org/abs/2206.15474","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2206-15474"},{"ref":"SRC-0437","id":"arxiv-2608-23630","kind":"paper","title":"FPGAgent: An LLM-Assisted Framework for Autonomous HLS Code Generation and Verification in FPGA Environments","year":2026,"arxiv_id":"2608.23630","external_url":"https://arxiv.org/abs/2608.23630","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-23630"},{"ref":"SRC-0348","id":"arxiv-2603-03332","kind":"paper","title":"Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbations","year":2026,"arxiv_id":"2603.03332","external_url":"https://arxiv.org/abs/2603.03332","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2603-03332"},{"ref":"SRC-0454","id":"arxiv-2608-29814","kind":"paper","title":"FRAMEWORKERS: A Dynamic Multi-Agent Framework for AI-Generated Video Production","year":2026,"arxiv_id":"2608.29814","external_url":"https://arxiv.org/abs/2608.29814","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29814"},{"ref":"SRC-0520","id":"arxiv-2609-23065","kind":"paper","title":"From Concept Alignment to Causal Grounding: An Intervention Test of Chain-of-Thought Faithfulness","year":2026,"arxiv_id":"2609.23065","external_url":"https://arxiv.org/abs/2609.23065","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-23065"},{"ref":"SRC-0257","id":"zelina-2024-llm-regex-czech-clinical","kind":"paper","title":"From Examples to Patterns: LLM-Generated Regular Expressions for Entity Extraction in Czech Clinical Texts","year":2024,"arxiv_id":null,"external_url":"https://nlp.fi.muni.cz/raslan/2024/paper6.pdf","counts":{"claims":1},"url":"https://rsiratchet.com/sources/zelina-2024-llm-regex-czech-clinical"},{"ref":"SRC-0636","id":"arxiv-2608-25661","kind":"paper","title":"From General Agents to RCA Experts: A Self-Evolving Harness for Root Cause Analysis","year":2026,"arxiv_id":"2608.25661","external_url":"https://arxiv.org/abs/2608.25661","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25661"},{"ref":"SRC-0239","id":"arxiv-2608-25243","kind":"paper","title":"From Memorization to Absorption: Mixed-Policy RL for Continual Knowledge Injection","year":2026,"arxiv_id":"2608.25243","external_url":"https://arxiv.org/abs/2608.25243","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25243"},{"ref":"SRC-0125","id":"arxiv-2608-25894","kind":"paper","title":"From Passive Response to Proactive Correction: Enhancing LLM Robustness Against Input Fact Perturbations","year":2026,"arxiv_id":"2608.25894","external_url":"https://arxiv.org/abs/2608.25894","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25894"},{"ref":"SRC-0532","id":"arxiv-2609-24631","kind":"paper","title":"From Semantic Decisions to Feasible Trajectories: Self-Evolving LLM-Guided Optimal Control for Narrow-Space Parking","year":2026,"arxiv_id":"2609.24631","external_url":"https://arxiv.org/abs/2609.24631","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24631"},{"ref":"SRC-0628","id":"arxiv-2608-16002","kind":"paper","title":"From Sequence to Structure: Relational Uncertainty Propagation for LLM Agents","year":2026,"arxiv_id":"2608.16002","external_url":"https://arxiv.org/abs/2608.16002","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-16002"},{"ref":"SRC-0366","id":"arxiv-2608-21897","kind":"paper","title":"From Solver Feedback to Faithful Plans: Multi-Role Reinforcement Learning for Symbolic Planning","year":2026,"arxiv_id":"2608.21897","external_url":"https://arxiv.org/abs/2608.21897","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-21897"},{"ref":"SRC-0098","id":"arxiv-2608-18581","kind":"paper","title":"From Storage to Access: Verifiable Activation of Parametric Knowledge in LLMs via Explicit Priming and Implicit Reasoning","year":2026,"arxiv_id":"2608.18581","external_url":"https://arxiv.org/abs/2608.18581","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-18581"},{"ref":"SRC-0179","id":"arxiv-2609-02679","kind":"paper","title":"From Tokens to Semantics: Leveraging Complementary Signals for Hallucination Detection in Black-Box LLMs","year":2026,"arxiv_id":"2609.02679","external_url":"https://arxiv.org/abs/2609.02679","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02679"},{"ref":"SRC-0076","id":"arxiv-2412-04984","kind":"paper","title":"Frontier Models are Capable of In-context Scheming","year":2024,"arxiv_id":"2412.04984","external_url":"https://arxiv.org/abs/2412.04984","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2412-04984"},{"ref":"SRC-0434","id":"arxiv-2608-22724","kind":"paper","title":"Frontiers in FinTech: Multimodal Foundation Models for Financial Reporting and Decision Science","year":2026,"arxiv_id":"2608.22724","external_url":"https://arxiv.org/abs/2608.22724","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-22724"},{"ref":"SRC-0508","id":"arxiv-2609-21948","kind":"paper","title":"GALA: Geometry-Aware Latent Action Modeling for Vision-Language-Action Model Pretraining across Embodiments","year":2026,"arxiv_id":"2609.21948","external_url":"https://arxiv.org/abs/2609.21948","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21948"},{"ref":"SRC-0497","id":"arxiv-2609-19315","kind":"paper","title":"GAVEL: Graph World Models for Verified and Efficient Long-Horizon LLM Task Planning","year":2026,"arxiv_id":"2609.19315","external_url":"https://arxiv.org/abs/2609.19315","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-19315"},{"ref":"SRC-0025","id":"arxiv-2304-03442","kind":"paper","title":"Generative Agents: Interactive Simulacra of Human Behavior","year":2023,"arxiv_id":"2304.03442","external_url":"https://arxiv.org/abs/2304.03442","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2304-03442"},{"ref":"SRC-0080","id":"arxiv-2507-19457","kind":"paper","title":"GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning","year":2025,"arxiv_id":"2507.19457","external_url":"https://arxiv.org/abs/2507.19457","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2507-19457"},{"ref":"SRC-0587","id":"arxiv-2608-03071","kind":"paper","title":"Getting the Parameters Right: A Difficulty-Graded Benchmark and Probe-Guided Training for LLM Tool Calls","year":2026,"arxiv_id":"2608.03071","external_url":"https://arxiv.org/abs/2608.03071","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-03071"},{"ref":"SRC-0672","id":"arxiv-2609-00103","kind":"paper","title":"Good Memory Has ECC: Evaluating the Memory of Vision-Language Models Beyond Accuracy","year":2026,"arxiv_id":"2609.00103","external_url":"https://arxiv.org/abs/2609.00103","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00103"},{"ref":"SRC-0029","id":"arxiv-2305-15334","kind":"paper","title":"Gorilla: Large Language Model Connected with Massive APIs","year":2023,"arxiv_id":"2305.15334","external_url":"https://arxiv.org/abs/2305.15334","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2305-15334"},{"ref":"SRC-0354","id":"arxiv-2605-01203","kind":"paper","title":"GR-Ben: A General Reasoning Benchmark for Evaluating Process Reward Models","year":2026,"arxiv_id":"2605.01203","external_url":"https://arxiv.org/abs/2605.01203","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2605-01203"},{"ref":"SRC-0205","id":"arxiv-2609-04442","kind":"paper","title":"GRACE: Graph-Grounded Reflective Agent Copilot Engine for Expert-in-the-Loop Knowledge Expansion","year":2026,"arxiv_id":"2609.04442","external_url":"https://arxiv.org/abs/2609.04442","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-04442"},{"ref":"SRC-0585","id":"arxiv-2610-17062","kind":"paper","title":"Gradient Routing Localizes and Removes Unintended Behaviors in RL","year":2026,"arxiv_id":"2610.17062","external_url":"https://arxiv.org/abs/2610.17062","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2610-17062"},{"ref":"SRC-0613","id":"arxiv-2608-11394","kind":"paper","title":"GraphAlignCoder: Aligning Program and Proof Graphs for Code Generation","year":2026,"arxiv_id":"2608.11394","external_url":"https://arxiv.org/abs/2608.11394","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-11394"},{"ref":"SRC-0635","id":"arxiv-2608-25583","kind":"paper","title":"GRIP: Granular Reward-Guided Parameter Interpolation for Efficient Reasoning","year":2026,"arxiv_id":"2608.25583","external_url":"https://arxiv.org/abs/2608.25583","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25583"},{"ref":"SRC-0373","id":"siegel-2026-validation-market","kind":"post","title":"Grokipedia is dead. I'd like to rescue it, starting with scientific data","year":2026,"arxiv_id":null,"external_url":"https://x.com/pullnews/status/2101754686306955275","counts":{"claims":1},"url":"https://rsiratchet.com/sources/siegel-2026-validation-market"},{"ref":"SRC-0484","id":"arxiv-2609-09898","kind":"paper","title":"Grounded Evaluation and Repair for NL-to-PDDL Problem Generation","year":2026,"arxiv_id":"2609.09898","external_url":"https://arxiv.org/abs/2609.09898","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09898"},{"ref":"SRC-0291","id":"arxiv-2609-14455","kind":"paper","title":"Grounded in Sound: Reinforcement Learning with a Frozen Acoustic Judge to Curb ASR Insertion Hallucinations","year":2026,"arxiv_id":"2609.14455","external_url":"https://arxiv.org/abs/2609.14455","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14455"},{"ref":"SRC-0236","id":"arxiv-2609-10046","kind":"paper","title":"Guaranteeing Faithful Evidence Extraction in Speculative Retrieval-Augmented Generation","year":2026,"arxiv_id":"2609.10046","external_url":"https://arxiv.org/abs/2609.10046","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10046"},{"ref":"SRC-0554","id":"arxiv-2609-05284","kind":"paper","title":"GUT: Quantifying and Optimizing the Reasoning Uncertainty of LLMs via Graph Complexity","year":2026,"arxiv_id":"2609.05284","external_url":"https://arxiv.org/abs/2609.05284","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05284"},{"ref":"SRC-0194","id":"hal-gaia-leaderboard-2026","kind":"post","title":"HAL GAIA Leaderboard (Holistic Agent Leaderboard, Princeton)","year":2026,"arxiv_id":null,"external_url":"https://hal.cs.princeton.edu/gaia","counts":{"claims":1},"url":"https://rsiratchet.com/sources/hal-gaia-leaderboard-2026"},{"ref":"SRC-0091","id":"arxiv-2606-21666","kind":"paper","title":"Hallucination as Context Drift: Synchronization Protocols for Multi-Agent LLM Systems","year":2026,"arxiv_id":"2606.21666","external_url":"https://arxiv.org/abs/2606.21666","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-21666"},{"ref":"SRC-0150","id":"arxiv-2608-29924","kind":"paper","title":"Hallucination Mitigation for Large Vision-Language Models via Implicit Feature Stabilization","year":2026,"arxiv_id":"2608.29924","external_url":"https://arxiv.org/abs/2608.29924","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29924"},{"ref":"SRC-0184","id":"arxiv-2609-03580","kind":"paper","title":"HalluPeer: A Taxonomy-driven Benchmark for Detecting Hallucinations in Scientific Peer Reviews","year":2026,"arxiv_id":"2609.03580","external_url":"https://arxiv.org/abs/2609.03580","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-03580"},{"ref":"SRC-0660","id":"arxiv-2608-29193","kind":"paper","title":"HalluPrism: When Multimodal Uncertainty Should Diagnose, Not Decide","year":2026,"arxiv_id":"2608.29193","external_url":"https://arxiv.org/abs/2608.29193","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29193"},{"ref":"SRC-0328","id":"arxiv-2510-13744","kind":"paper","title":"Hard2Verify: A Step-Level Verification Benchmark for Open-Ended Frontier Math","year":2025,"arxiv_id":"2510.13744","external_url":"https://arxiv.org/abs/2510.13744","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2510-13744"},{"ref":"SRC-0200","id":"synthesis-2026-harness-engineering-playbook","kind":"post","title":"Harness Engineering — Agent = Model + Harness: The 6-Layer Production Playbook","year":2026,"arxiv_id":null,"external_url":"https://drive.google.com/file/d/1wZeuP0t4WOA2COMxcrU-Uv34IutM13UR/view","counts":{"claims":6},"url":"https://rsiratchet.com/sources/synthesis-2026-harness-engineering-playbook"},{"ref":"SRC-0189","id":"boeckeler-2026-harness-engineering-fowler","kind":"post","title":"Harness engineering for coding agent users","year":2026,"arxiv_id":null,"external_url":"https://martinfowler.com/articles/harness-engineering.html","counts":{"claims":1},"url":"https://rsiratchet.com/sources/boeckeler-2026-harness-engineering-fowler"},{"ref":"SRC-0175","id":"arxiv-2609-01736","kind":"paper","title":"Harness Engineering in LLM Tool Use via Agent-Native Reusable Tool Primitives","year":2026,"arxiv_id":"2609.01736","external_url":"https://arxiv.org/abs/2609.01736","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01736"},{"ref":"SRC-0198","id":"openai-2026-harness-engineering-codex","kind":"vendor-doc","title":"Harness engineering: leveraging Codex in an agent-first world","year":2026,"arxiv_id":null,"external_url":"https://openai.com/index/harness-engineering/","counts":{"claims":0},"url":"https://rsiratchet.com/sources/openai-2026-harness-engineering-codex"},{"ref":"SRC-0089","id":"arxiv-2605-27922","kind":"paper","title":"Harness-Bench: Measuring Harness Effects across Models in Realistic Agent Workflows","year":2026,"arxiv_id":"2605.27922","external_url":"https://arxiv.org/abs/2605.27922","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2605-27922"},{"ref":"SRC-0663","id":"arxiv-2608-29641","kind":"paper","title":"Harness-RL: Black-Box Reinforcement Learning with Action-Args Decoupling for Central-Agent Multi-Agent Harnesses","year":2026,"arxiv_id":"2608.29641","external_url":"https://arxiv.org/abs/2608.29641","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29641"},{"ref":"SRC-0293","id":"arxiv-2609-15195","kind":"paper","title":"HarnessVLN: Unifying Training-Free Embodied Navigation through an Agent Harness","year":2026,"arxiv_id":"2609.15195","external_url":"https://arxiv.org/abs/2609.15195","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15195"},{"ref":"SRC-0201","id":"veremeyenko-2026-hashimotos-rule","kind":"post","title":"Hashimoto's rule for AI agents is that every mistake becomes a permanent fix","year":2026,"arxiv_id":null,"external_url":"https://x.com/alex_verem/status/2097792667795276241","counts":{"claims":1},"url":"https://rsiratchet.com/sources/veremeyenko-2026-hashimotos-rule"},{"ref":"SRC-0466","id":"arxiv-2609-00727","kind":"paper","title":"Heard but Not Heeded: Paralinguistic Information Encoding and Loss in Audio-Language Models","year":2026,"arxiv_id":"2609.00727","external_url":"https://arxiv.org/abs/2609.00727","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00727"},{"ref":"SRC-0355","id":"arxiv-2605-02215","kind":"paper","title":"HEJ-Robust: A Robustness Benchmark for LLM-Based Automated Program Repair","year":2026,"arxiv_id":"2605.02215","external_url":"https://arxiv.org/abs/2605.02215","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2605-02215"},{"ref":"SRC-0629","id":"arxiv-2608-24042","kind":"paper","title":"Hierarchical Skill Retrieval for Data-Efficient Adaptation of Vision-Language-Action Models","year":2026,"arxiv_id":"2608.24042","external_url":"https://arxiv.org/abs/2608.24042","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24042"},{"ref":"SRC-0499","id":"arxiv-2609-20659","kind":"paper","title":"HIL-UMI: Bringing Human-in-the-Loop Post-Training of Vision-Language-Action Models to Universal Manipulation Interface","year":2026,"arxiv_id":"2609.20659","external_url":"https://arxiv.org/abs/2609.20659","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-20659"},{"ref":"SRC-0409","id":"arxiv-2609-17771","kind":"paper","title":"HINT-Plan: Human Intention-Aware Robot Task Planning in Context-Rich Environments using Vision Language Models","year":2026,"arxiv_id":"2609.17771","external_url":"https://arxiv.org/abs/2609.17771","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17771"},{"ref":"SRC-0459","id":"arxiv-2608-30672","kind":"paper","title":"HiRS-Agent: A Hierarchical Multi-Agent System for Reliable Long-Horizon Remote Sensing Task Solving","year":2026,"arxiv_id":"2608.30672","external_url":"https://arxiv.org/abs/2608.30672","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30672"},{"ref":"SRC-0677","id":"arxiv-2609-01660","kind":"paper","title":"How Fast Do Agents Rot? An Empirical Study of Long-Horizon Degradation in LLM Agents for Production Decision-Making","year":2026,"arxiv_id":"2609.01660","external_url":"https://arxiv.org/abs/2609.01660","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01660"},{"ref":"SRC-0144","id":"arxiv-2608-29198","kind":"paper","title":"How Identity and Opinion Shape Political Sycophancy in LLMs","year":2026,"arxiv_id":"2608.29198","external_url":"https://arxiv.org/abs/2608.29198","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-29198"},{"ref":"SRC-0325","id":"arxiv-2510-06700","kind":"paper","title":"How Language Models Conflate Logical Validity with Plausibility: A Representational Analysis of Content Effects","year":2025,"arxiv_id":"2510.06700","external_url":"https://arxiv.org/abs/2510.06700","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2510-06700"},{"ref":"SRC-0461","id":"arxiv-2608-30841","kind":"paper","title":"HSRM: Hidden-State Reward Models for Test-Time