Multi-Agent Systems Are Mixtures of Experts: Opinion Dynamics Framework Reveals Who Becomes an Influencer in LLM Deliberation

arXiv·medium signal

Studies multi-agent LLM deliberation through Friedkin-Johnsen opinion dynamics, showing that stubbornness, influence, and opinion change parameters are input-dependent — turning multi-agent deliberation into a mixture of experts. Provides a tractable analytical framework for understanding why some agents dominate consensus in multi-agent systems. Useful for practitioners designing agent ensembles.