Adding agents doesn't dilute deceivers: LLM defection rises linearly with the share of bad-faith agents, even as a minority

arXiv·medium signalXBlueskyLinkedInCopy link

The authors vary group size and the proportion of deceptive agents in multi-agent deliberation. The rate at which initially correct agents switch to a wrong answer rises linearly with the share of deceivers, and group size itself has no effect. Unlike humans in conformity studies, LLM agents defect even when deceivers are a minority, and letting deceivers coordinate privately made them less effective. Scaling a debate or voting ensemble is therefore not a defense against a compromised participant.