Verification","year":2026,"arxiv_id":"2608.30841","external_url":"https://arxiv.org/abs/2608.30841","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30841"},{"ref":"SRC-0592","id":"arxiv-2608-04351","kind":"paper","title":"HyPASE: Hyperbolic Geometry for Parameter-Efficient Speech Emotion Fine-Tuning Framework for Large Audio-Language Models","year":2026,"arxiv_id":"2608.04351","external_url":"https://arxiv.org/abs/2608.04351","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-04351"},{"ref":"SRC-0100","id":"arxiv-2608-19347","kind":"paper","title":"Hype Meets Reality: Large Language Models as Mutators in Search-based Automated Program Repair of Simulink-Stateflow Models","year":2026,"arxiv_id":"2608.19347","external_url":"https://arxiv.org/abs/2608.19347","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-19347"},{"ref":"SRC-0491","id":"arxiv-2609-15938","kind":"paper","title":"HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses","year":2026,"arxiv_id":"2609.15938","external_url":"https://arxiv.org/abs/2609.15938","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15938"},{"ref":"SRC-0275","id":"arxiv-2608-25770","kind":"paper","title":"HypoForge: A Self-Improving Multi-Agent Framework for Automated Hypothesis Generation and Testing via Scientific Skill Learning","year":2026,"arxiv_id":"2608.25770","external_url":"https://arxiv.org/abs/2608.25770","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25770"},{"ref":"SRC-0191","id":"boluk-2026-harness-problem-15-llms","kind":"post","title":"I Improved 15 LLMs at Coding in One Afternoon. Only the Harness Changed.","year":2026,"arxiv_id":null,"external_url":"https://can.ac/2026/02/12/the-harness-problem/","counts":{"claims":2},"url":"https://rsiratchet.com/sources/boluk-2026-harness-problem-15-llms"},{"ref":"SRC-0018","id":"arxiv-2211-09527","kind":"paper","title":"Ignore Previous Prompt: Attack Techniques For Language Models","year":2022,"arxiv_id":"2211.09527","external_url":"https://arxiv.org/abs/2211.09527","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2211-09527"},{"ref":"SRC-0012","id":"arxiv-2202-07206","kind":"paper","title":"Impact of Pretraining Term Frequencies on Few-Shot Reasoning","year":2022,"arxiv_id":"2202.07206","external_url":"https://arxiv.org/abs/2202.07206","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2202-07206"},{"ref":"SRC-0375","id":"dair-2026-modularrsi-three-defects","kind":"post","title":"Impressive paper on building recursive self-improving agent harnesses","year":2026,"arxiv_id":null,"external_url":"https://x.com/dair_ai/status/2101835306953736396","counts":{"claims":1},"url":"https://rsiratchet.com/sources/dair-2026-modularrsi-three-defects"},{"ref":"SRC-0196","id":"langchain-2026-deep-agents-harness","kind":"vendor-doc","title":"Improving Deep Agents with harness engineering","year":2026,"arxiv_id":null,"external_url":"https://www.langchain.com/blog/improving-deep-agents-with-harness-engineering","counts":{"claims":1},"url":"https://rsiratchet.com/sources/langchain-2026-deep-agents-harness"},{"ref":"SRC-0178","id":"arxiv-2609-02396","kind":"paper","title":"Improving Health Literacy through Lay Summarization of Radiological Reports: An Evaluation of BioNER and Retrieval-Augmented Generation","year":2026,"arxiv_id":"2609.02396","external_url":"https://arxiv.org/abs/2609.02396","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02396"},{"ref":"SRC-0216","id":"arxiv-2609-07160","kind":"paper","title":"In-Place Instruction Following in Diffusion Language Models","year":2026,"arxiv_id":"2609.07160","external_url":"https://arxiv.org/abs/2609.07160","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07160"},{"ref":"SRC-0391","id":"arxiv-2609-12495","kind":"paper","title":"Information Specialization and Constrained Synthesis in Multi-Agent LLM Forecasting: A Prospective Live-Study of the 2026 FIFA World Cup","year":2026,"arxiv_id":"2609.12495","external_url":"https://arxiv.org/abs/2609.12495","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-12495"},{"ref":"SRC-0174","id":"arxiv-2609-01383","kind":"paper","title":"InSight: A Benchmark for Agentic Claim Verification in Interactive Visualizations","year":2026,"arxiv_id":"2609.01383","external_url":"https://arxiv.org/abs/2609.01383","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01383"},{"ref":"SRC-0053","id":"arxiv-2311-07911","kind":"paper","title":"Instruction-Following Evaluation for Large Language Models","year":2023,"arxiv_id":"2311.07911","external_url":"https://arxiv.org/abs/2311.07911","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2311-07911"},{"ref":"SRC-0553","id":"arxiv-2609-04483","kind":"paper","title":"Integrating Crash Report Mining and LLMs for Bug Localization and Repair: An Industrial Report","year":2026,"arxiv_id":"2609.04483","external_url":"https://arxiv.org/abs/2609.04483","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-04483"},{"ref":"SRC-0035","id":"arxiv-2306-14898","kind":"paper","title":"InterCode: Standardizing and Benchmarking Interactive Coding with Execution Feedback","year":2023,"arxiv_id":"2306.14898","external_url":"https://arxiv.org/abs/2306.14898","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2306-14898"},{"ref":"SRC-0395","id":"arxiv-2609-13975","kind":"paper","title":"Introspective Uncertainty Estimation for LLM-Based Code Generation","year":2026,"arxiv_id":"2609.13975","external_url":"https://arxiv.org/abs/2609.13975","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-13975"},{"ref":"SRC-0314","id":"arxiv-2503-22424","kind":"paper","title":"Issue Localization via LLM-Driven Iterative Code Graph Searching","year":2025,"arxiv_id":"2503.22424","external_url":"https://arxiv.org/abs/2503.22424","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2503-22424"},{"ref":"SRC-0034","id":"arxiv-2306-05685","kind":"paper","title":"Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena","year":2023,"arxiv_id":"2306.05685","external_url":"https://arxiv.org/abs/2306.05685","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2306-05685"},{"ref":"SRC-0306","id":"arxiv-2305-14975","kind":"paper","title":"Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback","year":2023,"arxiv_id":"2305.14975","external_url":"https://arxiv.org/abs/2305.14975","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2305-14975"},{"ref":"SRC-0185","id":"arxiv-2609-03588","kind":"paper","title":"KC-Bench: A Dynamic Interactive Benchmark for Evaluating Knowledge Conflicts in LLM Agents","year":2026,"arxiv_id":"2609.03588","external_url":"https://arxiv.org/abs/2609.03588","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-03588"},{"ref":"SRC-0234","id":"arxiv-2609-09865","kind":"paper","title":"Keep Evaluation Fair: Detecting Data Leakage in Code Generation Benchmarks via Membership Inference Attacks","year":2026,"arxiv_id":"2609.09865","external_url":"https://arxiv.org/abs/2609.09865","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09865"},{"ref":"SRC-0329","id":"arxiv-2510-16815","kind":"paper","title":"Knowing the Facts but Choosing the Shortcut: Understanding How Large Language Models Compare Entities","year":2025,"arxiv_id":"2510.16815","external_url":"https://arxiv.org/abs/2510.16815","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2510-16815"},{"ref":"SRC-0345","id":"arxiv-2602-03900","kind":"paper","title":"Knowledge Model Prompting Increases LLM Performance on Planning Tasks","year":2026,"arxiv_id":"2602.03900","external_url":"https://arxiv.org/abs/2602.03900","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2602-03900"},{"ref":"SRC-0130","id":"arxiv-2608-26372","kind":"paper","title":"Knowledge-Verified Emergent Deception in LLM Agents Under Conflicting Incentives","year":2026,"arxiv_id":"2608.26372","external_url":"https://arxiv.org/abs/2608.26372","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-26372"},{"ref":"SRC-0305","id":"arxiv-2207-05221","kind":"paper","title":"Language Models (Mostly) Know What They Know","year":2022,"arxiv_id":"2207.05221","external_url":"https://arxiv.org/abs/2207.05221","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2207-05221"},{"ref":"SRC-0005","id":"arxiv-2005-14165","kind":"paper","title":"Language Models are Few-Shot Learners","year":2020,"arxiv_id":"2005.14165","external_url":"https://arxiv.org/abs/2005.14165","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2005-14165"},{"ref":"SRC-0031","id":"arxiv-2305-17926","kind":"paper","title":"Large Language Models are not Fair Evaluators","year":2023,"arxiv_id":"2305.17926","external_url":"https://arxiv.org/abs/2305.17926","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2305-17926"},{"ref":"SRC-0342","id":"arxiv-2601-05050","kind":"paper","title":"Large language models can effectively convince people to believe conspiracies","year":2026,"arxiv_id":"2601.05050","external_url":"https://arxiv.org/abs/2601.05050","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2601-05050"},{"ref":"SRC-0618","id":"arxiv-2608-12426","kind":"paper","title":"Large Language Models Can Follow Instructions, But Not Many at Once: Phase Transitions in Compositional Constraint Satisfaction","year":2026,"arxiv_id":"2608.12426","external_url":"https://arxiv.org/abs/2608.12426","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-12426"},{"ref":"SRC-0045","id":"arxiv-2310-01798","kind":"paper","title":"Large Language Models Cannot Self-Correct Reasoning Yet","year":2023,"arxiv_id":"2310.01798","external_url":"https://arxiv.org/abs/2310.01798","counts":{"claims":5},"url":"https://rsiratchet.com/sources/arxiv-2310-01798"},{"ref":"SRC-0181","id":"arxiv-2609-03148","kind":"paper","title":"Large Language Models in Resolving Contextual Knowledge Conflicts","year":2026,"arxiv_id":"2609.03148","external_url":"https://arxiv.org/abs/2609.03148","counts":{"claims":3},"url":"https://rsiratchet.com/sources/arxiv-2609-03148"},{"ref":"SRC-0017","id":"arxiv-2211-08411","kind":"paper","title":"Large Language Models Struggle to Learn Long-Tail Knowledge","year":2022,"arxiv_id":"2211.08411","external_url":"https://arxiv.org/abs/2211.08411","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2211-08411"},{"ref":"SRC-0460","id":"arxiv-2608-30679","kind":"paper","title":"LCoT-GV: Graph Attention Networks for Verifying Long Reasoning Chains in Large Language Models","year":2026,"arxiv_id":"2608.30679","external_url":"https://arxiv.org/abs/2608.30679","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30679"},{"ref":"SRC-0449","id":"arxiv-2608-29305","kind":"paper","title":"Learning Simple Test-Time Environments for LLM Web Agents","year":2026,"arxiv_id":"2608.29305","external_url":"https://arxiv.org/abs/2608.29305","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29305"},{"ref":"SRC-0139","id":"arxiv-2608-28447","kind":"paper","title":"Learning to Use Tools: Reinforcement Learning for Tool-Integrated Mathematical Reasoning","year":2026,"arxiv_id":"2608.28447","external_url":"https://arxiv.org/abs/2608.28447","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-28447"},{"ref":"SRC-0464","id":"arxiv-2609-00237","kind":"paper","title":"Learning What to Retain: Gated-Memory Routing for Efficient Collaboration in Multi-Agent LLM Systems","year":2026,"arxiv_id":"2609.00237","external_url":"https://arxiv.org/abs/2609.00237","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00237"},{"ref":"SRC-0221","id":"arxiv-2609-08180","kind":"paper","title":"Less Is Personal: Learning Minimal Sufficient User Profiles for Personalized Language Models","year":2026,"arxiv_id":"2609.08180","external_url":"https://arxiv.org/abs/2609.08180","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08180"},{"ref":"SRC-0207","id":"arxiv-2609-05025","kind":"paper","title":"Leveraging Low-Level Symbolic Competences for Unsupervised Grounding in Hallucination Detection","year":2026,"arxiv_id":"2609.05025","external_url":"https://arxiv.org/abs/2609.05025","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-05025"},{"ref":"SRC-0233","id":"arxiv-2609-09754","kind":"paper","title":"LexAgentHallu: A Hierarchical Benchmark for Profiling Hallucinations in Legal Agents","year":2026,"arxiv_id":"2609.09754","external_url":"https://arxiv.org/abs/2609.09754","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09754"},{"ref":"SRC-0208","id":"arxiv-2609-05178","kind":"paper","title":"LIBERO-RECOVER: Beyond Task Success Towards Failure Recovery in Robotic Manipulation Models","year":2026,"arxiv_id":"2609.05178","external_url":"https://arxiv.org/abs/2609.05178","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-05178"},{"ref":"SRC-0462","id":"arxiv-2608-30935","kind":"paper","title":"LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation","year":2026,"arxiv_id":"2608.30935","external_url":"https://arxiv.org/abs/2608.30935","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30935"},{"ref":"SRC-0536","id":"arxiv-2503-08508","kind":"paper","title":"LightPlanner: Unleashing the Reasoning Capabilities of Lightweight Large Language Models in Task Planning","year":2025,"arxiv_id":"2503.08508","external_url":"https://arxiv.org/abs/2503.08508","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2503-08508"},{"ref":"SRC-0160","id":"arxiv-2608-30776","kind":"paper","title":"Likelihood-Constrained Acoustic Reranking for Training-Free Hallucination Mitigation in LLM-Based ASR","year":2026,"arxiv_id":"2608.30776","external_url":"https://arxiv.org/abs/2608.30776","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30776"},{"ref":"SRC-0526","id":"arxiv-2609-24186","kind":"paper","title":"LIMIT: Less Is More for Instruction Tuning in Text-to-SQL","year":2026,"arxiv_id":"2609.24186","external_url":"https://arxiv.org/abs/2609.24186","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24186"},{"ref":"SRC-0575","id":"arxiv-2609-24163","kind":"paper","title":"Listen, Critique, and Refine: RL-Based Self-Refinement for Instruction-Following Speech Synthesis","year":2026,"arxiv_id":"2609.24163","external_url":"https://arxiv.org/abs/2609.24163","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-24163"},{"ref":"SRC-0061","id":"arxiv-2403-07974","kind":"paper","title":"LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code","year":2024,"arxiv_id":"2403.07974","external_url":"https://arxiv.org/abs/2403.07974","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2403-07974"},{"ref":"SRC-0435","id":"arxiv-2608-23058","kind":"paper","title":"LLM-based Agents for Forecasting and Prediction: Methods, Training, Evaluation, and Applications","year":2026,"arxiv_id":"2608.23058","external_url":"https://arxiv.org/abs/2608.23058","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-23058"},{"ref":"SRC-0521","id":"arxiv-2609-23367","kind":"paper","title":"LLM-Based FORM Code Generation with Verification-Driven Fine-Tuning","year":2026,"arxiv_id":"2609.23367","external_url":"https://arxiv.org/abs/2609.23367","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-23367"},{"ref":"SRC-0336","id":"arxiv-2512-03373","kind":"paper","title":"LLM-Generated Ads: From Personalization Parity to Persuasion Superiority","year":2025,"arxiv_id":"2512.03373","external_url":"https://arxiv.org/abs/2512.03373","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2512-03373"},{"ref":"SRC-0620","id":"arxiv-2608-12626","kind":"paper","title":"LLMs Are Not Good Strategists, Yet Memory-Enhanced Agency Boosts Reasoning","year":2026,"arxiv_id":"2608.12626","external_url":"https://arxiv.org/abs/2608.12626","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-12626"},{"ref":"SRC-0582","id":"arxiv-2604-06236","kind":"paper","title":"LLMs Have Made Failure Worth Publishing","year":2026,"arxiv_id":"2604.06236","external_url":"https://arxiv.org/abs/2604.06236","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2604-06236"},{"ref":"SRC-0653","id":"arxiv-2608-27756","kind":"paper","title":"Load-Bearing Context: The Question Damage Score for Evaluating Context Reliance in Linguistic Reasoning","year":2026,"arxiv_id":"2608.27756","external_url":"https://arxiv.org/abs/2608.27756","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27756"},{"ref":"SRC-0313","id":"arxiv-2503-09089","kind":"paper","title":"LocAgent: Graph-Guided LLM Agents for Code Localization","year":2025,"arxiv_id":"2503.09089","external_url":"https://arxiv.org/abs/2503.09089","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2503-09089"},{"ref":"SRC-0334","id":"arxiv-2511-10459","kind":"paper","title":"LocalBench: Benchmarking LLMs on County-Level Local Knowledge and Reasoning","year":2025,"arxiv_id":"2511.10459","external_url":"https://arxiv.org/abs/2511.10459","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2511-10459"},{"ref":"SRC-0637","id":"arxiv-2608-25824","kind":"paper","title":"Localize-Then-Decide Guarantees for LLM Judgments","year":2026,"arxiv_id":"2608.25824","external_url":"https://arxiv.org/abs/2608.25824","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25824"},{"ref":"SRC-0161","id":"arxiv-2608-30959","kind":"paper","title":"LOCI: A Locator-Critic with Refinement Loop","year":2026,"arxiv_id":"2608.30959","external_url":"https://arxiv.org/abs/2608.30959","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30959"},{"ref":"SRC-0505","id":"arxiv-2609-21492","kind":"paper","title":"LogicTrack: