Saying it does not know

Claims

• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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).
• observationsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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).
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.
• mechanismsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.
• observationsingle paper↑1↓0
  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.

Techniques

• Abstain unless one answer dominates the LoRA ensemble's probdecoding
  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
• Abstain when answers fail to track entity substitutionsprompting
  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.
• Ask about a generic AI agent, not the model itselfprompting
  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.
• Ask for P(True) after showing the model several of its own sprompting
  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
• Calibrate confidence per atomic claim, not per responseprocess
  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
• Conformal abstention from self-consistencydecoding
  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
• Conformal shortlist first, then calibrated margin abstentionprocess
  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
• Delete the one context example a question uniquely needsprocess
  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.
• Gate the answer on a separate self-verification call for conprocess
  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.
• Gate the answer on self-scored information needsprompting
  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.
• Pool semantic opinions across models, then split entropy intdecoding
  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
• Probe failure structure before scoring abstentionprocess
  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
• Read impossibility from the answerable-null residual subspacarchitecture
  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.
• Refusal-aware instruction tuningtraining
  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
• Replay matched trajectories to compare confidence-to-action process
  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
• Report calibration at trajectory checkpoints, split by horizprocess
  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.
• Score restraint and confidence separately against a matched process
  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
• Separate boundary-awareness from boundary-compliance probesprocess
  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
• Verbalized confidence after listing alternativesprompting
  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.

Sources

Every paper, post or observation behind this capability. Ones cited by a claim show the claim and its stance; ones filed directly are about the capability but have no claim drafted from them yet. Open one to add support or contest it.

• When Should LLMs Abstain? Chain-of-Self-Questioning for Selective Risk Control2026-09-15
  ↑ supports a claim here
• Strangers to Themselves: What Language Models Say About Themselves Is Generic2026-09-09
  ↑ supports a claim here
• Do Large Language Models Know What They Don't Know II? A Fully Behavioral, Non-Cognitive Measure of Epistemic Honesty2026-09-07
  ↑ supports a claim here
• Uncertainty Quantification for LLM Agents: A Taxonomy, an Evaluation Protocol, and an Empirical Study2026-09-07
  ↑ supports a claim here
• Solving versus Verifying: Catching Contradictions in Tax Reasoning Systems2026-09-05
  ↑ supports a claim here
• CUSP: Decomposable Collective Uncertainty for Multi-Agent Multimodal Reasoning2026-09-04
  ↑ supports a claim here
• CHARM: Character Hallucination for Multicultural Role Play Benchmark2026-09-01
  ↑ supports a claim here
• Good Memory Has ECC: Evaluating the Memory of Vision-Language Models Beyond Accuracy2026-08-31
  ↑ supports a claim here
• Recognition-Refusal Misalignment in LLMs: Why Models Answer Structurally Unanswerable Questions2026-08-29
  ↑ supports a claim here
• HalluPrism: When Multimodal Uncertainty Should Diagnose, Not Decide2026-08-29
  ↑ supports a claim here
• When Linguistic and Internal Confidence Diverge in Large Language Models2026-08-28
  ↑ supports a claim here
• Twin Worlds: Equivariance-Based Abstention for Evidence-Grounded Reasoning2026-08-28
  ↑ supports a claim here
• Evaluating Confidence-Gated Retrieval with Matched Trajectory Replay2026-08-27
  ↑ supports a claim here
• Load-Bearing Context: The Question Damage Score for Evaluating Context Reliance in Linguistic Reasoning2026-08-27
  ↑ supports a claim here
• Localize-Then-Decide Guarantees for LLM Judgments2026-08-26
  ↑ supports a claim here
• Credal Large Language Models for Semantic Commitment under Uncertainty2026-08-24
  ↑ supports a claim here · ↑ supports a claim here
• Claim-Level Confidence Calibration for Reliable Decision Making with Large Language Models2026-08-23
  ↑ supports a claim here · ↑ supports a claim here
• Training-Free Inference-Time Self-Reflection and Cost-Bounded Early Stopping for Large Language Models2026-08-19
  ↑ supports a claim here
• Mitigating LLM Hallucinations via Conformal Abstention2024-04-04
  ↑ supports a claim here
• R-Tuning: Instructing Large Language Models to Say `I Don't Know'2023-11-16
  ↑ supports a claim here
• Do Large Language Models Know What They Don't Know?2023-05-29
  ↑ supports a claim here
• Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback2023-05-24
  ↑ supports a claim here
• Language Models (Mostly) Know What They Know2022-07-11
  ↑ supports a claim here