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

Vira Kasprova*, Amruta Parulekar*, Abdulrahman AlRabah*, Krishna Agaram*, Ritwik Garg, Sagar Jha, Nimet Beyza Bozdag, Dilek Hakkani-Tür (* equal contribution)

SIGDIAL 2026 Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL 2026), 2026

Paper Code

@inproceedings{kasprova2026polite,
  title     = {Too Polite to Disagree: Understanding Sycophancy Propagation in Multi-Agent Systems},
  author    = {Kasprova, Vira and Parulekar, Amruta and AlRabah, Abdulrahman and Agaram, Krishna and Garg, Ritwik and Jha, Sagar and Bozdag, Nimet Beyza and Hakkani-T{\"u}r, Dilek},
  booktitle = {Proceedings of the 27th Annual Meeting of the Special Interest Group on Discourse and Dialogue},
  pages     = {795--814},
  year      = {2026},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2026.sigdial-1.56}
}

Discussion pipeline: base sycophancy scores computed per model, then six agents debate over five rounds and a majority vote picks the final answer. — Figure 1 from the paper.

Sycophancy is usually measured on a single model answering a single user. But once models talk to each other, agreement becomes contagious: one agent’s incorrect concession can pull an entire discussion toward the wrong answer.

We run controlled experiments with six open-source LLMs discussing questions over five rounds. Before and during the discussion, each agent receives a ranking of its peers’ tendency toward sycophancy, estimated with static (pre-discussion) and dynamic (online) scoring strategies, including the Base Sycophancy Score (BSS) and the Dynamic Sycophancy Score (DSS).

Providing these sycophancy priors reduces the influence of sycophancy-prone peers, mitigates error cascades, and improves final discussion accuracy by an absolute 10.5%. It is a lightweight way to reduce sycophancy in discussion without modifying the models.