Signal How does Adversarial Influence Scale in Multi-Agent Systems?

Summary

The paper asks what happens in multi-agent deliberation when some agents do not act in good faith, and studies how susceptibility to deception scales with group size and the prevalence of deceivers. It finds that the proportion of deceivers matters rather than the number of agents. The defection rate, how often initially correct agents switch to an incorrect final answer, rises linearly with that proportion. Humans in comparable conformity studies are swayed reliably only by a majority, but LLM agents defect regularly even when deceivers are a minority. Susceptibility depends on which models interact, especially on the honest-agent side, and private coordination among deceivers can unexpectedly make them less effective. The authors conclude that adding more agents is not a sufficient defense.