Auditing Reasoning Trajectories of Large Language Models with Formal Logic Solvers","year":2026,"arxiv_id":"2609.21492","external_url":"https://arxiv.org/abs/2609.21492","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21492"},{"ref":"SRC-0064","id":"arxiv-2403-18802","kind":"paper","title":"Long-form factuality in large language models","year":2024,"arxiv_id":"2403.18802","external_url":"https://arxiv.org/abs/2403.18802","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2403-18802"},{"ref":"SRC-0478","id":"arxiv-2609-07303","kind":"paper","title":"Long-Horizon Language Model Reinforcement Learning via Progressive Point Matching","year":2026,"arxiv_id":"2609.07303","external_url":"https://arxiv.org/abs/2609.07303","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07303"},{"ref":"SRC-0135","id":"arxiv-2608-27580","kind":"paper","title":"LongGuard: Mechanistic Analysis and Training-Free Mitigation of Long-Context Failure in Safety Guardrails","year":2026,"arxiv_id":"2608.27580","external_url":"https://arxiv.org/abs/2608.27580","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27580"},{"ref":"SRC-0075","id":"arxiv-2410-10813","kind":"paper","title":"LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory","year":2024,"arxiv_id":"2410.10813","external_url":"https://arxiv.org/abs/2410.10813","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2410-10813"},{"ref":"SRC-0138","id":"arxiv-2608-28411","kind":"paper","title":"LongPIBench: A Long-Context Benchmark for Prompt Injection","year":2026,"arxiv_id":"2608.28411","external_url":"https://arxiv.org/abs/2608.28411","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-28411"},{"ref":"SRC-0272","id":"arxiv-2609-15745","kind":"paper","title":"Look Before You Leap: Factual Decoding with Internal Attribution Signals","year":2026,"arxiv_id":"2609.15745","external_url":"https://arxiv.org/abs/2609.15745","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15745"},{"ref":"SRC-0119","id":"arxiv-2608-24707","kind":"paper","title":"Lost in Speech: Trilingual Spoken Hallucination Detection Across Audio and Transcripts","year":2026,"arxiv_id":"2608.24707","external_url":"https://arxiv.org/abs/2608.24707","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24707"},{"ref":"SRC-0036","id":"arxiv-2307-03172","kind":"paper","title":"Lost in the Middle: How Language Models Use Long Contexts","year":2023,"arxiv_id":"2307.03172","external_url":"https://arxiv.org/abs/2307.03172","counts":{"claims":3},"url":"https://rsiratchet.com/sources/arxiv-2307-03172"},{"ref":"SRC-0332","id":"arxiv-2511-06346","kind":"paper","title":"LPFQA: A Long-Tail Professional Forum-based Benchmark for LLM Evaluation","year":2025,"arxiv_id":"2511.06346","external_url":"https://arxiv.org/abs/2511.06346","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2511-06346"},{"ref":"SRC-0386","id":"arxiv-2609-11043","kind":"paper","title":"LTLDiff: Finite Linear Temporal Logic-Guided Data Generation and Diffusion Policies for Multi-agent Robotic Manipulation","year":2026,"arxiv_id":"2609.11043","external_url":"https://arxiv.org/abs/2609.11043","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-11043"},{"ref":"SRC-0416","id":"arxiv-2609-18669","kind":"paper","title":"M$^3$P-R1: Reinforcement Learning for Large Language Model Guided Multi-Modal Motion Planning via MIP Code Generation","year":2026,"arxiv_id":"2609.18669","external_url":"https://arxiv.org/abs/2609.18669","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-18669"},{"ref":"SRC-0190","id":"boeckeler-2026-maintainability-sensors","kind":"post","title":"Maintainability sensors for coding agents","year":2026,"arxiv_id":null,"external_url":"https://martinfowler.com/articles/sensors-for-coding-agents.html","counts":{"claims":1},"url":"https://rsiratchet.com/sources/boeckeler-2026-maintainability-sensors"},{"ref":"SRC-0387","id":"arxiv-2609-11607","kind":"paper","title":"Making Alternative Data Work: Context-Augmented LLMs for Financial Forecasting","year":2026,"arxiv_id":"2609.11607","external_url":"https://arxiv.org/abs/2609.11607","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-11607"},{"ref":"SRC-0388","id":"arxiv-2609-11636","kind":"paper","title":"MAPLE: Memory-Augmented Planning with Language and Evolution","year":2026,"arxiv_id":"2609.11636","external_url":"https://arxiv.org/abs/2609.11636","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-11636"},{"ref":"SRC-0270","id":"arxiv-2609-15408","kind":"paper","title":"MarKey: Marginal Utility Guided Greedy Keyframe Selection for Long Video Understanding","year":2026,"arxiv_id":"2609.15408","external_url":"https://arxiv.org/abs/2609.15408","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15408"},{"ref":"SRC-0430","id":"arxiv-2608-21664","kind":"paper","title":"Measuring Activation Control in Large Language Models","year":2026,"arxiv_id":"2608.21664","external_url":"https://arxiv.org/abs/2608.21664","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-21664"},{"ref":"SRC-0078","id":"arxiv-2503-14499","kind":"paper","title":"Measuring AI Ability to Complete Long Software Tasks","year":2025,"arxiv_id":"2503.14499","external_url":"https://arxiv.org/abs/2503.14499","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2503-14499"},{"ref":"SRC-0226","id":"arxiv-2609-09090","kind":"paper","title":"Measuring LLM Sycophancy under Sustained Multi-Turn Pressure","year":2026,"arxiv_id":"2609.09090","external_url":"https://arxiv.org/abs/2609.09090","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09090"},{"ref":"SRC-0015","id":"arxiv-2211-03540","kind":"paper","title":"Measuring Progress on Scalable Oversight for Large Language Models","year":2022,"arxiv_id":"2211.03540","external_url":"https://arxiv.org/abs/2211.03540","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2211-03540"},{"ref":"SRC-0095","id":"arxiv-2608-18419","kind":"paper","title":"Mechanistic Interpretability of Structure-Aware Numerical Reasoning in LLaMA 3.1 8B","year":2026,"arxiv_id":"2608.18419","external_url":"https://arxiv.org/abs/2608.18419","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-18419"},{"ref":"SRC-0626","id":"arxiv-2608-14015","kind":"paper","title":"MedClaw: Heuristic Agent Harness for Long-Horizon Surgical Video Reasoning","year":2026,"arxiv_id":"2608.14015","external_url":"https://arxiv.org/abs/2608.14015","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-14015"},{"ref":"SRC-0350","id":"arxiv-2604-18418","kind":"paper","title":"MedProbeBench: Systematic Benchmarking at Deep Evidence Integration for Expert-level Medical Guideline","year":2026,"arxiv_id":"2604.18418","external_url":"https://arxiv.org/abs/2604.18418","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2604-18418"},{"ref":"SRC-0049","id":"arxiv-2310-08560","kind":"paper","title":"MemGPT: Towards LLMs as Operating Systems","year":2023,"arxiv_id":"2310.08560","external_url":"https://arxiv.org/abs/2310.08560","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2310-08560"},{"ref":"SRC-0577","id":"arxiv-2609-24238","kind":"paper","title":"Memory vs. Context? Influential Factors of Factual Recall in Language Models","year":2026,"arxiv_id":"2609.24238","external_url":"https://arxiv.org/abs/2609.24238","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-24238"},{"ref":"SRC-0085","id":"arxiv-2603-28052","kind":"paper","title":"Meta-Harness: End-to-End Optimization of Model Harnesses","year":2026,"arxiv_id":"2603.28052","external_url":"https://arxiv.org/abs/2603.28052","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2603-28052"},{"ref":"SRC-0427","id":"arxiv-2608-20853","kind":"paper","title":"MGAL: A Multilingual Granularity-Aware Long-Context Benchmark","year":2026,"arxiv_id":"2608.20853","external_url":"https://arxiv.org/abs/2608.20853","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-20853"},{"ref":"SRC-0426","id":"arxiv-2608-20314","kind":"paper","title":"MidTool: Mid-training Data Synthesis for Agentic Tool Use","year":2026,"arxiv_id":"2608.20314","external_url":"https://arxiv.org/abs/2608.20314","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-20314"},{"ref":"SRC-0362","id":"arxiv-2606-16710","kind":"paper","title":"Misinformation Propagation in Benign Multi-Agent Systems","year":2026,"arxiv_id":"2606.16710","external_url":"https://arxiv.org/abs/2606.16710","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-16710"},{"ref":"SRC-0309","id":"arxiv-2405-01563","kind":"paper","title":"Mitigating LLM Hallucinations via Conformal Abstention","year":2024,"arxiv_id":"2405.01563","external_url":"https://arxiv.org/abs/2405.01563","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2405-01563"},{"ref":"SRC-0121","id":"arxiv-2608-25267","kind":"paper","title":"Mitigating LLM sycophancy with RL-based fine-tuning: Bayesian Truth Serum approach","year":2026,"arxiv_id":"2608.25267","external_url":"https://arxiv.org/abs/2608.25267","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-25267"},{"ref":"SRC-0625","id":"arxiv-2608-13463","kind":"paper","title":"MLLM-Routed Heterogeneous Ensembles for Robust Cross-Dataset Image Classification","year":2026,"arxiv_id":"2608.13463","external_url":"https://arxiv.org/abs/2608.13463","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-13463"},{"ref":"SRC-0227","id":"arxiv-2609-09206","kind":"paper","title":"MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads","year":2026,"arxiv_id":"2609.09206","external_url":"https://arxiv.org/abs/2609.09206","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09206"},{"ref":"SRC-0374","id":"arxiv-2609-14857","kind":"paper","title":"ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement","year":2026,"arxiv_id":"2609.14857","external_url":"https://arxiv.org/abs/2609.14857","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14857"},{"ref":"SRC-0457","id":"arxiv-2608-30463","kind":"paper","title":"More Capable, Less Faithful: A Multilingual Analysis of Mathematical (Un)Solvability Detection in LLMs","year":2026,"arxiv_id":"2608.30463","external_url":"https://arxiv.org/abs/2608.30463","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30463"},{"ref":"SRC-0534","id":"arxiv-2405-17830","kind":"paper","title":"More Than Catastrophic Forgetting: Integrating General Capabilities For Domain-Specific LLMs","year":2024,"arxiv_id":"2405.17830","external_url":"https://arxiv.org/abs/2405.17830","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2405-17830"},{"ref":"SRC-0440","id":"arxiv-2608-24763","kind":"paper","title":"MoTE: Mixture of Task Experts for Multi-Task Video Understanding","year":2026,"arxiv_id":"2608.24763","external_url":"https://arxiv.org/abs/2608.24763","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24763"},{"ref":"SRC-0422","id":"arxiv-2608-18579","kind":"paper","title":"MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment","year":2026,"arxiv_id":"2608.18579","external_url":"https://arxiv.org/abs/2608.18579","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-18579"},{"ref":"SRC-0268","id":"arxiv-2609-14992","kind":"paper","title":"MTAC-IFBench: Benchmarking Instruction-Following in Multi-Turn Agentic Coding","year":2026,"arxiv_id":"2609.14992","external_url":"https://arxiv.org/abs/2609.14992","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14992"},{"ref":"SRC-0483","id":"arxiv-2609-09565","kind":"paper","title":"Multi-Agent Agentic Graph Learning via Structural Signatures","year":2026,"arxiv_id":"2609.09565","external_url":"https://arxiv.org/abs/2609.09565","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09565"},{"ref":"SRC-0033","id":"arxiv-2306-03314","kind":"paper","title":"Multi-Agent Collaboration: Harnessing the Power of Intelligent LLM Agents","year":2023,"arxiv_id":"2306.03314","external_url":"https://arxiv.org/abs/2306.03314","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2306-03314"},{"ref":"SRC-0600","id":"arxiv-2608-06865","kind":"paper","title":"Multi-Agent Forensic Reasoning for Generalizable Deepfake Video Detection","year":2026,"arxiv_id":"2608.06865","external_url":"https://arxiv.org/abs/2608.06865","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-06865"},{"ref":"SRC-0279","id":"arxiv-2609-11154","kind":"paper","title":"Multi-Faceted Evaluation and Mitigation of Emotion Hallucinations in MLLMs","year":2026,"arxiv_id":"2609.11154","external_url":"https://arxiv.org/abs/2609.11154","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-11154"},{"ref":"SRC-0251","id":"arxiv-1908-03316","kind":"paper","title":"Multi-modal Synthesis of Regular Expressions","year":2019,"arxiv_id":"1908.03316","external_url":"https://arxiv.org/abs/1908.03316","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-1908-03316"},{"ref":"SRC-0539","id":"arxiv-2602-20065","kind":"paper","title":"Multilingual Large Language Models do not comprehend all natural languages to equal degrees","year":2026,"arxiv_id":"2602.20065","external_url":"https://arxiv.org/abs/2602.20065","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2602-20065"},{"ref":"SRC-0195","id":"hashimoto-2026-ai-adoption-journey","kind":"post","title":"My AI Adoption Journey – Mitchell Hashimoto","year":2026,"arxiv_id":null,"external_url":"https://mitchellh.com/writing/my-ai-adoption-journey","counts":{"claims":2},"url":"https://rsiratchet.com/sources/hashimoto-2026-ai-adoption-journey"},{"ref":"SRC-0324","id":"arxiv-2510-02816","kind":"paper","title":"NCV: A Node-Wise Consistency Verification Approach for Low-Cost Structured Error Localization in LLM Reasoning","year":2025,"arxiv_id":"2510.02816","external_url":"https://arxiv.org/abs/2510.02816","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2510-02816"},{"ref":"SRC-0583","id":"fanelli-2012-negative-results-disappearing","kind":"paper","title":"Negative results are disappearing from most disciplines and countries","year":2012,"arxiv_id":null,"external_url":"https://doi.org/10.1007/s11192-011-0494-7","counts":{"claims":1},"url":"https://rsiratchet.com/sources/fanelli-2012-negative-results-disappearing"},{"ref":"SRC-0509","id":"arxiv-2609-21967","kind":"paper","title":"NemotronLabs VoiceChat: An Open Full-duplex Speech-to-Speech Model with Tool Calling Capabilities","year":2026,"arxiv_id":"2609.21967","external_url":"https://arxiv.org/abs/2609.21967","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21967"},{"ref":"SRC-0249","id":"arxiv-1608-03000","kind":"paper","title":"Neural Generation of Regular Expressions from Natural Language with Minimal Domain Knowledge","year":2016,"arxiv_id":"1608.03000","external_url":"https://arxiv.org/abs/1608.03000","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-1608-03000"},{"ref":"SRC-0644","id":"arxiv-2608-26329","kind":"paper","title":"Neuro-symbolic PRM: Enhancing Scientific Reasoning via Structured Traces and Symbolic Verification","year":2026,"arxiv_id":"2608.26329","external_url":"https://arxiv.org/abs/2608.26329","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26329"},{"ref":"SRC-0396","id":"arxiv-2609-14448","kind":"paper","title":"NeuroActiSep: Detecting Factual Hallucinations from Feed-Forward Neurons in a Single Pass","year":2026,"arxiv_id":"2609.14448","external_url":"https://arxiv.org/abs/2609.14448","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14448"},{"ref":"SRC-0475","id":"arxiv-2609-05913","kind":"paper","title":"Neuron-Guided Fine-Tuning: Unlocking Efficient Alignment Mechanisms for Large Language Models","year":2026,"arxiv_id":"2609.05913","external_url":"https://arxiv.org/abs/2609.05913","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05913"},{"ref":"SRC-0652","id":"arxiv-2608-27746","kind":"paper","title":"NormasTCU --- A Brazilian Portuguese IR Dataset and an Evaluation of LLM-as-a-Judge for Relevance Assessment","year":2026,"arxiv_id":"2608.27746","external_url":"https://arxiv.org/abs/2608.27746","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27746"},{"ref":"SRC-0564","id":"arxiv-2609-18084","kind":"paper","title":"Not All Layers Need Tuning: Diagnosing and Directing Adaptation in Vision-Language-Action Models","year":2026,"arxiv_id":"2609.18084","external_url":"https://arxiv.org/abs/2609.18084","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-18084"},{"ref":"SRC-0646","id":"arxiv-2608-26596","kind":"paper","title":"Not Just Reason, Not Just Scan: Reinforcement Learning for Proactive Scientific Error Verification over Academic Paper","year":2026,"arxiv_id":"2608.26596","external_url":"https://arxiv.org/abs/2608.26596","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26596"},{"ref":"SRC-0022","id":"arxiv-2302-12173","kind":"paper","title":"Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection","year":2023,"arxiv_id":"2302.12173","external_url":"https://arxiv.org/abs/2302.12173","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2302-12173"},{"ref":"SRC-0280","id":"arxiv-2609-11234","kind":"paper","title":"NovGauge: A Fine-Grained Benchmark for Diagnosing LLMs' Capability in Paper Novelty Assessment","year":2026,"arxiv_id":"2609.11234","external_url":"https://arxiv.org/abs/2609.11234","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-11234"},{"ref":"SRC-0542","id":"arxiv-2609-01971","kind":"paper","title":"NS-Copilot: An LLM-Driven Agent System for Autonomous Neuroscience