Do multi-agent LLM systems develop differentiated roles?

Determine whether multi-agent systems composed of large language model agents develop differentiated roles (role specialization) rather than remaining undifferentiated collections of agents.

Background

The paper investigates emergent coordination and synergy in multi-agent systems built from LLMs, using information-theoretic tools such as partial information decomposition and time-delayed mutual information. A central theme is whether groups of LLM agents exhibit higher-order structure beyond individual behavior.

While the authors present evidence of emergent structure in a controlled group guessing task and show that prompting can influence coordination regimes, they identify a broader uncertainty in the literature as to whether LLM-based multi-agent systems generally develop stable, differentiated roles across settings.