AVCORP trains an LLM to recognize deception using the Avalon corpus
Start here
In November 2026, arXiv published a study on the AVCORP corpus, which used 5,000 annotated utterances from 250 Avalon games to train Qwen3-8B to recognize deception. The corpus maps 37 strategies onto a 4x4 behavior matrix and includes reasoning traces with abductive hypotheses, suspicion levels and nested theory-of-mind beliefs. The study claims that its best detector surpassed frontier models such as GPT-5.6 Sol and matched Claude Opus 5.5 on 25 held-out games; on 20 human games, all trained detectors had higher F1 across all rounds than untrained models.
Generated by AI from reports · Updated 2 days ago
Attention and developments
Attention, hourly
Now 25, -2% in 24 hours·Peak 28 (Nov 10 20:00)
Related research
This event's paper
Progress timeline
- arXiv: Alignment, Deception & OversightThe AVCORP corpus: identifying and defending against deception based on propagation theory and theory of mind
The AVCORP corpus uses 5,000 annotated utterances from 250 Avalon games to train Qwen3-8B to identify deception; its best detector surpasses frontier models such as GPT-5.6 Sol and matches Claude Opus 5.5 on 25 held-out games. The corpus maps 37 strategies onto a 4x4 behavior matrix and includes reasoning traces with abductive hypotheses, suspicion levels and nested ToM beliefs. In 20 human games, every trained detector's F1 across all rounds was higher than that of untrained models.
Related discussion
Attention trend
New sources per day
- Nov 10: 1 new sources (1 outlets, 0 accounts)
- Nov 11: 0 new sources (0 outlets, 0 accounts)
- Nov 12: 0 new sources (0 outlets, 0 accounts)
- Nov 13: 0 new sources (0 outlets, 0 accounts)