多智能体LLM在推理努力与通信拓扑下的集体状态
原标题：Collective Regimes in Multi-Agent LLMs under Reasoning Effort and Communication Topology
AI 摘要
该研究将多智能体LLM系统建模为无状态智能体在局部可见邻居间反复更新预测，并用全局与局部一致性指标识别出三种集体状态：同步、扭曲（局部有序但全局不一致）和嵌合体（一致与不一致子群共存）。实验发现，提高gpt-5-mini的推理努力会把面板从碎片化结果推向局部有序的扭曲状态，而增加通信连通性则推动系统走向全局同步；在非环形评判任务和三个提供商的模型上拓扑效应同样出现。作者指出，仅凭聚合一致性不足以刻画多智能体LLM的集体行为。
正文节选
Collective Regimes in Multi-Agent LLMs under Reasoning Effort and Communication Topology Abstract Multi-agent LLM systems are increasingly used for deliberation and evaluation, often under the assumption that greater peer interaction leads to more reliable consensus. Existing work largely evaluates these systems through final accuracy or aggregate agreement. However, such measures do not reveal how agreement is organized in the panel. In this paper, we study stateless LLM agents that update thei