What they found
The authors built AVCORP, a corpus of 5,000 annotated utterances from 250 games of Avalon, a social game built on hidden roles and lying. It labels 37 deception tactics, mapped to a 4x4 behavior matrix. It also includes reasoning traces with abductive hypotheses (best-guess explanations), suspicion levels, and nested Theory of Mind beliefs (what one player thinks another player believes). They trained Qwen3-8B on it and tested on 25 held-out games and 20 games played by humans. On held-out games, their best trained detector beat its untrained baseline and the frontier models GPT-5.6 Sol, Gemini 3.1 Pro and DeepSeek-V4-Pro, and matched Claude Opus 5.5. On human games, every trained detector beat the untrained model in F1 score across all rounds. The abstract gives no exact scores.
Why it matters
SI agents are starting to talk with users and with other agents, and either may try to deceive them. This work offers a training set and a method aimed at spotting deception. It also shows that a small 8B model can compete with much larger ones on this narrow task. The abstract does not say how well results carry over from a game to real-world deception.
How it changes the race
If the result holds, deception detection may not need the largest models, which would lower the cost of building it. It is a narrow, game-based result, so it is an incremental step.
Safety and control
The paper bears directly on defending SI agents against manipulation by users and other agents. The abstract does not say whether skills learned in Avalon transfer to other settings.
The abstract, by the authors
Today's LLM-based agents are vulnerable to deception by users and other agents. Defending against deception requires Theory of Mind (ToM) reasoning to infer the beliefs and goals of others as well as modeling tactics based on communication theory. To enable training models to detect deception, we adopt the well-known deception game, Avalon. We present a corpus, AVCORP, of 5,000 annotated utterances from 250 games of Avalon. AVCORP includes (1) 37 tactics mapped to a 4x4 behavior matrix, and (2) reasoning traces capturing abductive hypotheses, suspicion levels, and nested ToM beliefs. We train Qwen3-8B on AVCORP and evaluate it on 25 held-out games and 20 games played by humans. On held-out games, our best trained detector outperforms its untrained baseline and frontier models (GPT-5.6 Sol, Gemini 3.1 Pro, DeepSeek-V4-Pro), matching Claude Opus 5.5. On human games, every trained detector exceeds the untrained model in F1 score across all rounds.
Quoted from arXiv:2611.10176. The research and its words belong to its authors.