AVCORP: Identifying and Resisting Deception via Communication Theory and Theory of Mind
Summary
- Researchers introduce AVCORP, a corpus of 5,000 annotated utterances from 250 Avalon games, designed to study deception detection in AI agents.
- AVCORP integrates communication theory tactics and Theory of Mind reasoning traces, including abductive hypotheses, suspicion levels, and nested beliefs.
- The authors train Qwen3-8B on the corpus and evaluate it on held-out simulated games and human-played games.
- The trained Qwen3-8B detector outperforms untrained baselines and certain frontier models, matching Claude Opus 5.5 on held-out games.
Why it matters
Enhancing LLM agent safety and robustness against user and multi-agent deception requires advanced Theory of Mind reasoning and tactical communication modeling.