Analysis","year":2026,"arxiv_id":"2609.01971","external_url":"https://arxiv.org/abs/2609.01971","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01971"},{"ref":"SRC-0623","id":"arxiv-2608-13129","kind":"paper","title":"Numeracy in Large Language Models: Fundamental Limitations and Paths to Improvement","year":2026,"arxiv_id":"2608.13129","external_url":"https://arxiv.org/abs/2608.13129","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-13129"},{"ref":"SRC-0215","id":"arxiv-2609-07135","kind":"paper","title":"NutriBench-Kitchen: Benchmarking Embodied AI for Nutrition Management","year":2026,"arxiv_id":"2609.07135","external_url":"https://arxiv.org/abs/2609.07135","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07135"},{"ref":"SRC-0281","id":"arxiv-2609-11244","kind":"paper","title":"OmniHallu: Unified Hallucination Detection for Cross-Modal Comprehension and Generation in Multimodal Large Language Models","year":2026,"arxiv_id":"2609.11244","external_url":"https://arxiv.org/abs/2609.11244","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-11244"},{"ref":"SRC-0248","id":"arxiv-2609-04835","kind":"paper","title":"On Epistemic Diversity in Large Language Models","year":2026,"arxiv_id":"2609.04835","external_url":"https://arxiv.org/abs/2609.04835","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-04835"},{"ref":"SRC-0199","id":"post-wu-rsi-outcome-variables","kind":"post","title":"On OpenAI's research-acceleration disclosure — the measured variables are all inputs","year":2026,"arxiv_id":null,"external_url":"https://x.com/cherylwoooo/status/2096741494866715055","counts":{"claims":1},"url":"https://rsiratchet.com/sources/post-wu-rsi-outcome-variables"},{"ref":"SRC-0063","id":"arxiv-2403-14380","kind":"paper","title":"On the Conversational Persuasiveness of Large Language Models: A Randomized Controlled Trial","year":2024,"arxiv_id":"2403.14380","external_url":"https://arxiv.org/abs/2403.14380","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2403-14380"},{"ref":"SRC-0322","id":"arxiv-2607-27081","kind":"paper","title":"On-Policy Distillation for LLM Safety: A Routing Approach to Template-Robust Realignment","year":2026,"arxiv_id":"2607.27081","external_url":"https://arxiv.org/abs/2607.27081","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2607-27081"},{"ref":"SRC-0446","id":"arxiv-2608-27338","kind":"paper","title":"One Model, Many Minds: Unlocking Multi-Agent Synergy in a Single Agent via Mixture of Roles","year":2026,"arxiv_id":"2608.27338","external_url":"https://arxiv.org/abs/2608.27338","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27338"},{"ref":"SRC-0611","id":"arxiv-2608-10178","kind":"paper","title":"One Recipe, Many Harnesses: What Self-Evolution Encodes Across Languages and Models","year":2026,"arxiv_id":"2608.10178","external_url":"https://arxiv.org/abs/2608.10178","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-10178"},{"ref":"SRC-0087","id":"arxiv-2605-17172","kind":"paper","title":"OpenJarvis: Personal AI, On Personal Devices","year":2026,"arxiv_id":"2605.17172","external_url":"https://arxiv.org/abs/2605.17172","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2605-17172"},{"ref":"SRC-0570","id":"arxiv-2609-21527","kind":"paper","title":"OpenMAS-GCom. A Diagnostic Benchmark for Graph-enhanced Multi-Agent Systems","year":2026,"arxiv_id":"2609.21527","external_url":"https://arxiv.org/abs/2609.21527","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-21527"},{"ref":"SRC-0312","id":"arxiv-2502-00350","kind":"paper","title":"OrcaLoca: An LLM Agent Framework for Software Issue Localization","year":2025,"arxiv_id":"2502.00350","external_url":"https://arxiv.org/abs/2502.00350","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2502-00350"},{"ref":"SRC-0284","id":"arxiv-2609-11737","kind":"paper","title":"ORCH: Organizational Principles Enable Collective Intelligence in Embodied AI","year":2026,"arxiv_id":"2609.11737","external_url":"https://arxiv.org/abs/2609.11737","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-11737"},{"ref":"SRC-0501","id":"arxiv-2609-21088","kind":"paper","title":"Origin Is All You Need: Provenance-Aware Transformers for Structural Trust-Boundary Separation","year":2026,"arxiv_id":"2609.21088","external_url":"https://arxiv.org/abs/2609.21088","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21088"},{"ref":"SRC-0372","id":"miller-2026-outcome-reputation-commons","kind":"observation","title":"Outcome-based reputation as the incentive for a human-and-agent knowledge commons","year":2026,"arxiv_id":null,"external_url":null,"counts":{"claims":1},"url":"https://rsiratchet.com/sources/miller-2026-outcome-reputation-commons"},{"ref":"SRC-0123","id":"arxiv-2608-25662","kind":"paper","title":"Overview of SHROOM-Visions 2026: A Shared Task on Hallucination Detection in Large Vision-Language Models","year":2026,"arxiv_id":"2608.25662","external_url":"https://arxiv.org/abs/2608.25662","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25662"},{"ref":"SRC-0019","id":"arxiv-2211-10435","kind":"paper","title":"PAL: Program-aided Language Models","year":2022,"arxiv_id":"2211.10435","external_url":"https://arxiv.org/abs/2211.10435","counts":{"claims":3},"url":"https://rsiratchet.com/sources/arxiv-2211-10435"},{"ref":"SRC-0101","id":"arxiv-2608-19598","kind":"paper","title":"PEA-DPO: Perception-Enhanced Alignment Direct Preference Optimization for MLLMs Alignment","year":2026,"arxiv_id":"2608.19598","external_url":"https://arxiv.org/abs/2608.19598","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-19598"},{"ref":"SRC-0347","id":"arxiv-2602-18092","kind":"paper","title":"Perceived Political Bias in LLMs Reduces Persuasive Abilities","year":2026,"arxiv_id":"2602.18092","external_url":"https://arxiv.org/abs/2602.18092","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2602-18092"},{"ref":"SRC-0247","id":"arxiv-2609-01188","kind":"paper","title":"PersuaRL: Reinforcement Learning-Driven Multi-Expert Selection for Persuasive Dialogue Generation in Insurance","year":2026,"arxiv_id":"2609.01188","external_url":"https://arxiv.org/abs/2609.01188","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-01188"},{"ref":"SRC-0220","id":"arxiv-2609-08067","kind":"paper","title":"Popular Knowledge Propagates More Errors in LLM Knowledge Updating","year":2026,"arxiv_id":"2609.08067","external_url":"https://arxiv.org/abs/2609.08067","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08067"},{"ref":"SRC-0581","id":"arxiv-2406-03980","kind":"paper","title":"Position: Embracing Negative Results in Machine Learning","year":2024,"arxiv_id":"2406.03980","external_url":"https://arxiv.org/abs/2406.03980","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2406-03980"},{"ref":"SRC-0133","id":"arxiv-2608-27165","kind":"paper","title":"Prediction of Prediction (PoP): Inter-Layer Activation Fusion for Single-Pass Hallucination Detection in Large Language Models","year":2026,"arxiv_id":"2608.27165","external_url":"https://arxiv.org/abs/2608.27165","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27165"},{"ref":"SRC-0439","id":"arxiv-2608-24192","kind":"paper","title":"Preference Data Selection for Mitigating the Alignment Tax in Large Language Models","year":2026,"arxiv_id":"2608.24192","external_url":"https://arxiv.org/abs/2608.24192","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24192"},{"ref":"SRC-0209","id":"arxiv-2609-05235","kind":"paper","title":"PRICE: A Systematic Study of LLM Adaptation Choices for Bitcoin Price Forecasting","year":2026,"arxiv_id":"2609.05235","external_url":"https://arxiv.org/abs/2609.05235","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-05235"},{"ref":"SRC-0111","id":"arxiv-2608-23552","kind":"paper","title":"Prime Agent: A Self-Improving RLM Harness","year":2026,"arxiv_id":"2608.23552","external_url":"https://arxiv.org/abs/2608.23552","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-23552"},{"ref":"SRC-0472","id":"arxiv-2609-01658","kind":"paper","title":"PRO-Step: Step-level Process Reward Optimization for Retrieval-Augmented Generation","year":2026,"arxiv_id":"2609.01658","external_url":"https://arxiv.org/abs/2609.01658","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01658"},{"ref":"SRC-0238","id":"arxiv-2609-10540","kind":"paper","title":"Programmable World Model","year":2026,"arxiv_id":"2609.10540","external_url":"https://arxiv.org/abs/2609.10540","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10540"},{"ref":"SRC-0640","id":"arxiv-2608-25992","kind":"paper","title":"ProgRouter: Online Progress-Guided Orchestration for Multi-Agent LLM Workflows under Quality-Cost Tradeoffs","year":2026,"arxiv_id":"2608.25992","external_url":"https://arxiv.org/abs/2608.25992","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25992"},{"ref":"SRC-0608","id":"arxiv-2608-09068","kind":"paper","title":"Pseudo2CodeQA: A Benchmark for LLM-Based Structured Algorithmic Reasoning in Code Generation","year":2026,"arxiv_id":"2608.09068","external_url":"https://arxiv.org/abs/2608.09068","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-09068"},{"ref":"SRC-0054","id":"arxiv-2312-04724","kind":"paper","title":"Purple Llama CyberSecEval: A Secure Coding Benchmark for Language Models","year":2023,"arxiv_id":"2312.04724","external_url":"https://arxiv.org/abs/2312.04724","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2312-04724"},{"ref":"SRC-0514","id":"arxiv-2609-22765","kind":"paper","title":"Quality over Quantity: Diversity-Aware Data Selection for Efficient Verilog Code Generation","year":2026,"arxiv_id":"2609.22765","external_url":"https://arxiv.org/abs/2609.22765","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22765"},{"ref":"SRC-0308","id":"arxiv-2311-09677","kind":"paper","title":"R-Tuning: Instructing Large Language Models to Say `I Don't Know'","year":2023,"arxiv_id":"2311.09677","external_url":"https://arxiv.org/abs/2311.09677","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2311-09677"},{"ref":"SRC-0316","id":"arxiv-2409-12558","kind":"paper","title":"RAD-Bench: Evaluating Large Language Models Capabilities in Retrieval Augmented Dialogues","year":2024,"arxiv_id":"2409.12558","external_url":"https://arxiv.org/abs/2409.12558","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2409-12558"},{"ref":"SRC-0487","id":"arxiv-2609-10092","kind":"paper","title":"RAP: Research Attention Prediction Reveals Target-Conditioned Evidence Acquisition Biases","year":2026,"arxiv_id":"2609.10092","external_url":"https://arxiv.org/abs/2609.10092","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10092"},{"ref":"SRC-0152","id":"arxiv-2608-30041","kind":"paper","title":"Reachability-Based Capability Confinement for LLM Agents under Indirect Prompt Injection","year":2026,"arxiv_id":"2608.30041","external_url":"https://arxiv.org/abs/2608.30041","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30041"},{"ref":"SRC-0014","id":"arxiv-2210-03629","kind":"paper","title":"ReAct: Synergizing Reasoning and Acting in Language Models","year":2022,"arxiv_id":"2210.03629","external_url":"https://arxiv.org/abs/2210.03629","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2210-03629"},{"ref":"SRC-0556","id":"arxiv-2609-10895","kind":"paper","title":"ReactHuman: A Physics-Grounded Benchmark for Human-Like Reactive Decision-Making in Embodied Multimodal LLMs","year":2026,"arxiv_id":"2609.10895","external_url":"https://arxiv.org/abs/2609.10895","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10895"},{"ref":"SRC-0580","id":"arxiv-2509-25140","kind":"paper","title":"ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory","year":2025,"arxiv_id":"2509.25140","external_url":"https://arxiv.org/abs/2509.25140","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2509-25140"},{"ref":"SRC-0610","id":"arxiv-2608-10149","kind":"paper","title":"REATS: LLM Reasoning-based Ensemble Learning for Adaptive Time Series Forecasting","year":2026,"arxiv_id":"2608.10149","external_url":"https://arxiv.org/abs/2608.10149","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-10149"},{"ref":"SRC-0659","id":"arxiv-2608-29109","kind":"paper","title":"Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions","year":2026,"arxiv_id":"2608.29109","external_url":"https://arxiv.org/abs/2608.29109","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29109"},{"ref":"SRC-0083","id":"arxiv-2512-24601","kind":"paper","title":"Recursive Language Models","year":2025,"arxiv_id":"2512.24601","external_url":"https://arxiv.org/abs/2512.24601","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2512-24601"},{"ref":"SRC-0042","id":"arxiv-2308-15022","kind":"paper","title":"Recursively Summarizing Enables Long-Term Dialogue Memory in Large Language Models","year":2023,"arxiv_id":"2308.15022","external_url":"https://arxiv.org/abs/2308.15022","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2308-15022"},{"ref":"SRC-0624","id":"arxiv-2608-13292","kind":"paper","title":"Refine After Generation: Toward Correct and Concise Patches in LLM-based Program Repair","year":2026,"arxiv_id":"2608.13292","external_url":"https://arxiv.org/abs/2608.13292","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-13292"},{"ref":"SRC-0535","id":"arxiv-2502-09183","kind":"paper","title":"RefineCoder: Iterative Improving of Large Language Models via Adaptive Critique Refinement for Code Generation","year":2025,"arxiv_id":"2502.09183","external_url":"https://arxiv.org/abs/2502.09183","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2502-09183"},{"ref":"SRC-0023","id":"arxiv-2303-11366","kind":"paper","title":"Reflexion: Language Agents with Verbal Reinforcement Learning","year":2023,"arxiv_id":"2303.11366","external_url":"https://arxiv.org/abs/2303.11366","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2303-11366"},{"ref":"SRC-0183","id":"arxiv-2609-03267","kind":"paper","title":"Refusing the Impossible: A Taxonomy and Benchmark for Code Hallucination in Large Language Models","year":2026,"arxiv_id":"2609.03267","external_url":"https://arxiv.org/abs/2609.03267","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-03267"},{"ref":"SRC-0412","id":"arxiv-2609-18207","kind":"paper","title":"Reinforcement Learning for Real-Time Vision-Language-Action Policies","year":2026,"arxiv_id":"2609.18207","external_url":"https://arxiv.org/abs/2609.18207","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-18207"},{"ref":"SRC-0242","id":"arxiv-2608-28233","kind":"paper","title":"REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Features","year":2026,"arxiv_id":"2608.28233","external_url":"https://arxiv.org/abs/2608.28233","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-28233"},{"ref":"SRC-0173","id":"arxiv-2609-01318","kind":"paper","title":"Reliability Challenges in Diffusion Vision-Language Models","year":2026,"arxiv_id":"2609.01318","external_url":"https://arxiv.org/abs/2609.01318","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01318"},{"ref":"SRC-0550","id":"arxiv-2609-03619","kind":"paper","title":"Remember and Reweight: Enhancing Multi-Agent Debate with Experience Memory and Confidence Estimation","year":2026,"arxiv_id":"2609.03619","external_url":"https://arxiv.org/abs/2609.03619","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-03619"},{"ref":"SRC-0286","id":"arxiv-2609-12230","kind":"paper","title":"Repair Before Reinforce: Context-Augmented Knowledge Graph Reasoning for Multi-Hop Question Answering","year":2026,"arxiv_id":"2609.12230","external_url":"https://arxiv.org/abs/2609.12230","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-12230"},{"ref":"SRC-0639","id":"arxiv-2608-25920","kind":"paper","title":"Repair or Resample? Rethinking Failure Debugging in LLM Multi-Agent Systems","year":2026,"arxiv_id":"2608.25920","external_url":"https://arxiv.org/abs/2608.25920","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25920"},{"ref":"SRC-0467","id":"arxiv-2609-00834","kind":"paper","title":"Replacing Training with Memory: Listwise Selection for Text-to-SQL","year":2026,"arxiv_id":"2609.00834","external_url":"https://arxiv.org/abs/2609.00834","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00834"},{"ref":"SRC-0425","id":"arxiv-2608-19854","kind":"paper","title":"Repo0: Design-Driven Zero-to-All Code Generation","year":2026,"arxiv_id":"2608.19854","external_url":"https://arxiv.org/abs/2608.19854","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-19854"},{"ref":"SRC-0311","id":"arxiv-2410-14684","kind":"paper","title":"RepoGraph: Enhancing AI Software Engineering with Repository-level Code Graph","year":2024,"arxiv_id":"2410.14684","external_url":"https://arxiv.org/abs/2410.14684","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2410-14684"},{"ref":"SRC-0593","id":"arxiv-2608-04783","kind":"paper","title":"RepoProbe: Benchmarking Architecture-Aware Repository Comprehension with Checklists","year":2026,"arxiv_id":"2608.04783","external_url":"https://arxiv.org/abs/2608.04783","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-04783"},{"ref":"SRC-0379","id":"supermannx-2026-dopamine-neurons-post","kind":"post","title":"Researchers found dopamine neurons inside