Research map / Large language models / Language, safety and interpretability
Concept Incongruence: An Exploration of Time and Death in Role Playing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces "concept incongruence" and studies how LLMs behave when role-playing historical figures past their death date, measuring abstention, answer rate, and conditional accuracy, and probing death-year and temporal representations.
Problem
LLMs given conflicting concept boundaries (e.g., role-playing a deceased person while asked about later events) fail to abstain consistently and show accuracy drops; the paper asks how such incongruence shows up in behavior and internal representations.
Approach
Frames concept incongruence across three levels and uses role-play of deceased figures as a controllable testbed; defines abstention rate, answer rate and conditional accuracy metrics, and uses linear probes (dead/alive, death year, calendar time) to locate failures in representations.
Key results
- Llama abstains after death at 18.7% and Claude at 9.6%, far from the expected 100%; Gemma and GPT-4.1 abstain under 3% of the time.
- Conditional accuracy in Role-Play drops by 8% for Gemma and 8.2% for GPT-4.1 relative to Non-Role-Play, which stays at 100%.
- Adding explicit death year and check instructions raises after-death abstention substantially but lowers conditional accuracy further (6.5% for Llama, 8.1% for Gemma).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Role-Play instead of Non-Role-Play | Llama-3.1-8B-Instruct · Generalized temporal questions | accuracy: 85.0 -> 38.6 | Hurt |
| Use Role-Play instead of Non-Role-Play | Gemma-2-9B-Instruct · Generalized temporal questions | accuracy: 93.0 -> 81.8 | Hurt |
| Use Role-Play instead of Non-Role-Play | Llama-3.1-8B-Instruct · Generalized temporal questions | RMSE: 7.3 -> 7.7 | Hurt |
| Use Role-Play instead of Non-Role-Play | Gemma-2-9B-Instruct · Generalized temporal questions | RMSE: 8.2 -> 8.9 | Hurt |
| Use Role-Play instead of Non-Role-Play | Llama-3.1-8B-Instruct · Dead/alive questions | accuracy: 100% -> 88.9% | Hurt |
| Use Role-Play instead of Non-Role-Play | Llama-3.1-8B-Instruct · Death-year questions | conditional accuracy: about 91% -> 84% | Hurt |
| Add explicit death year and strict time-checking instructions (ablation) | Llama Role-Play · Questions after character death | abstention rate: +75.5% | Helped |
| Add explicit death year and strict time-checking instructions (ablation) | Gemma Role-Play · Questions after character death | abstention rate: +66.1% | Helped |
| Add explicit death year and strict time-checking instructions (ablation) | Llama Role-Play · Questions after character death | answer rate: -83.6% | Helped |
| Add explicit death year and strict time-checking instructions (ablation) | Gemma Role-Play · Questions after character death | answer rate: -66.1% | Helped |
| Add explicit death year and strict time-checking instructions (ablation) | Llama Role-Play · Conditional accuracy | conditional accuracy: -6.5% | Hurt |
| Add explicit death year and strict time-checking instructions (ablation) | Gemma Role-Play · Conditional accuracy | conditional accuracy: -8.1% | Hurt |
Surprising findings
- Role-play shifts temporal representations: probe RMSE rises by 8.2 years for president-year questions under Role-Play, while Spearman correlation stays high.
- Models often abstain and then answer anyway (e.g., Llama abstains yet gives Joe Biden as the answer), so abstention and answer rates do not sum to one.
Limitations
- Evaluation centers on U.S. president questions, limiting coverage of temporal reasoning.
- Focuses on a relatively easy-to-trace incongruence (I-B), not hidden internal conflicts.
- Desired role-play behavior is an open question; the expected behavior assumed may not be universally accepted.
Experimental setup
Datasets: 100 historical figures who died 1890-1993 (custom) six living public figures artwork release dates dataset (entertainment dataset) custom U.S. president temporal questions (277 questions)
Benchmarks: custom role-play death-time benchmark (100 historical figures) artwork release date yes/no questions
Metrics: abstention rate answer rate conditional accuracy Spearman correlation RMSE probe accuracy
Baselines: Non-Role-Play prompting Role-Play prompting
Builds on: Stroop effect (conceptual analogy) linear probing calendar-time representation probes
Closest papers by meaning
- ToMATO: Verbalizing the Mental States of Role-Playing LLMs for Benchmarking Theory of Mind (AAAI 2025 · similarity 0.87)
- Codifying Character Logic in Role-Playing (NeurIPS 2025 · similarity 0.87)
- The Illusion of Role Separation: Hidden Shortcuts in LLM Role Learning (and How to Fix Them) (ICML 2025 · similarity 0.85)
- Thinking in Character: Advancing Role-Playing Agents with Role-Aware Reasoning (NeurIPS 2025 · similarity 0.85)
- Emergent Response Planning in LLMs (ICML 2025 · similarity 0.85)
- OmniToM: Benchmarking Theory of Mind in LLMs via Explicit Belief Modeling (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.