New Paper Says LLM Agents Fail to Explore Peers, Offers MACE

When two language model agents sit down to work together, do they actually probe each other's strengths and weaknesses before locking in a plan? A new arXiv preprint from Hyeong Kyu Choi, Jiatong Li, Wendi Li, Xin Eric Wang and Sharon Li argues that they do not. The authors report that modern LLM agents exhibit 'myopic and polarized interaction patterns' that leave them with worse coordination and higher regret than they should have.

The paper frames the challenge as a partially observable stochastic game, where each agent has to probe its peers to infer what they can do before settling on a strategy. The complaint is that current agents skip that probing step, converge early, and leave value on the table. The authors' proposed fix is Multi-Agent Contextual Exploration, or MACE, described as a lightweight framework that explicitly promotes exploration through structured peer selection. Across both contextual and parametric diversity settings, they say MACE substantially improves exploration behavior and downstream task performance, and they argue theoretically that the value of exploration grows as agent diversity grows.

Why this is worth watching if you are shipping agent stacks: as teams wire several LLM agents together into pipelines, the default social behavior of those agents becomes part of the system's reliability profile. If the base tendency is to collapse toward the same plan rather than to test collaborators, that is a failure mode you inherit whether you designed for it or not.

The honest caveat is that this is a preprint abstract, and the specifics readers will most want — which frontier models were tested, how large the regret gap actually was, how much MACE moved the numbers — are not in the material I could retrieve this run. Take the framing as a research claim rather than a settled measurement.

For anyone building agent orchestration, the practical takeaway is smaller than the framing suggests: if you are composing agents that differ in capability, do not assume they will figure each other out on their own. Wire the probing step in explicitly.