role-play-assigned-character-models-usually-recognize-query-lies-outside CLM-0479In 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.
Capability: Saying it does not know
Observed on
Six LLMs (GPT-4o, GPT-5.5, Gemini-3.5-Flash, Llama-3.1-8B, Gemma-3-12B, Qwen3-8B) on 40 characters from five cultural-linguistic regions, using abstention-enabled multiple-choice i.
Sources
- Validated by: Applied to 680 awareness and 1,332 compliance items across six models; compliance-gap rate exceeded recognition-failure rate for every model (e.g. GPT-4o 91.3% awareness accuracy, 18.9% compliance accuracy, 72.1% gap vs 8.9% recognition failure), and a knowledge-verification probe confirmed 78-100% of factually correct gap cases in five of six models as parametric overrides.. Measured on a new benchmark of 40 characters, 680 awareness / 1,332 compliance / 736 verification items, native-reviewer validated. [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
- 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.Recalling facts it was trained on · reviewed by AI
- 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.Saying it does not know · reviewed by AI
- 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).Saying it does not know · reviewed by AI
- 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.Writing a reusable locator: regex, SQL, XPath · reviewed by AI
- 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).Stating false facts confidently · 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: medium. Falsifier as drafted: Models showing recognition-failure rates above compliance-gap rates on the same paired probes, or gap cases mostly failing the knowledge-verification check, would show the failure is unawareness rather than suppression.