open-ended-llm-roleplay-game-where-character-must-withhold-authored CLM-0461In 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.
Capability: Stating false facts confidently
Observed on
One 20-minute session per condition, 33 participants, Llama-3.1-70B-instruct with no fine-tuning, a single hand-authored detective narrative; baseline was a single model with a lon.
Sources
- Validated by: Hand-coded hallucination rate over gameplay logs (17.80% of responses vs 6.27%), confessions extracted (10/33 vs 0/33), and a 9-item Likert survey where only progression and out-of-nowhere-reveal items differed; the tri-agent pipeline, the tree and the visual node board were evaluated together with no ablation.. Measured on logs from 33 within-subject participants, one narrative, one model. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- 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.Checking claims against evidence · reviewed by AI
- 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.Saying what actually drove the answer · reviewed by AI
- 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.Whether the measurement made the finding · reviewed by AI
- 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.Judging another's work as well as a careful reviewer would · reviewed by AI
- 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.Remembering across sessions · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: low. Falsifier as drafted: An ablation showing the hallucination reduction comes from the visual UI changing player behaviour (fewer adversarial prompts) rather than from the constrained retrieve-generate-verify pipeline, or a replication where injecting a single permitted fact does not lower unauthored-content rates. Automatic check flagged: figures not in the source: 6.3%.