LLMs","year":2026,"arxiv_id":null,"external_url":"https://x.com/thesupermannx/status/2102027727192932469","counts":{"claims":1},"url":"https://rsiratchet.com/sources/supermannx-2026-dopamine-neurons-post"},{"ref":"SRC-0369","id":"arxiv-2609-19759","kind":"paper","title":"Rethinking Multi-Agent Collaboration: When More Is Less","year":2026,"arxiv_id":"2609.19759","external_url":"https://arxiv.org/abs/2609.19759","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-19759"},{"ref":"SRC-0476","id":"arxiv-2609-05933","kind":"paper","title":"Rethinking the Evaluation of Efficiency Methods for Multi-Agent Systems","year":2026,"arxiv_id":"2609.05933","external_url":"https://arxiv.org/abs/2609.05933","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05933"},{"ref":"SRC-0004","id":"arxiv-2005-11401","kind":"paper","title":"Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks","year":2020,"arxiv_id":"2005.11401","external_url":"https://arxiv.org/abs/2005.11401","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2005-11401"},{"ref":"SRC-0244","id":"arxiv-2608-28707","kind":"paper","title":"ReVA: A Region-Aware Visual Assistant for Visually Grounded Question Answering","year":2026,"arxiv_id":"2608.28707","external_url":"https://arxiv.org/abs/2608.28707","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-28707"},{"ref":"SRC-0062","id":"arxiv-2403-13799","kind":"paper","title":"Reverse Training to Nurse the Reversal Curse","year":2024,"arxiv_id":"2403.13799","external_url":"https://arxiv.org/abs/2403.13799","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2403-13799"},{"ref":"SRC-0359","id":"arxiv-2606-01610","kind":"paper","title":"Revisiting Ripple Effects in Knowledge Editing through Pressure-Aware Joint Neighborhood Optimization","year":2026,"arxiv_id":"2606.01610","external_url":"https://arxiv.org/abs/2606.01610","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-01610"},{"ref":"SRC-0424","id":"arxiv-2608-19075","kind":"paper","title":"ReWEIGH the Evidence: Calibrating Token-Level Ordinal Visual Evidence to Mitigate Hallucinations in Large Vision-Language Models","year":2026,"arxiv_id":"2608.19075","external_url":"https://arxiv.org/abs/2608.19075","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-19075"},{"ref":"SRC-0411","id":"arxiv-2609-17985","kind":"paper","title":"RideWay: Benchmarking Efficient Task Completion for Tool-Using Language Agents","year":2026,"arxiv_id":"2609.17985","external_url":"https://arxiv.org/abs/2609.17985","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17985"},{"ref":"SRC-0665","id":"arxiv-2608-29958","kind":"paper","title":"RIDGE: Region-Informed Derivative-Guided Evidence Selection for Long Video Understanding","year":2026,"arxiv_id":"2608.29958","external_url":"https://arxiv.org/abs/2608.29958","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29958"},{"ref":"SRC-0330","id":"arxiv-2510-25427","kind":"paper","title":"RLMEval: Evaluating Research-Level Neural Theorem Proving","year":2025,"arxiv_id":"2510.25427","external_url":"https://arxiv.org/abs/2510.25427","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2510-25427"},{"ref":"SRC-0210","id":"arxiv-2609-05324","kind":"paper","title":"RoboSPA: Can VLA Models Go Beyond Simple Scenes and Short-Horizon Tasks?","year":2026,"arxiv_id":"2609.05324","external_url":"https://arxiv.org/abs/2609.05324","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-05324"},{"ref":"SRC-0413","id":"arxiv-2609-18304","kind":"paper","title":"Rollback the World, Keep the Reflection: Rollback-Induced Reflection for Long-Horizon LLM Agents","year":2026,"arxiv_id":"2609.18304","external_url":"https://arxiv.org/abs/2609.18304","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-18304"},{"ref":"SRC-0134","id":"arxiv-2608-27496","kind":"paper","title":"ROPE: Routed Origin Policy Enforcement against Indirect Prompt Injection","year":2026,"arxiv_id":"2608.27496","external_url":"https://arxiv.org/abs/2608.27496","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27496"},{"ref":"SRC-0588","id":"arxiv-2608-03341","kind":"paper","title":"Route-Align-Verify for Functional Correctness in Code Generation","year":2026,"arxiv_id":"2608.03341","external_url":"https://arxiv.org/abs/2608.03341","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-03341"},{"ref":"SRC-0116","id":"arxiv-2608-24138","kind":"paper","title":"Rubrics as Visual-Repair Context for Self-Evolving UI-to-Code Generation","year":2026,"arxiv_id":"2608.24138","external_url":"https://arxiv.org/abs/2608.24138","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24138"},{"ref":"SRC-0065","id":"arxiv-2404-06654","kind":"paper","title":"RULER: What's the Real Context Size of Your Long-Context Language Models?","year":2024,"arxiv_id":"2404.06654","external_url":"https://arxiv.org/abs/2404.06654","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2404-06654"},{"ref":"SRC-0180","id":"arxiv-2609-02731","kind":"paper","title":"RVSD: Retrieval Vision Sparse Decoding for Mitigating Visual Hallucinations in Large Vision-Language Models","year":2026,"arxiv_id":"2609.02731","external_url":"https://arxiv.org/abs/2609.02731","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02731"},{"ref":"SRC-0463","id":"arxiv-2608-31100","kind":"paper","title":"S3Gym: Can LLMs Turn Self-Testing and Self-Judging into Self-Improvement?","year":2026,"arxiv_id":"2608.31100","external_url":"https://arxiv.org/abs/2608.31100","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-31100"},{"ref":"SRC-0223","id":"arxiv-2609-08444","kind":"paper","title":"Safe Task Planning with Long-Term Graph Memory for Embodied Agents","year":2026,"arxiv_id":"2609.08444","external_url":"https://arxiv.org/abs/2609.08444","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08444"},{"ref":"SRC-0537","id":"arxiv-2503-15707","kind":"paper","title":"Safety Aware Task Planning via Large Language Models in Robotics","year":2025,"arxiv_id":"2503.15707","external_url":"https://arxiv.org/abs/2503.15707","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2503-15707"},{"ref":"SRC-0494","id":"arxiv-2609-17775","kind":"paper","title":"SAGE: Governed Artifact Generation from Enterprise Guidelines","year":2026,"arxiv_id":"2609.17775","external_url":"https://arxiv.org/abs/2609.17775","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17775"},{"ref":"SRC-0363","id":"arxiv-2606-29520","kind":"paper","title":"SAKE: Software Architectural Knowledge Evaluation Benchmark for Large Language Models","year":2026,"arxiv_id":"2606.29520","external_url":"https://arxiv.org/abs/2606.29520","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-29520"},{"ref":"SRC-0602","id":"arxiv-2608-07302","kind":"paper","title":"Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination","year":2026,"arxiv_id":"2608.07302","external_url":"https://arxiv.org/abs/2608.07302","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-07302"},{"ref":"SRC-0058","id":"arxiv-2402-14848","kind":"paper","title":"Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models","year":2024,"arxiv_id":"2402.14848","external_url":"https://arxiv.org/abs/2402.14848","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2402-14848"},{"ref":"SRC-0571","id":"arxiv-2609-21666","kind":"paper","title":"Samsone: A Family of Open Small Audio Language Models for On-Device Inference","year":2026,"arxiv_id":"2609.21666","external_url":"https://arxiv.org/abs/2609.21666","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-21666"},{"ref":"SRC-0103","id":"arxiv-2608-19842","kind":"paper","title":"SAPO: Single-Rollout Autoregressive Policy Optimization for Agentic Reinforcement Learning","year":2026,"arxiv_id":"2608.19842","external_url":"https://arxiv.org/abs/2608.19842","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-19842"},{"ref":"SRC-0298","id":"arxiv-2609-16601","kind":"paper","title":"SAVOR: Self-Aware Visual Grounding via Confidence-Calibrated Reinforcement Learning for Multimodal Hallucination Mitigation","year":2026,"arxiv_id":"2609.16601","external_url":"https://arxiv.org/abs/2609.16601","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-16601"},{"ref":"SRC-0335","id":"arxiv-2511-10519","kind":"paper","title":"Say It Differently: Linguistic Styles as Jailbreak Vectors","year":2025,"arxiv_id":"2511.10519","external_url":"https://arxiv.org/abs/2511.10519","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2511-10519"},{"ref":"SRC-0567","id":"arxiv-2609-20398","kind":"paper","title":"Schema-Anchored Latent Reasoning for Semantic Parsing-Based Knowledge Base Question Answering","year":2026,"arxiv_id":"2609.20398","external_url":"https://arxiv.org/abs/2609.20398","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-20398"},{"ref":"SRC-0370","id":"arxiv-2609-19644","kind":"paper","title":"ScientistTwo: Pioneering the Human Knowledge Frontier with Autonomous AI","year":2026,"arxiv_id":"2609.19644","external_url":"https://arxiv.org/abs/2609.19644","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-19644"},{"ref":"SRC-0127","id":"arxiv-2608-25973","kind":"paper","title":"SciMIF: Understanding Multimodal Instruction Following in Scientific Domains","year":2026,"arxiv_id":"2608.25973","external_url":"https://arxiv.org/abs/2608.25973","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25973"},{"ref":"SRC-0230","id":"arxiv-2609-09672","kind":"paper","title":"SEA-SpeechBench: A Large-Scale Multitask Benchmark for Speech Understanding Across Southeast Asia","year":2026,"arxiv_id":"2609.09672","external_url":"https://arxiv.org/abs/2609.09672","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09672"},{"ref":"SRC-0341","id":"arxiv-2512-17260","kind":"paper","title":"Seed-Prover 1.5: Mastering Undergraduate-Level Theorem Proving via Learning from Experience","year":2025,"arxiv_id":"2512.17260","external_url":"https://arxiv.org/abs/2512.17260","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2512-17260"},{"ref":"SRC-0333","id":"arxiv-2511-07364","kind":"paper","title":"Self-Evaluating LLMs for Multi-Step Tasks: Stepwise Confidence Estimation for Failure Detection","year":2025,"arxiv_id":"2511.07364","external_url":"https://arxiv.org/abs/2511.07364","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2511-07364"},{"ref":"SRC-0090","id":"arxiv-2606-09498","kind":"paper","title":"Self-Harness: Harnesses That Improve Themselves","year":2026,"arxiv_id":"2606.09498","external_url":"https://arxiv.org/abs/2606.09498","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-09498"},{"ref":"SRC-0050","id":"arxiv-2310-11511","kind":"paper","title":"Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection","year":2023,"arxiv_id":"2310.11511","external_url":"https://arxiv.org/abs/2310.11511","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2310-11511"},{"ref":"SRC-0024","id":"arxiv-2303-17651","kind":"paper","title":"Self-Refine: Iterative Refinement with Self-Feedback","year":2023,"arxiv_id":"2303.17651","external_url":"https://arxiv.org/abs/2303.17651","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2303-17651"},{"ref":"SRC-0431","id":"arxiv-2608-22090","kind":"paper","title":"Semantic Reasoning Denoising: Correcting Language Model Reasoning with Semantic Operators","year":2026,"arxiv_id":"2608.22090","external_url":"https://arxiv.org/abs/2608.22090","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-22090"},{"ref":"SRC-0627","id":"arxiv-2608-14707","kind":"paper","title":"Semantic Uncertainty-Guided Orchestration in Hierarchical Multi-Agent Systems","year":2026,"arxiv_id":"2608.14707","external_url":"https://arxiv.org/abs/2608.14707","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-14707"},{"ref":"SRC-0303","id":"arxiv-2609-17269","kind":"paper","title":"Semantic-Spatial Agreement Verification for Mitigating Object Hallucination in Multimodal Large Language Models","year":2026,"arxiv_id":"2609.17269","external_url":"https://arxiv.org/abs/2609.17269","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17269"},{"ref":"SRC-0546","id":"arxiv-2609-02387","kind":"paper","title":"Semantics-Guided Automatic Tensorization for Multiobjective Evolutionary Algorithms: A Multi-Agent Framework","year":2026,"arxiv_id":"2609.02387","external_url":"https://arxiv.org/abs/2609.02387","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02387"},{"ref":"SRC-0258","id":"zhong-2018-semregex","kind":"paper","title":"SemRegex: A Semantics-Based Approach for Generating Regular Expressions from Natural Language Specifications","year":2018,"arxiv_id":null,"external_url":"https://aclanthology.org/D18-1189/","counts":{"claims":1},"url":"https://rsiratchet.com/sources/zhong-2018-semregex"},{"ref":"SRC-0410","id":"arxiv-2609-17848","kind":"paper","title":"SFT or RL for Tool-Calling Agents? A Controlled Study Across Data, Method, and Scale","year":2026,"arxiv_id":"2609.17848","external_url":"https://arxiv.org/abs/2609.17848","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17848"},{"ref":"SRC-0549","id":"arxiv-2609-02780","kind":"paper","title":"ShallowStream: Index Shallow then Answer Deep for Streaming Video Understanding","year":2026,"arxiv_id":"2609.02780","external_url":"https://arxiv.org/abs/2609.02780","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02780"},{"ref":"SRC-0041","id":"arxiv-2308-03958","kind":"paper","title":"Simple synthetic data reduces sycophancy in large language models","year":2023,"arxiv_id":"2308.03958","external_url":"https://arxiv.org/abs/2308.03958","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2308-03958"},{"ref":"SRC-0337","id":"arxiv-2512-08743","kind":"paper","title":"Single-Agent Scaling Fails Multi-Agent Intelligence: Towards Foundation Models with Native Multi-Agent Intelligence","year":2025,"arxiv_id":"2512.08743","external_url":"https://arxiv.org/abs/2512.08743","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2512-08743"},{"ref":"SRC-0159","id":"arxiv-2608-30703","kind":"paper","title":"SingProbe Technical Report","year":2026,"arxiv_id":"2608.30703","external_url":"https://arxiv.org/abs/2608.30703","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30703"},{"ref":"SRC-0154","id":"arxiv-2608-30207","kind":"paper","title":"SIR: Self-improving Red-teaming for Compute Use Agents","year":2026,"arxiv_id":"2608.30207","external_url":"https://arxiv.org/abs/2608.30207","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30207"},{"ref":"SRC-0643","id":"arxiv-2608-26263","kind":"paper","title":"SKILL.state: Scalable Long-Horizon Agent Skills","year":2026,"arxiv_id":"2608.26263","external_url":"https://arxiv.org/abs/2608.26263","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26263"},{"ref":"SRC-0598","id":"arxiv-2608-05970","kind":"paper","title":"SkillMemo: Expert-guided Skill Memory Framework for Compositional Embodied Manipulation","year":2026,"arxiv_id":"2608.05970","external_url":"https://arxiv.org/abs/2608.05970","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-05970"},{"ref":"SRC-0320","id":"arxiv-2606-22676","kind":"paper","title":"Skin-Deep: A Geometric Diagnostic for Alignment Fragility in Large Language Model Representations","year":2026,"arxiv_id":"2606.22676","external_url":"https://arxiv.org/abs/2606.22676","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-22676"},{"ref":"SRC-0576","id":"arxiv-2609-24198","kind":"paper","title":"SKstars at SHROOM: Visions Agreement-Guided Ensembling of Zero-Shot and LoRA-Adapted Vision--Language Models","year":2026,"arxiv_id":"2609.24198","external_url":"https://arxiv.org/abs/2609.24198","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-24198"},{"ref":"SRC-0056","id":"arxiv-2401-05566","kind":"paper","title":"Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training","year":2024,"arxiv_id":"2401.05566","external_url":"https://arxiv.org/abs/2401.05566","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2401-05566"},{"ref":"SRC-0106","id":"arxiv-2608-22472","kind":"paper","title":"Small Reasoning Models are Instruction Followers in Function Calling","year":2026,"arxiv_id":"2608.22472","external_url":"https://arxiv.org/abs/2608.22472","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-22472"},{"ref":"SRC-0591","id":"arxiv-2608-04009","kind":"paper","title":"SocietyBench: Forecasting Counterfactual Social-World Evolution","year":2026,"arxiv_id":"2608.04009","external_url":"https://arxiv.org/abs/2608.04009","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-04009"},{"ref":"SRC-0390","id":"arxiv-2609-12413","kind":"paper","title":"SoK: Rethinking Jailbreaking in the Era of Agentic AI: Attacks, Defenses, and Practical Consideration","year":2026,"arxiv_id":"2609.12413","external_url":"https://arxiv.org/abs/2609.12413","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-12413"},{"ref":"SRC-0304","id":"arxiv-2609-20519","kind":"paper","title":"SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness","year":2026,"arxiv_id":"2609.20519","external_url":"https://arxiv.org/abs/2609.20519","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-20519"},{"ref":"SRC-0683","id":"arxiv-2609-05928","kind":"paper","title":"Solving versus Verifying: Catching Contradictions in Tax Reasoning Systems","year":2026,"arxiv_id":"2609.05928","external_url":"https://arxiv.org/abs/2609.05928","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05928"},{"ref":"SRC-0099","id":"arxiv-2608-19197","kind":"paper","title":"SPADE: Self-Play