Multi-Agent LLMs Fail to Explore Each Other

This paper studies whether LLM agents can explore one another’s capabilities and interaction strategies in multi-agent settings. It formalizes the problem as Multi-Agent Exploration in a partially observable stochastic game, and reports that modern agents often develop myopic and polarized interaction patterns, producing suboptimal coordination and higher regret. The authors introduce Multi-Agent Contextual Exploration (MACE), a lightweight framework that encourages structured peer selection. According to the abstract, MACE improves exploration behavior and downstream task performance across contextual and parametric diversity settings, while the paper theoretically argues that exploration becomes more valuable as agent diversity increases.