Multi-Agent LLMs Lack Exploration Capabilities

uw-madison · hf · 2026-07-15

This work investigates: why multi-agent LLMs fail to "explore each other".

The authors point out that current LLM agents often exhibit the following behaviors when interacting:

Being overly myopic
Polarization of interaction patterns
Poor coordination, leading to higher regret

They formalize the problem as Multi-Agent Exploration and model it as a Partially Observable Stochastic Game (POSG): agents must actively probe their opponents to infer capability boundaries and find better collaboration strategies.

To address this, the paper proposes MACE (Multi-Agent Contextual Exploration):

A lightweight framework
Explicitly encourages exploration through structured peer selection

Experimental results show that MACE significantly improves exploration behavior and downstream task performance across different contextual / parametric diversity scenarios. The authors also provide theoretical analysis, noting that higher diversity among agents leads to greater exploration value.

The paper also emphasizes: a fundamental limitation of current LLM agents is the lack of an explicitly guided exploration mechanism. Code will be open-sourced on GitHub.

Original post →

Multi-Agent LLMs Lack Exploration Capabilities

More from coding & agent