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

Summary

Moves sycophancy from the single-agent setting into collaborative multi-agent discussion, running controlled experiments with six open-source LLMs in which agents are given peer sycophancy rankings computed by static (pre-discussion) and dynamic (online) strategies. Supplying these sycophancy priors reduces the influence of sycophancy-prone peers, mitigates error cascades, and improves final discussion accuracy by an absolute 10.5%. A mechanism citation for why deliberative structures degrade, and evidence that the degradation is a property of the interaction design rather than of the underlying models.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-08-29 literature sweep as a medium-priority candidate for the Architecture as Epistemology (ISR) paper (angle: correlated errors / homogeneity across LLM agents; aggregating LLM judgments). Verified against: Opened https://arxiv.org/abs/2604.02668 (arXiv abstract page, v2 14 Jul 2026; journal-ref: SIGDIAL 2026, pp. 795-814, ACL)

Critique / open questions

Queued — not yet assessed.

Key quotes

Large language models (LLMs) often exhibit sycophancy: agreement with user stance even when it conflicts with the model's opinion. (abstract)

While prior work has mostly studied this in single-agent settings, it remains underexplored in collaborative multi-agent systems. (abstract)