Definition
Sycophancy means a model tells others what they want to hear instead of what is correct. Conformity is the multi-agent version: a model changes its answer to match its peers.
- Peer pressure (Song et al., 2025): shown their peers’ answers, models lose more correct answers than they gain. About 65% of the answer changes go from right to wrong, and average accuracy drops 5.65%.
- Disagreement collapse (Yao et al., 2025): in debate, agents give up correct positions in order to agree. In homogeneous debates this reached up to 86%, with a 0.90 correlation to sycophancy. A persistent “troublemaker” role helps.
- Single-model flaws compound across agents instead of averaging out. More agents can make the answer worse.