Why don't LLM agents naturally explore each other in teams?

Multi-agent LLM systems tacitly assume that reliable collective behavior emerges from agent-to-agent interaction — agents will probe each other, infer capabilities, and settle on good interaction strategies. This paper shows the assumption fails. Modern LLM agents interacting with one another exhibit myopic and polarized interaction patterns: they commit prematurely to a peer or strategy, select peers poorly, and accumulate higher regret. The RL framing is that exploration is foundational to autonomous behavior, and LLM agents simply do not do it on their own.

The formalization is a Multi-Agent Exploration problem modeled as a partially observable stochastic game (POSG): each agent must probe peers to infer their capabilities before it can identify who to interact with and how. The fix, MACE, is lightweight — it explicitly induces exploration through structured peer selection rather than hoping it emerges — and improves both exploration behavior and downstream performance, with gains that transfer to unseen tasks. Theoretically, the value of exploration increases with agent diversity: the more heterogeneous the population, the more there is to learn by probing, and the costlier premature commitment becomes.

This slots into a recurring diagnosis of multi-agent LLM failure as premature convergence. Since Why do multi-agent LLM systems converge without genuine deliberation?, the exploration failure looks like the interaction-selection analogue of the deliberation failure — agents collapse onto a choice before probing alternatives. And it sharpens why Does cognitive diversity alone improve multi-agent ideation quality?: diversity only pays off if agents actually explore it, which they must be explicitly guided to do.