[2506.17788] Bayesian Social Deduction with Graph-Informed Language Models

Computer Science > Artificial Intelligence arXiv:2506.17788 (cs) [Submitted on 21 Jun 2025 (v1), last revised 10 Apr 2026 (this version, v2)] Title:Bayesian Social Deduction with Graph-Informed Language Models Authors:Shahab Rahimirad, Guven Gergerli, Lucia Romero, Angela Qian, Matthew Lyle Olson, Simon Stepputtis, Joseph Campbell View a PDF of the paper titled Bayesian Social Deduction with Graph-Informed Language Models, by Shahab Rahimirad and 6 other authors View PDF HTML (experimental) Abstract:Social reasoning - inferring unobservable beliefs and intentions from partial observations of other agents - remains a challenging task for large language models (LLMs). We evaluate the limits of current reasoning language models in the social deduction game Avalon and find that while the largest models demonstrate strong performance, they require extensive test-time inference and degrade sharply when distilled to smaller, real-time-capable variants. To address this, we introduce a hybrid reasoning framework that externalizes belief inference to a structured probabilistic model, while using an LLM for language understanding and interaction. Our approach achieves competitive performance with much larger models in Agent-Agent play and, notably, is the first language agent to defeat human players in a controlled study - achieving a 67% win rate and receiving higher qualitative ratings than both reasoning baselines and human teammates. We release code, models, and a dataset to supp...