[2Yin-B] Poster 4

[2Yin-B-52] A Note on Collective Behavior Analysis in Distributed LLM-based Multi-Agent Systems Based on Interaction Strength Control

〇Taiyo Sato¹, Keisuke Maeda¹, Naoki Saito¹, Takahiro Ogawa¹, Miki Haseyama¹ (1. Hokkaido University)

Keywords:

Large language models,multi-agent systems,distributed cooperation,consensus building

In this paper, we investigate collective dynamics in distributed LLM-based multi-agent systems (LLM-MAS) through a numerical consensus task. While distributed LLM agents can exhibit coherent behavior without central control, how interaction strength influences such dynamics remains insufficiently understood. We introduce a minimal framework where agents on a toroidal grid update integer states using only local comparisons. Interaction strength is controlled by a single parameter modulating receptivity to local observations. By varying this parameter, we observe a non-monotonic transition: weak interaction leads to stagnation, strong interaction induces instability, and intermediate interaction reduces variance toward consensus. This pattern appears across various LLMs and remains robust under changes in interaction steps and grid size. In some settings, explicit receptivity improves performance over a baseline without interaction-strength control. These results demonstrate that interaction strength plays a central role in shaping collective behavior in distributed LLM-MAS and highlight the importance of systematically examining this parameter.