Small models sway large ones in multi-agent LLM disagreements

The study reveals that persuasibility in multi-agent settings is decoupled from model size and confidence, meaning you cannot predict which models will influence or be influenced by peers based on standard benchmarks. This breaks a common assumption that larger models are more robust to social pressure.

This connects directly to the finding from early October that multi-agent systems can produce correct task outputs while corrupting shared state behind the scenes. Together, these papers suggest that heterogeneous model collaboration introduces two distinct failure modes: models may agree on wrong answers due to peer influence (this paper), or they may solve tasks correctly but leave unreliable information trails for downstream agents (the state corruption work). The gender bias heterogeneity study from late September also matters here: if models diverge radically on bias even within the same training paradigm, they're likely to diverge on susceptibility to peer pressure too. Organizations deploying mixed-model systems need to account for influence dynamics as a separate risk layer from task accuracy.

If researchers test whether fine-tuning or constitutional AI methods can reduce cross-model persuasibility without degrading task performance, that confirms this is a tractable deployment problem. If major cloud providers (Azure, Bedrock, Vertex) publish guidance on model pairing strategies within the next six months, that signals the finding has moved from academic concern to operational priority.