Misinformation Propagation in Benign Multi-Agent Systems
This study investigated how contextual misinformation affects multi-agent debate (MAD) where agents follow a benign protocol and none intend to deceive. Researchers tested three open-weight LLMs using 10,278 misinformation passages across eight styles. Findings show that while single agents are vulnerable, MAD is more robust; group composition and decision protocols significantly influence this robustness.