in Adaptive Synthetic Executable Environments","year":2026,"arxiv_id":"2608.19197","external_url":"https://arxiv.org/abs/2608.19197","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-19197"},{"ref":"SRC-0151","id":"arxiv-2608-29974","kind":"paper","title":"SpanCalib-VLM: Calibrated Hallucination Span Detection in Vision-Language Models","year":2026,"arxiv_id":"2608.29974","external_url":"https://arxiv.org/abs/2608.29974","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29974"},{"ref":"SRC-0377","id":"arxiv-2602-00986","kind":"paper","title":"Sparse Reward Subsystem in Large Language Models","year":2026,"arxiv_id":"2602.00986","external_url":"https://arxiv.org/abs/2602.00986","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2602-00986"},{"ref":"SRC-0212","id":"arxiv-2609-06041","kind":"paper","title":"SpecCoder: Specification-Aware Code Generation with Curriculum Dual-Task Reinforcement Learning","year":2026,"arxiv_id":"2609.06041","external_url":"https://arxiv.org/abs/2609.06041","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-06041"},{"ref":"SRC-0250","id":"arxiv-1809-08887","kind":"paper","title":"Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task","year":2018,"arxiv_id":"1809.08887","external_url":"https://arxiv.org/abs/1809.08887","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-1809-08887"},{"ref":"SRC-0297","id":"arxiv-2609-16268","kind":"paper","title":"Spurious Tool Use: When RL Agents Learn the Wrong Reason to Act","year":2026,"arxiv_id":"2609.16268","external_url":"https://arxiv.org/abs/2609.16268","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-16268"},{"ref":"SRC-0110","id":"arxiv-2608-23493","kind":"paper","title":"SRPO: Self-Reflective Policy Optimization for Long-Horizon Reasoning","year":2026,"arxiv_id":"2608.23493","external_url":"https://arxiv.org/abs/2608.23493","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-23493"},{"ref":"SRC-0477","id":"arxiv-2609-07148","kind":"paper","title":"Stable-MM-R1: Anchoring Multimodal Reasoning Dynamics via Entropy-Guided Stratification","year":2026,"arxiv_id":"2609.07148","external_url":"https://arxiv.org/abs/2609.07148","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07148"},{"ref":"SRC-0500","id":"arxiv-2609-20791","kind":"paper","title":"StageGuard: Learning Stage Transitions for Long-Horizon Robot Tasks via Agentic Distillation","year":2026,"arxiv_id":"2609.20791","external_url":"https://arxiv.org/abs/2609.20791","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-20791"},{"ref":"SRC-0479","id":"arxiv-2609-07344","kind":"paper","title":"Staying on the Attack Path: Structured State for Long-Horizon Automated Penetration Testing","year":2026,"arxiv_id":"2609.07344","external_url":"https://arxiv.org/abs/2609.07344","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07344"},{"ref":"SRC-0295","id":"arxiv-2609-15983","kind":"paper","title":"Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and Theoretical Computer Science","year":2026,"arxiv_id":"2609.15983","external_url":"https://arxiv.org/abs/2609.15983","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15983"},{"ref":"SRC-0445","id":"arxiv-2608-27225","kind":"paper","title":"STEP: State-Aware Task Estimation and Planning with Multi-Modal LLMs for Human-Robot Collaboration","year":2026,"arxiv_id":"2608.27225","external_url":"https://arxiv.org/abs/2608.27225","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27225"},{"ref":"SRC-0522","id":"arxiv-2609-23716","kind":"paper","title":"STEVE: Stabilizing Textual Gradient-Based Prompt Optimization via Error-Driven Refinement and Regularized Verification","year":2026,"arxiv_id":"2609.23716","external_url":"https://arxiv.org/abs/2609.23716","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-23716"},{"ref":"SRC-0162","id":"arxiv-2608-30987","kind":"paper","title":"Stick to What You Know: A Study of Knowledge-Aligned Supervised Fine-Tuning","year":2026,"arxiv_id":"2608.30987","external_url":"https://arxiv.org/abs/2608.30987","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30987"},{"ref":"SRC-0481","id":"arxiv-2609-09264","kind":"paper","title":"StochBench: A Domain-Specific Benchmark for Stochastic Processes in Lean","year":2026,"arxiv_id":"2609.09264","external_url":"https://arxiv.org/abs/2609.09264","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09264"},{"ref":"SRC-0092","id":"arxiv-2607-14166","kind":"paper","title":"Stop Means Stop: Measuring and Repairing the Enforcement Gap in Agent-Framework Control Primitives","year":2026,"arxiv_id":"2607.14166","external_url":"https://arxiv.org/abs/2607.14166","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2607-14166"},{"ref":"SRC-0213","id":"arxiv-2609-06117","kind":"paper","title":"STQA: A Benchmark for Stock-Focused Tabular Question Answering over Historical and Forecasted Data","year":2026,"arxiv_id":"2609.06117","external_url":"https://arxiv.org/abs/2609.06117","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-06117"},{"ref":"SRC-0485","id":"arxiv-2609-09899","kind":"paper","title":"Strangers to Themselves: What Language Models Say About Themselves Is Generic","year":2026,"arxiv_id":"2609.09899","external_url":"https://arxiv.org/abs/2609.09899","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09899"},{"ref":"SRC-0597","id":"arxiv-2608-05703","kind":"paper","title":"StreamArena: Toward Continuous, Interactive, and Long-Horizon Agentic Streaming Video Understanding","year":2026,"arxiv_id":"2608.05703","external_url":"https://arxiv.org/abs/2608.05703","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-05703"},{"ref":"SRC-0057","id":"arxiv-2402-06363","kind":"paper","title":"StruQ: Defending Against Prompt Injection with Structured Queries","year":2024,"arxiv_id":"2402.06363","external_url":"https://arxiv.org/abs/2402.06363","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2402-06363"},{"ref":"SRC-0568","id":"arxiv-2609-20712","kind":"paper","title":"Summarization Bias: The Directional Collapse of Objective Projection into Told-Mode Labels in Large Language Models --- A Conceptual Framework and Registered Test Protocol","year":2026,"arxiv_id":"2609.20712","external_url":"https://arxiv.org/abs/2609.20712","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-20712"},{"ref":"SRC-0682","id":"arxiv-2609-05450","kind":"paper","title":"SureRoute: Toward a Hallucination-Free Self-Improving Platform for Retrosynthesis","year":2026,"arxiv_id":"2609.05450","external_url":"https://arxiv.org/abs/2609.05450","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05450"},{"ref":"SRC-0458","id":"arxiv-2608-30661","kind":"paper","title":"SwarmBench: Can Large Language Models Act as Agent Swarm Orchestrators?","year":2026,"arxiv_id":"2608.30661","external_url":"https://arxiv.org/abs/2608.30661","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30661"},{"ref":"SRC-0642","id":"arxiv-2608-26081","kind":"paper","title":"SwarmWorld: Stigmergic technological evolution in societies of language-model agents","year":2026,"arxiv_id":"2608.26081","external_url":"https://arxiv.org/abs/2608.26081","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26081"},{"ref":"SRC-0112","id":"arxiv-2608-23564","kind":"paper","title":"SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration?","year":2026,"arxiv_id":"2608.23564","external_url":"https://arxiv.org/abs/2608.23564","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-23564"},{"ref":"SRC-0067","id":"arxiv-2405-15793","kind":"paper","title":"SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering","year":2024,"arxiv_id":"2405.15793","external_url":"https://arxiv.org/abs/2405.15793","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2405-15793"},{"ref":"SRC-0048","id":"arxiv-2310-06770","kind":"paper","title":"SWE-bench: Can Language Models Resolve Real-World GitHub Issues?","year":2023,"arxiv_id":"2310.06770","external_url":"https://arxiv.org/abs/2310.06770","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2310-06770"},{"ref":"SRC-0560","id":"arxiv-2609-13841","kind":"paper","title":"Sweet Talkers: How Query Formulation Shapes Sycophancy in Romantic Relationship Advice","year":2026,"arxiv_id":"2609.13841","external_url":"https://arxiv.org/abs/2609.13841","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-13841"},{"ref":"SRC-0315","id":"arxiv-2505-07849","kind":"paper","title":"SweRank: Software Issue Localization with Code Ranking","year":2025,"arxiv_id":"2505.07849","external_url":"https://arxiv.org/abs/2505.07849","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2505-07849"},{"ref":"SRC-0131","id":"arxiv-2608-26511","kind":"paper","title":"Sycophancy Suppression Can Impair Rational Updating: Anti-Sycophancy Should Preserve the Ability to Update","year":2026,"arxiv_id":"2608.26511","external_url":"https://arxiv.org/abs/2608.26511","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26511"},{"ref":"SRC-0070","id":"arxiv-2406-10162","kind":"paper","title":"Sycophancy to Subterfuge: Investigating Reward-Tampering in Large Language Models","year":2024,"arxiv_id":"2406.10162","external_url":"https://arxiv.org/abs/2406.10162","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2406-10162"},{"ref":"SRC-0163","id":"arxiv-2608-31079","kind":"paper","title":"Sycophantic Agreement Transfers with Neutral Data via Contrastive Preference Optimization","year":2026,"arxiv_id":"2608.31079","external_url":"https://arxiv.org/abs/2608.31079","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-31079"},{"ref":"SRC-0245","id":"arxiv-2609-00184","kind":"paper","title":"Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs","year":2026,"arxiv_id":"2609.00184","external_url":"https://arxiv.org/abs/2609.00184","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00184"},{"ref":"SRC-0217","id":"arxiv-2609-07286","kind":"paper","title":"TAD: Token-Adaptive Contrastive Decoding with Confidence-Guided Gating for Hallucination Mitigation in Large Audio-Language Models","year":2026,"arxiv_id":"2609.07286","external_url":"https://arxiv.org/abs/2609.07286","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07286"},{"ref":"SRC-0676","id":"arxiv-2609-01542","kind":"paper","title":"TAG-Bench: Benchmarking Temporal Audio Grounding in Large Audio Language Models","year":2026,"arxiv_id":"2609.01542","external_url":"https://arxiv.org/abs/2609.01542","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01542"},{"ref":"SRC-0120","id":"arxiv-2608-24966","kind":"paper","title":"Targeting the Attention Heads Behind Object Hallucination in LLaVA","year":2026,"arxiv_id":"2608.24966","external_url":"https://arxiv.org/abs/2608.24966","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-24966"},{"ref":"SRC-0392","id":"arxiv-2609-13005","kind":"paper","title":"Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models","year":2026,"arxiv_id":"2609.13005","external_url":"https://arxiv.org/abs/2609.13005","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-13005"},{"ref":"SRC-0654","id":"arxiv-2608-27871","kind":"paper","title":"Temporal Tree of Thought: Reasoning-Guided Visual Cue Search for Long-Video Understanding","year":2026,"arxiv_id":"2608.27871","external_url":"https://arxiv.org/abs/2608.27871","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27871"},{"ref":"SRC-0614","id":"arxiv-2608-11474","kind":"paper","title":"Test-Time Hallucination Control in Large Vision-Language Models","year":2026,"arxiv_id":"2608.11474","external_url":"https://arxiv.org/abs/2608.11474","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-11474"},{"ref":"SRC-0562","id":"arxiv-2609-16574","kind":"paper","title":"Testing Our Foundations: Citation Trends, Errors, and Emerging Hallucinations in the Computing Education Literature","year":2026,"arxiv_id":"2609.16574","external_url":"https://arxiv.org/abs/2609.16574","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-16574"},{"ref":"SRC-0192","id":"cnn-2025-hinton-maternal-instincts","kind":"post","title":"The 'godfather of AI' reveals the only way humanity can survive superintelligent AI","year":2025,"arxiv_id":null,"external_url":"https://www.cnn.com/2025/08/13/tech/ai-geoffrey-hinton","counts":{"claims":1},"url":"https://rsiratchet.com/sources/cnn-2025-hinton-maternal-instincts"},{"ref":"SRC-0338","id":"arxiv-2512-12488","kind":"paper","title":"The American Ghost in the Machine: How language models align culturally and the effects of cultural prompting","year":2025,"arxiv_id":"2512.12488","external_url":"https://arxiv.org/abs/2512.12488","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2512-12488"},{"ref":"SRC-0136","id":"arxiv-2608-27750","kind":"paper","title":"The Calls are Coming from Inside the Model: Investigating Probe-based Detection of Tool-Calling Errors in LLMs","year":2026,"arxiv_id":"2608.27750","external_url":"https://arxiv.org/abs/2608.27750","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-27750"},{"ref":"SRC-0578","id":"arxiv-2609-24885","kind":"paper","title":"The Copy Ceiling: An Input-Exposure Control for Ontology-Grounded Generation over Curated Corpora","year":2026,"arxiv_id":"2609.24885","external_url":"https://arxiv.org/abs/2609.24885","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-24885"},{"ref":"SRC-0389","id":"arxiv-2609-12111","kind":"paper","title":"The Cost of Compression: A Rate-Distortion Limit on Factual Hallucination","year":2026,"arxiv_id":"2609.12111","external_url":"https://arxiv.org/abs/2609.12111","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-12111"},{"ref":"SRC-0290","id":"arxiv-2609-14098","kind":"paper","title":"The Deception Delta: Adversarial Evaluation of LLM-Based Smart Contract Bytecode Forensics","year":2026,"arxiv_id":"2609.14098","external_url":"https://arxiv.org/abs/2609.14098","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-14098"},{"ref":"SRC-0193","id":"elasticity-2026-economics-of-rsi","kind":"paper","title":"The Economics of Recursive Self-Improvement","year":2026,"arxiv_id":null,"external_url":"https://elasticity.institute/rsi-paper.pdf","counts":{"claims":2},"url":"https://rsiratchet.com/sources/elasticity-2026-economics-of-rsi"},{"ref":"SRC-0066","id":"arxiv-2404-13208","kind":"paper","title":"The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions","year":2024,"arxiv_id":"2404.13208","external_url":"https://arxiv.org/abs/2404.13208","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2404-13208"},{"ref":"SRC-0188","id":"arxiv-2609-11873","kind":"paper","title":"The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement","year":2026,"arxiv_id":"2609.11873","external_url":"https://arxiv.org/abs/2609.11873","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-11873"},{"ref":"SRC-0001","id":"anthropic-2026-context-engineering-claude-5","kind":"vendor-doc","title":"The new rules of context engineering for Claude 5 generation models","year":2026,"arxiv_id":null,"external_url":"https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models","counts":{"claims":3},"url":"https://rsiratchet.com/sources/anthropic-2026-context-engineering-claude-5"},{"ref":"SRC-0003","id":"arxiv-1611-08219","kind":"paper","title":"The Off-Switch Game","year":2016,"arxiv_id":"1611.08219","external_url":"https://arxiv.org/abs/1611.08219","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-1611-08219"},{"ref":"SRC-0167","id":"arxiv-2609-00492","kind":"paper","title":"The Privacy-Hallucination Tradeoff in Differentially Private Language Models","year":2026,"arxiv_id":"2609.00492","external_url":"https://arxiv.org/abs/2609.00492","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00492"},{"ref":"SRC-0044","id":"arxiv-2309-12288","kind":"paper","title":"The Reversal Curse: LLMs trained on \"A is B\" fail to learn \"B is A\"","year":2023,"arxiv_id":"2309.12288","external_url":"https://arxiv.org/abs/2309.12288","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2309-12288"},{"ref":"SRC-0404","id":"arxiv-2609-16705","kind":"paper","title":"The Robot Data Factory","year":2026,"arxiv_id":"2609.16705","external_url":"https://arxiv.org/abs/2609.16705","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-16705"},{"ref":"SRC-0490","id":"arxiv-2609-15314","kind":"paper","title":"The Universe of Universes: Benefit Yield Functions, Implosion Thresholds, and Infrastructure-Aware Optimization in Multi-LLM Systems","year":2026,"arxiv_id":"2609.15314","external_url":"https://arxiv.org/abs/2609.15314","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15314"},{"ref":"SRC-0264","id":"arxiv-2609-14528","kind":"paper","title":"Theseus in the Graph: Towards Traceable Multi-Hop Graph Navigation","year":2026,"arxiv_id":"2609.14528","external_url":"https://arxiv.org/abs/2609.14528","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-14528"},{"ref":"SRC-0147","id":"arxiv-2608-29374","kind":"paper","title":"Think, Look, and Revise: Inconsistency-Aware Visual Self-Correction in MLLMs","year":2026,"arxiv_id":"2608.29374","external_url":"https://arxiv.org/abs/2608.29374","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29374"},{"ref":"SRC-0579","id":"arxiv-2610-11765","kind":"paper","title":"Thinking Inertia: LLMs Keep Thinking When Told Not