Larger models dominate consensus in multi-agent LLM systems
October 23, 2026
In multi-agent systems, the final consensus is driven by the model with the highest parameter count, which acts as an 'influencer' by absorbing the opinions of smaller models. This phenomenon is attributed to larger models having stronger in-context learning capabilities, effectively creating a deeper well in the system's energy landscape.
HOW THIS AFFECTS YOU
●
builderBe aware that adding smaller models to a multi-agent swarm may not provide diversity if a much larger model is present.
●
researcherInvestigate the dynamics of in-context learning weight in multi-agent consensus formation.