Bayesian Social Deduction with Graph-Informed Language Models
Summary: arXiv:2506.17788v2 Announce Type: replace
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 support future work on social reasoning in LLM agents, which can be found at https://camp-lab-purdue.github.io/bayesian-social-deduction/.
Introduction
In recent years, the field of artificial intelligence has made significant strides, particularly in the development of large language models (LLMs). These models have shown remarkable capabilities in understanding and generating human-like text. However, a key area where they still struggle is in social reasoning—specifically, the ability to infer beliefs and intentions from limited observations of others.
Challenges in Social Reasoning
To investigate the effectiveness of existing reasoning models, researchers assessed their performance in the social deduction game Avalon, a game that inherently requires players to deduce the motivations and intentions of others. The findings indicated that although the largest LLMs performed well, they faced significant limitations:
- Extensive test-time inference was necessary for optimal performance.
- Performance degraded sharply when models were distilled to smaller versions, which are more suitable for real-time applications.
A Hybrid Reasoning Framework
In response to these challenges, the researchers introduced a novel hybrid reasoning framework. This approach separates belief inference from the language understanding task:
- A structured probabilistic model is employed for belief inference.
- An LLM is utilized for language understanding and interaction.
This innovative combination has led to competitive performance in agent-agent play scenarios, marking a significant advancement in the capabilities of language agents.
Achievements and Future Work
One of the most notable achievements of this research is that the newly developed language agent has successfully defeated human players in controlled studies. It achieved a remarkable 67% win rate, outperforming both reasoning baselines and even human teammates in qualitative assessments.
To support ongoing research in this area, the authors have made their code, models, and dataset publicly available. This resource is expected to facilitate further advancements in social reasoning within LLM agents.
Conclusion
The integration of Bayesian social deduction techniques with graph-informed language models represents a promising direction for improving social reasoning in AI. As researchers continue to explore this domain, we can anticipate significant advancements that could enhance how AI systems understand and interact in complex social environments.