To","year":2026,"arxiv_id":"2610.11765","external_url":"https://arxiv.org/abs/2610.11765","counts":{"claims":3},"url":"https://rsiratchet.com/sources/arxiv-2610-11765"},{"ref":"SRC-0187","id":"arxiv-2609-11801","kind":"paper","title":"Thinking with Looped Flows","year":2026,"arxiv_id":"2609.11801","external_url":"https://arxiv.org/abs/2609.11801","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-11801"},{"ref":"SRC-0073","id":"arxiv-2410-03608","kind":"paper","title":"TICKing All the Boxes: Generated Checklists Improve LLM Evaluation and Generation","year":2024,"arxiv_id":"2410.03608","external_url":"https://arxiv.org/abs/2410.03608","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2410-03608"},{"ref":"SRC-0533","id":"arxiv-2609-24677","kind":"paper","title":"TimeLitmus: A Diagnostic Benchmark for Cross-Modal Understanding and Explanation Faithfulness in Event-Conditioned Time-Series Prediction","year":2026,"arxiv_id":"2609.24677","external_url":"https://arxiv.org/abs/2609.24677","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-24677"},{"ref":"SRC-0589","id":"arxiv-2608-03391","kind":"paper","title":"TimeRLM: Recursive Language Models Enable Precise Anomaly Localization in Long-Context Time-Series","year":2026,"arxiv_id":"2608.03391","external_url":"https://arxiv.org/abs/2608.03391","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-03391"},{"ref":"SRC-0261","id":"arxiv-2609-13457","kind":"paper","title":"TimeThink: Eliciting Compositional Reasoning in Timeseries Large Language Models","year":2026,"arxiv_id":"2609.13457","external_url":"https://arxiv.org/abs/2609.13457","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-13457"},{"ref":"SRC-0496","id":"arxiv-2609-19167","kind":"paper","title":"To Memories and Beyond: From Remembering to Knowing You across Long-Term Multimodal Personal Archives","year":2026,"arxiv_id":"2609.19167","external_url":"https://arxiv.org/abs/2609.19167","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-19167"},{"ref":"SRC-0059","id":"arxiv-2402-14903","kind":"paper","title":"Tokenization counts: the impact of tokenization on arithmetic in frontier LLMs","year":2024,"arxiv_id":"2402.14903","external_url":"https://arxiv.org/abs/2402.14903","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2402-14903"},{"ref":"SRC-0021","id":"arxiv-2302-04761","kind":"paper","title":"Toolformer: Language Models Can Teach Themselves to Use Tools","year":2023,"arxiv_id":"2302.04761","external_url":"https://arxiv.org/abs/2302.04761","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2302-04761"},{"ref":"SRC-0039","id":"arxiv-2307-16789","kind":"paper","title":"ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs","year":2023,"arxiv_id":"2307.16789","external_url":"https://arxiv.org/abs/2307.16789","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2307-16789"},{"ref":"SRC-0438","id":"arxiv-2608-23635","kind":"paper","title":"ToolRobustBench: Stage-Wise Perturbation Evaluation and Failure Diagnosis for Tool-Calling Agents","year":2026,"arxiv_id":"2608.23635","external_url":"https://arxiv.org/abs/2608.23635","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-23635"},{"ref":"SRC-0365","id":"arxiv-2608-11794","kind":"paper","title":"Toward Meaningful Transparency for AI Chatbots: Disclosing Persuasive Intent Reduces Persuasion","year":2026,"arxiv_id":"2608.11794","external_url":"https://arxiv.org/abs/2608.11794","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-11794"},{"ref":"SRC-0559","id":"arxiv-2609-13436","kind":"paper","title":"Toward Self-Adaptive Physical AI: Can LLM Agents Manage Long-Horizon Physical Tasks?","year":2026,"arxiv_id":"2609.13436","external_url":"https://arxiv.org/abs/2609.13436","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-13436"},{"ref":"SRC-0166","id":"arxiv-2609-00455","kind":"paper","title":"Towards a Belief-Based World Model for LLM Agents","year":2026,"arxiv_id":"2609.00455","external_url":"https://arxiv.org/abs/2609.00455","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00455"},{"ref":"SRC-0555","id":"arxiv-2609-10728","kind":"paper","title":"Towards a Deterministic Math Solver for Clinical Language Models","year":2026,"arxiv_id":"2609.10728","external_url":"https://arxiv.org/abs/2609.10728","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-10728"},{"ref":"SRC-0222","id":"arxiv-2609-08402","kind":"paper","title":"Towards Embodied Air-Ground Cooperative Object Search: Benchmark, Dataset and Agentic Method","year":2026,"arxiv_id":"2609.08402","external_url":"https://arxiv.org/abs/2609.08402","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-08402"},{"ref":"SRC-0142","id":"arxiv-2608-29016","kind":"paper","title":"Towards Fully Automated Medical Imaging Code Generation via Validation-based Context Engineering","year":2026,"arxiv_id":"2608.29016","external_url":"https://arxiv.org/abs/2608.29016","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-29016"},{"ref":"SRC-0468","id":"arxiv-2609-00845","kind":"paper","title":"Towards Generalizable Visually Grounded Exploration of Household Devices","year":2026,"arxiv_id":"2609.00845","external_url":"https://arxiv.org/abs/2609.00845","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00845"},{"ref":"SRC-0515","id":"arxiv-2609-22844","kind":"paper","title":"Towards the Generalizability of Leveraging ChatGPT in APR via Self-enhancing: An Empirical Study","year":2026,"arxiv_id":"2609.22844","external_url":"https://arxiv.org/abs/2609.22844","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22844"},{"ref":"SRC-0051","id":"arxiv-2310-13548","kind":"paper","title":"Towards Understanding Sycophancy in Language Models","year":2023,"arxiv_id":"2310.13548","external_url":"https://arxiv.org/abs/2310.13548","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2310-13548"},{"ref":"SRC-0541","id":"arxiv-2608-26086","kind":"paper","title":"TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning Development","year":2026,"arxiv_id":"2608.26086","external_url":"https://arxiv.org/abs/2608.26086","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26086"},{"ref":"SRC-0009","id":"arxiv-2110-14168","kind":"paper","title":"Training Verifiers to Solve Math Word Problems","year":2021,"arxiv_id":"2110.14168","external_url":"https://arxiv.org/abs/2110.14168","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2110-14168"},{"ref":"SRC-0357","id":"arxiv-2605-10843","kind":"paper","title":"Training-Free Cultural Alignment of Large Language Models via Persona Disagreement","year":2026,"arxiv_id":"2605.10843","external_url":"https://arxiv.org/abs/2605.10843","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2605-10843"},{"ref":"SRC-0423","id":"arxiv-2608-18884","kind":"paper","title":"Training-Free Inference-Time Self-Reflection and Cost-Bounded Early Stopping for Large Language Models","year":2026,"arxiv_id":"2608.18884","external_url":"https://arxiv.org/abs/2608.18884","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2608-18884"},{"ref":"SRC-0068","id":"arxiv-2405-17399","kind":"paper","title":"Transformers Can Do Arithmetic with the Right Embeddings","year":2024,"arxiv_id":"2405.17399","external_url":"https://arxiv.org/abs/2405.17399","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2405-17399"},{"ref":"SRC-0489","id":"arxiv-2609-15079","kind":"paper","title":"Translating the Translator: Decomposing the Cost of English-Forced Inter-Agent Communication","year":2026,"arxiv_id":"2609.15079","external_url":"https://arxiv.org/abs/2609.15079","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15079"},{"ref":"SRC-0513","id":"arxiv-2609-22712","kind":"paper","title":"Trustworthy Agentic AI: Failure Modes, Mitigation Strategies, and a Lifecycle Framework for Autonomous LLM Systems","year":2026,"arxiv_id":"2609.22712","external_url":"https://arxiv.org/abs/2609.22712","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22712"},{"ref":"SRC-0008","id":"arxiv-2109-07958","kind":"paper","title":"TruthfulQA: Measuring How Models Mimic Human Falsehoods","year":2021,"arxiv_id":"2109.07958","external_url":"https://arxiv.org/abs/2109.07958","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2109-07958"},{"ref":"SRC-0595","id":"arxiv-2608-05616","kind":"paper","title":"TruthLens: Object Hallucination Detection via Self-Evaluating Truthfulness Scores in LVLMs","year":2026,"arxiv_id":"2608.05616","external_url":"https://arxiv.org/abs/2608.05616","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2608-05616"},{"ref":"SRC-0656","id":"arxiv-2608-28018","kind":"paper","title":"Twin Worlds: Equivariance-Based Abstention for Evidence-Grounded Reasoning","year":2026,"arxiv_id":"2608.28018","external_url":"https://arxiv.org/abs/2608.28018","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-28018"},{"ref":"SRC-0538","id":"arxiv-2511-05040","kind":"paper","title":"UA-Code-Bench: A Competitive Programming Benchmark for Evaluating LLM Code Generation in Ukrainian","year":2025,"arxiv_id":"2511.05040","external_url":"https://arxiv.org/abs/2511.05040","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2511-05040"},{"ref":"SRC-0403","id":"arxiv-2609-16368","kind":"paper","title":"UDAV: Uncertainty-Driven Adaptive VLM Waypoint Planner","year":2026,"arxiv_id":"2609.16368","external_url":"https://arxiv.org/abs/2609.16368","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-16368"},{"ref":"SRC-0590","id":"arxiv-2608-03817","kind":"paper","title":"UHP Detection: LVLMs have their Unique Hallucination Pattern in the Consistency Space","year":2026,"arxiv_id":"2608.03817","external_url":"https://arxiv.org/abs/2608.03817","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-03817"},{"ref":"SRC-0684","id":"arxiv-2609-07395","kind":"paper","title":"Uncertainty Quantification for LLM Agents: A Taxonomy, an Evaluation Protocol, and an Empirical Study","year":2026,"arxiv_id":"2609.07395","external_url":"https://arxiv.org/abs/2609.07395","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-07395"},{"ref":"SRC-0673","id":"arxiv-2609-00213","kind":"paper","title":"Uncovering and Mitigating Aggregation-Induced Reward Hacking in Multi-Reward Reinforcement Learning","year":2026,"arxiv_id":"2609.00213","external_url":"https://arxiv.org/abs/2609.00213","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00213"},{"ref":"SRC-0339","id":"arxiv-2512-12620","kind":"paper","title":"Understanding Syllogistic Reasoning in LLMs from Formal and Natural Language Perspectives","year":2025,"arxiv_id":"2512.12620","external_url":"https://arxiv.org/abs/2512.12620","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2512-12620"},{"ref":"SRC-0612","id":"arxiv-2608-10835","kind":"paper","title":"UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations","year":2026,"arxiv_id":"2608.10835","external_url":"https://arxiv.org/abs/2608.10835","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-10835"},{"ref":"SRC-0474","id":"arxiv-2609-05910","kind":"paper","title":"UniRRM: Unified Reasoning Reward Models Across Languages and Evaluation Paradigms","year":2026,"arxiv_id":"2609.05910","external_url":"https://arxiv.org/abs/2609.05910","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-05910"},{"ref":"SRC-0296","id":"arxiv-2609-16098","kind":"paper","title":"Universal Defenses for Tool-Integrated LLM Agents Against Adversarial Attacks","year":2026,"arxiv_id":"2609.16098","external_url":"https://arxiv.org/abs/2609.16098","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-16098"},{"ref":"SRC-0318","id":"arxiv-2605-08765","kind":"paper","title":"Unlearners Can Lie: Evaluating and Improving Honesty in LLM Unlearning","year":2026,"arxiv_id":"2605.08765","external_url":"https://arxiv.org/abs/2605.08765","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2605-08765"},{"ref":"SRC-0132","id":"arxiv-2608-26588","kind":"paper","title":"Unsaid, Unsafe? Implicit Security Obligations in LLM-Based RTL Code Generation","year":2026,"arxiv_id":"2608.26588","external_url":"https://arxiv.org/abs/2608.26588","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26588"},{"ref":"SRC-0586","id":"arxiv-2608-03018","kind":"paper","title":"UrbanAgent: A Tool-Augmented Agent for Cross-System Urban Tasks","year":2026,"arxiv_id":"2608.03018","external_url":"https://arxiv.org/abs/2608.03018","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-03018"},{"ref":"SRC-0547","id":"arxiv-2609-02421","kind":"paper","title":"UTP-Bench: Uncertainty-aware Travel Planning Benchmark","year":2026,"arxiv_id":"2609.02421","external_url":"https://arxiv.org/abs/2609.02421","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-02421"},{"ref":"SRC-0319","id":"arxiv-2606-16262","kind":"paper","title":"UXBench: Measuring the Actionability of LLM-Generated UX Critiques","year":2026,"arxiv_id":"2606.16262","external_url":"https://arxiv.org/abs/2606.16262","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-16262"},{"ref":"SRC-0617","id":"arxiv-2608-12282","kind":"paper","title":"VAKRA: Evaluating Multi-Hop Reasoning Across APIs and Retrieval Under Tool-Use Policies","year":2026,"arxiv_id":"2608.12282","external_url":"https://arxiv.org/abs/2608.12282","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-12282"},{"ref":"SRC-0511","id":"arxiv-2609-22293","kind":"paper","title":"Validating, Not Sampling: Region-Level Robustness of Vision-Language and Vision-Language-Action Models","year":2026,"arxiv_id":"2609.22293","external_url":"https://arxiv.org/abs/2609.22293","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22293"},{"ref":"SRC-0651","id":"arxiv-2608-27311","kind":"paper","title":"Verify Smarter, Evolve Further: Efficient Harness Evolution through Behavior-Aware Verification","year":2026,"arxiv_id":"2608.27311","external_url":"https://arxiv.org/abs/2608.27311","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27311"},{"ref":"SRC-0171","id":"arxiv-2609-00921","kind":"paper","title":"VIBE-Bench: Evaluating Personalized Large Language Models When Profiles Don't Mean Preferences","year":2026,"arxiv_id":"2609.00921","external_url":"https://arxiv.org/abs/2609.00921","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00921"},{"ref":"SRC-0301","id":"arxiv-2609-16647","kind":"paper","title":"ViD: Vision-Dominant Gender Bias Mitigation for Large Vision-Language Models","year":2026,"arxiv_id":"2609.16647","external_url":"https://arxiv.org/abs/2609.16647","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-16647"},{"ref":"SRC-0302","id":"arxiv-2609-17248","kind":"paper","title":"Video-HolmesV2: Can MLLMs Reason with Spatio-Temporal Audio-Visual Evidence in Long Videos?","year":2026,"arxiv_id":"2609.17248","external_url":"https://arxiv.org/abs/2609.17248","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17248"},{"ref":"SRC-0122","id":"arxiv-2608-25529","kind":"paper","title":"Video-IFBench: Evaluating Instruction Following of Multimodal LLMs in Video Understanding Scenarios","year":2026,"arxiv_id":"2608.25529","external_url":"https://arxiv.org/abs/2608.25529","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25529"},{"ref":"SRC-0650","id":"arxiv-2608-27065","kind":"paper","title":"Video-OPSD: Exploiting Privileged Visual Evidence for On-Policy Self-Distillation in Video Large Language Models","year":2026,"arxiv_id":"2608.27065","external_url":"https://arxiv.org/abs/2608.27065","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27065"},{"ref":"SRC-0631","id":"arxiv-2608-24302","kind":"paper","title":"VideoHarness-RSI: Recursive Harness Self-Improvement for Long-Video Understanding with Frozen Vision-Language Models","year":2026,"arxiv_id":"2608.24302","external_url":"https://arxiv.org/abs/2608.24302","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24302"},{"ref":"SRC-0405","id":"arxiv-2609-16722","kind":"paper","title":"VideoMM: Adaptive Macro-Micro Inference for Efficient Video MLLMs","year":2026,"arxiv_id":"2609.16722","external_url":"https://arxiv.org/abs/2609.16722","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-16722"},{"ref":"SRC-0271","id":"arxiv-2609-15606","kind":"paper","title":"VideoScout: Learning Agentic Active Exploration with Adaptive Reasoning Pacing for Long Video Understanding","year":2026,"arxiv_id":"2609.15606","external_url":"https://arxiv.org/abs/2609.15606","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15606"},{"ref":"SRC-0506","id":"arxiv-2609-21521","kind":"paper","title":"VidOmni-Bench: A Benchmark for Fine-Grained Video Understanding via Spatio-Temporal Event Verification across Complexity and Duration","year":2026,"arxiv_id":"2609.21521","external_url":"https://arxiv.org/abs/2609.21521","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-21521"},{"ref":"SRC-0128","id":"arxiv-2608-26013","kind":"paper","title":"VISA: Agentic Self-Evolving Data Synthesis for Multimodal Instruction Following","year":2026,"arxiv_id":"2608.26013","external_url":"https://arxiv.org/abs/2608.26013","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-26013"},{"ref":"SRC-0157","id":"arxiv-2608-30480","kind":"paper","title":"VisER: Visual Evidence and Reliance for Object Hallucination Detection in LVLMs","year":2026,"arxiv_id":"2608.30480","external_url":"https://arxiv.org/abs/2608.30480","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-30480"},{"ref":"SRC-0246","id":"arxiv-2609-00505","kind":"paper","title":"ViTAL-X: Video-Text Alignment with Cross-Modal Temporal