Presentation Information
[1M4-GS-5x-01]Analyzing the Beauty Contest Game with LLM-Based Multi-Agent Systems: A Focus on the Distinct Roles of Cognitive Hierarchy and Personality Traits
〇Kazuki Matsui1 (1. JAIST Support Organization)
[[online]]
Keywords:
LLM,Multi Agent,Cognitive Hierarchy,Big Five Personality Traits
Recent advances in Large Language Model–based multi-agent simulations (LLM-MAS) have raised interest in their ability to reproduce human strategic decision-making. This study examines this potential using the beauty contest game. We construct an LLM-MAS in which reasoning depth is controlled through cognitive hierarchy (CH), while Extraversion, a Big Five personality trait, is introduced as an agent characteristic. Simulation results indicate that Extraversion does not predict reasoning depth, nor does it show a direct statistical effect on prediction error. However, descriptive analyses suggest that agents with higher Extraversion tend to adjust their behavior more strongly over repeated rounds. These findings indicate that while cognitive hierarchy shapes strategic inference, personality traits may be related to how inferred social information is translated into action, particularly in adjustment dynamics. This study provides an exploratory framework for incorporating personality traits into LLM-based social simulations.