Micro-Macro Links Between LLM Agent Personality and Collective Dynamics

Abstract

In recent years, the collaborative use of multiple large language models (LLMs) has attracted increasing attention for its potential to enhance problem-solving capabilities and enable high-fidelity social simulations. While these efforts have yielded valuable insights, a unified understanding of the underlying phenomena requires clarifying the mathematical principles that govern the collective behavior of LLMs.In this study, we systematically examine the dynamics of LLM agent populations through a framework grounded in statistical mechanics. Focusing on consensus formation via dialogue between agents holding opposing views, we uncover phase transitions induced by variations in network structure and agent personality. Moreover, we introduce a method for decomposing pairwise agent dynamics into interpretable components that capture individual personality traits.