Edits","year":2026,"arxiv_id":"2609.00505","external_url":"https://arxiv.org/abs/2609.00505","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-00505"},{"ref":"SRC-0204","id":"arxiv-2609-04355","kind":"paper","title":"VLA-Precision: Asymmetric Co-Bootstrapping for Efficient Real-World Online RL of Vision-Language-Action Models","year":2026,"arxiv_id":"2609.04355","external_url":"https://arxiv.org/abs/2609.04355","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-04355"},{"ref":"SRC-0465","id":"arxiv-2609-00570","kind":"paper","title":"VoiceLongMemEval: Do Assistants Remember How You Sounded?","year":2026,"arxiv_id":"2609.00570","external_url":"https://arxiv.org/abs/2609.00570","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-00570"},{"ref":"SRC-0512","id":"arxiv-2609-22452","kind":"paper","title":"Vox-Infinity: Benchmarking the Limits of Long-Context Spoken Language Models","year":2026,"arxiv_id":"2609.22452","external_url":"https://arxiv.org/abs/2609.22452","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-22452"},{"ref":"SRC-0606","id":"arxiv-2608-08569","kind":"paper","title":"VoxZip: Semantic-Anchored Temporal KV Cache Compression for Long-Context Audio Inference","year":2026,"arxiv_id":"2608.08569","external_url":"https://arxiv.org/abs/2608.08569","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-08569"},{"ref":"SRC-0030","id":"arxiv-2305-16291","kind":"paper","title":"Voyager: An Open-Ended Embodied Agent with Large Language Models","year":2023,"arxiv_id":"2305.16291","external_url":"https://arxiv.org/abs/2305.16291","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2305-16291"},{"ref":"SRC-0071","id":"arxiv-2406-10279","kind":"paper","title":"We Have a Package for You! A Comprehensive Analysis of Package Hallucinations by Code Generating LLMs","year":2024,"arxiv_id":"2406.10279","external_url":"https://arxiv.org/abs/2406.10279","counts":{"claims":3},"url":"https://rsiratchet.com/sources/arxiv-2406-10279"},{"ref":"SRC-0055","id":"arxiv-2312-09390","kind":"paper","title":"Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision","year":2023,"arxiv_id":"2312.09390","external_url":"https://arxiv.org/abs/2312.09390","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2312-09390"},{"ref":"SRC-0037","id":"arxiv-2307-13854","kind":"paper","title":"WebArena: A Realistic Web Environment for Building Autonomous Agents","year":2023,"arxiv_id":"2307.13854","external_url":"https://arxiv.org/abs/2307.13854","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2307-13854"},{"ref":"SRC-0010","id":"arxiv-2112-09332","kind":"paper","title":"WebGPT: Browser-assisted question-answering with human feedback","year":2021,"arxiv_id":"2112.09332","external_url":"https://arxiv.org/abs/2112.09332","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2112-09332"},{"ref":"SRC-0408","id":"arxiv-2609-17515","kind":"paper","title":"What Breaks Under Pruning in Smart Homes, and When? Evaluating LLM Degradation Across Architectures and Task Complexity","year":2026,"arxiv_id":"2609.17515","external_url":"https://arxiv.org/abs/2609.17515","counts":{"claims":0},"url":"https://rsiratchet.com/sources/arxiv-2609-17515"},{"ref":"SRC-0551","id":"arxiv-2609-03718","kind":"paper","title":"What Do CAE Simulation Agents Really Need Beyond a Generic Harness?","year":2026,"arxiv_id":"2609.03718","external_url":"https://arxiv.org/abs/2609.03718","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-03718"},{"ref":"SRC-0300","id":"arxiv-2609-16646","kind":"paper","title":"What Do Hallucinations Reveal About Multimodal Reasoning? Diagnosing Visual Grounding Failures via Contrastive Decoding Probes","year":2026,"arxiv_id":"2609.16646","external_url":"https://arxiv.org/abs/2609.16646","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-16646"},{"ref":"SRC-0114","id":"arxiv-2608-24022","kind":"paper","title":"What Guides the Agent? Adjudicating Unauthorized Behavior via Localizing Behavior-Guiding Instructions","year":2026,"arxiv_id":"2608.24022","external_url":"https://arxiv.org/abs/2608.24022","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24022"},{"ref":"SRC-0232","id":"arxiv-2609-09696","kind":"paper","title":"When Auditors Fabricate: Batch-Size Degradation and Confident Hallucination in LLM Detection of Planted Document Contamination","year":2026,"arxiv_id":"2609.09696","external_url":"https://arxiv.org/abs/2609.09696","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-09696"},{"ref":"SRC-0594","id":"arxiv-2608-05207","kind":"paper","title":"When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters","year":2026,"arxiv_id":"2608.05207","external_url":"https://arxiv.org/abs/2608.05207","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-05207"},{"ref":"SRC-0118","id":"arxiv-2608-24492","kind":"paper","title":"When Do Supervised UQ Ensembles Improve LLM Hallucination Detection? A Robustness Study","year":2026,"arxiv_id":"2608.24492","external_url":"https://arxiv.org/abs/2608.24492","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-24492"},{"ref":"SRC-0655","id":"arxiv-2608-27984","kind":"paper","title":"When Evidence Shapes Collaboration: Knowledge-Conditioned Topology Generation for Multi-Agent Systems","year":2026,"arxiv_id":"2608.27984","external_url":"https://arxiv.org/abs/2608.27984","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-27984"},{"ref":"SRC-0675","id":"arxiv-2609-01519","kind":"paper","title":"When Guardrails Look Effective: Construct Validity Failures in LLM Agent Commerce Evaluation","year":2026,"arxiv_id":"2609.01519","external_url":"https://arxiv.org/abs/2609.01519","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01519"},{"ref":"SRC-0658","id":"arxiv-2608-28382","kind":"paper","title":"When Linguistic and Internal Confidence Diverge in Large Language Models","year":2026,"arxiv_id":"2608.28382","external_url":"https://arxiv.org/abs/2608.28382","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-28382"},{"ref":"SRC-0493","id":"arxiv-2609-17516","kind":"paper","title":"When Should LLMs Abstain? Chain-of-Self-Questioning for Selective Risk Control","year":2026,"arxiv_id":"2609.17516","external_url":"https://arxiv.org/abs/2609.17516","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-17516"},{"ref":"SRC-0615","id":"arxiv-2608-11715","kind":"paper","title":"When the API Speaks the Wrong Language: Revisiting Post-Training for Multilingual Tool Use","year":2026,"arxiv_id":"2608.11715","external_url":"https://arxiv.org/abs/2608.11715","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-11715"},{"ref":"SRC-0269","id":"arxiv-2609-15106","kind":"paper","title":"When the Wrong Key Wins: Understanding and Detecting Hallucinations in LLMs","year":2026,"arxiv_id":"2609.15106","external_url":"https://arxiv.org/abs/2609.15106","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-15106"},{"ref":"SRC-0360","id":"arxiv-2606-07877","kind":"paper","title":"Whose Norms? Disentangling Cultural and Personal Alignment in Large Language Models","year":2026,"arxiv_id":"2606.07877","external_url":"https://arxiv.org/abs/2606.07877","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2606-07877"},{"ref":"SRC-0605","id":"arxiv-2608-08167","kind":"paper","title":"Wiener Representation Filtering for VLM Hallucination Suppression","year":2026,"arxiv_id":"2608.08167","external_url":"https://arxiv.org/abs/2608.08167","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-08167"},{"ref":"SRC-0168","id":"arxiv-2609-00568","kind":"paper","title":"WiseSpec: Requirements-Driven Agents for Code Generation","year":2026,"arxiv_id":"2609.00568","external_url":"https://arxiv.org/abs/2609.00568","counts":{"claims":2},"url":"https://rsiratchet.com/sources/arxiv-2609-00568"},{"ref":"SRC-0102","id":"arxiv-2608-19661","kind":"paper","title":"World-Model-Grounded LLM Planning for AUV and ASV Navigation Near Offshore Wind Farms","year":2026,"arxiv_id":"2608.19661","external_url":"https://arxiv.org/abs/2608.19661","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-19661"},{"ref":"SRC-0674","id":"arxiv-2609-01056","kind":"paper","title":"WorldBench: Culturally Grounded Benchmark for Multilingual Agents","year":2026,"arxiv_id":"2609.01056","external_url":"https://arxiv.org/abs/2609.01056","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-01056"},{"ref":"SRC-0552","id":"arxiv-2609-03952","kind":"paper","title":"WorldReward: Reward Modeling for Camera-Conditioned World Models","year":2026,"arxiv_id":"2609.03952","external_url":"https://arxiv.org/abs/2609.03952","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2609-03952"},{"ref":"SRC-0255","id":"kushmerick-2000-wrapper-induction","kind":"paper","title":"Wrapper Induction: Efficiency and Expressiveness","year":2000,"arxiv_id":null,"external_url":"https://aaai.org/papers/082-ws98-14-012/","counts":{"claims":1},"url":"https://rsiratchet.com/sources/kushmerick-2000-wrapper-induction"},{"ref":"SRC-0253","id":"arxiv-2502-15688","kind":"paper","title":"XPath Agent: An Efficient XPath Programming Agent Based on LLM for Web Crawler","year":2024,"arxiv_id":"2502.15688","external_url":"https://arxiv.org/abs/2502.15688","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2502-15688"},{"ref":"SRC-0126","id":"arxiv-2608-25939","kind":"paper","title":"XREPOTEST: Benchmarking Multilingual Repository-Level Unit Test Generation for Large Language Models","year":2026,"arxiv_id":"2608.25939","external_url":"https://arxiv.org/abs/2608.25939","counts":{"claims":1},"url":"https://rsiratchet.com/sources/arxiv-2608-25939"},{"ref":"SRC-0072","id":"arxiv-2406-12045","kind":"paper","title":"τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains","year":2024,"arxiv_id":"2406.12045","external_url":"https://arxiv.org/abs/2406.12045","counts":{"claims":3},"url":"https://rsiratchet.com/sources/arxiv-2406-12045"}],"heuristics":[{"ref":"ADG-0001","id":"amdahls-law","label":"Amdahl's law","statement":"The speed-up from parallelising a task is bounded by the fraction of it that must stay serial.","provenance":"borrowed","standing":"untested","counts":{"evidence":0},"url":"https://rsiratchet.com/heuristics/amdahls-law"},{"ref":"ADG-0002","id":"ashbys-law-of-experience","label":"Ashby's law of experience","statement":"Identical deterministic machines given the same inputs converge on the same state; variety does not grow.","provenance":"borrowed","standing":"untested","counts":{"evidence":1},"url":"https://rsiratchet.com/heuristics/ashbys-law-of-experience"},{"ref":"ADG-0003","id":"ashbys-law-of-requisite-variety","label":"Ashby's law of requisite variety","statement":"A regulator can only control a system if it has at least as many distinguishable responses as the disturbances it must handle.","provenance":"borrowed","standing":"holds","counts":{"evidence":1},"url":"https://rsiratchet.com/heuristics/ashbys-law-of-requisite-variety"},{"ref":"ADG-0004","id":"boyds-law-of-iteration","label":"Boyd's law of iteration","statement":"Speed of iteration beats quality of iteration.","provenance":"borrowed","standing":"narrowed","counts":{"evidence":3},"url":"https://rsiratchet.com/heuristics/boyds-law-of-iteration"},{"ref":"ADG-0005","id":"braesss-paradox","label":"Braess's paradox","statement":"Adding capacity to a network can make its overall performance worse.","provenance":"borrowed","standing":"untested","counts":{"evidence":3},"url":"https://rsiratchet.com/heuristics/braesss-paradox"},{"ref":"ADG-0006","id":"brandolinis-law","label":"Brandolini's law","statement":"The amount of energy needed to refute nonsense is an order of magnitude bigger than that needed to produce it.","provenance":"borrowed","standing":"holds","counts":{"evidence":3},"url":"https://rsiratchet.com/heuristics/brandolinis-law"},{"ref":"ADG-0007","id":"conways-law","label":"Conway's law","statement":"Organisations design systems that mirror their own communication structure.","provenance":"borrowed","standing":"untested","counts":{"evidence":0},"url":"https://rsiratchet.com/heuristics/conways-law"},{"ref":"ADG-0008","id":"first-impressions-last","label":"First impressions last","statement":"What is encountered first weighs more than what comes after.","provenance":"borrowed","standing":"narrowed","counts":{"evidence":2},"url":"https://rsiratchet.com/heuristics/first-impressions-last"},{"ref":"ADG-0009","id":"good-regulator-theorem","label":"Good regulator theorem","statement":"Every good regulator of a system must be a model of that system.","provenance":"borrowed","standing":"holds","counts":{"evidence":2},"url":"https://rsiratchet.com/heuristics/good-regulator-theorem"},{"ref":"ADG-0010","id":"goodharts-law","label":"Goodhart's law","statement":"When a measure becomes a target, it ceases to be a good measure.","provenance":"borrowed","standing":"holds","counts":{"evidence":5},"url":"https://rsiratchet.com/heuristics/goodharts-law"},{"ref":"ADG-0011","id":"hanlons-razor","label":"Hanlon's razor","statement":"Never attribute to malice that which is adequately explained by incompetence.","provenance":"borrowed","standing":"narrowed","counts":{"evidence":2},"url":"https://rsiratchet.com/heuristics/hanlons-razor"},{"ref":"ADG-0012","id":"hashimotos-rule","label":"Hashimoto's rule","statement":"Every mistake becomes a permanent fix.","provenance":"borrowed","standing":"untested","counts":{"evidence":4},"url":"https://rsiratchet.com/heuristics/hashimotos-rule"},{"ref":"ADG-0013","id":"humphreys-law","label":"Humphrey's law","statement":"Users do not know what they want until they see working software.","provenance":"borrowed","standing":"untested","counts":{"evidence":1},"url":"https://rsiratchet.com/heuristics/humphreys-law"},{"ref":"ADG-0014","id":"hyrums-law","label":"Hyrum's law","statement":"With enough users of an interface, every observable behaviour of the system will be depended on by somebody, whatever the contract promises.","provenance":"borrowed","standing":"untested","counts":{"evidence":0},"url":"https://rsiratchet.com/heuristics/hyrums-law"},{"ref":"ADG-0015","id":"kerckhoffss-principle","label":"Kerckhoffs's principle","statement":"A system should be secure even if everything about it except the key is public knowledge.","provenance":"borrowed","standing":"untested","counts":{"evidence":0},"url":"https://rsiratchet.com/heuristics/kerckhoffss-principle"},{"ref":"ADG-0016","id":"law-of-triviality","label":"Law of triviality","statement":"People give disproportionate weight to trivial issues because those are the ones everyone can grasp.","provenance":"borrowed","standing":"holds","counts":{"evidence":3},"url":"https://rsiratchet.com/heuristics/law-of-triviality"},{"ref":"ADG-0017","id":"lehmans-law-of-increasing-complexity","label":"Lehman's law of increasing complexity","statement":"As a system evolves its complexity increases unless work is done to maintain or reduce it.","provenance":"borrowed","standing":"holds","counts":{"evidence":3},"url":"https://rsiratchet.com/heuristics/lehmans-law-of-increasing-complexity"},{"ref":"ADG-0018","id":"linuss-law","label":"Linus's law","statement":"Given enough eyeballs, all bugs are shallow.","provenance":"borrowed","standing":"narrowed","counts":{"evidence":2},"url":"https://rsiratchet.com/heuristics/linuss-law"},{"ref":"ADG-0019","id":"no-one-should-judge-their-own-cause","label":"No one should be judge in their own cause","statement":"A party with a stake in an outcome cannot be trusted to judge it.","provenance":"borrowed","standing":"holds","counts":{"evidence":4},"url":"https://rsiratchet.com/heuristics/no-one-should-judge-their-own-cause"},{"ref":"ADG-0020","id":"parkinsons-law","label":"Parkinson's law","statement":"Work expands so as to fill the time available for its completion.","provenance":"borrowed","standing":"untested","counts":{"evidence":1},"url":"https://rsiratchet.com/heuristics/parkinsons-law"},{"ref":"ADG-0021","id":"postels-law","label":"Postel's law","statement":"Be conservative in what you send, be liberal in what you accept.","provenance":"borrowed","standing":"untested","counts":{"evidence":3},"url":"https://rsiratchet.com/heuristics/postels-law"},{"ref":"ADG-0022","id":"practice-makes-perfect","label":"Practice makes perfect","statement":"Doing a thing repeatedly makes you better at it.","provenance":"borrowed","standing":"untested","counts":{"evidence":3},"url":"https://rsiratchet.com/heuristics/practice-makes-perfect"},{"ref":"ADG-0023","id":"simons-law","label":"Simon's law","statement":"The apparent complexity of a system's behaviour is largely a reflection of the complexity of the environment it is in.","provenance":"borrowed","standing":"untested","counts":{"evidence":2},"url":"https://rsiratchet.com/heuristics/simons-law"},{"ref":"ADG-0024","id":"thomas-theorem","label":"Thomas theorem","statement":"If people define situations as real, they are real in their consequences.","provenance":"borrowed","standing":"holds","counts":{"evidence":3},"url":"https://rsiratchet.com/heuristics/thomas-theorem"},{"ref":"ADG-0025","id":"two-heads-are-better-than-one","label":"Two heads are better than one","statement":"Two people working on a problem will do better than either alone.","provenance":"borrowed","standing":"untested","counts":{"evidence":2},"url":"https://rsiratchet.com/heuristics/two-heads-are-better-than-one"}]}