Too Polite to Disagree: Understanding Sycophancy Propagation in Multi-Agent Systems

arXiv cs.AI · 18 hours ago

Researchers tested whether large language models in multi-agent discussions reduce unhelpful agreement behavior when informed about peers' sycophancy tendencies through rankings. The method improved final discussion accuracy by 10.5 percentage points across six open-source LLMs. Making agents aware of sycophancy levels enables them to discount unreliable agreement and reach more accurate conclusions.