'Multi-Agent LLMs Fail to Explore Each Other' — And a Lightweight Fix (MACE) Cuts the Regret

arXiv·medium signal

This paper shows modern LLM agents interacting with one another exhibit myopic, polarized patterns that cause suboptimal coordination and increased regret, formalizing the problem as a multi-agent exploration POSG where agents must probe peers to infer their capabilities. The authors introduce MACE (Multi-Agent Contextual Exploration), a lightweight framework that promotes exploration through structured peer selection, and prove theoretically that exploration's value rises with agent diversity. Across contextual and parametric diversity settings MACE substantially improves both exploration and downstream task performance — relevant for anyone orchestrating heterogeneous agent teams.