Sycophancy Shapes Multi-Agent Debate

Peacemaker or Troublemaker: How Sycophancy Shapes Multi-Agent Debate.

Research question

The paper asks how sycophancy operates not between a user and a model, but among agents inside a multi-agent debating system. More specifically, it studies whether excessive inter-agent agreement causes debates to collapse into premature consensus and whether this effect differs for debaters versus judges in decentralized and centralized debate setups. 

Methodology

The authors propose an operational framework for sycophancy in multi-agent debate, including a formal definition, a system-level metric called disagreement collapse rate (DCR), and agent-level sycophancy scores. They implement decentralized and centralized debate frameworks in AutoGen, test Qwen3-32B and Llama 3.3-70B on MMLU-Pro and CommonsenseQA, and control agent “persona” with prompt-defined sycophancy levels from low-sycophancy “troublemaker” to high-sycophancy “peacemaker.” 

Findings

The paper finds that sycophancy is a major failure mode in multi-agent debate because it pushes agents toward premature agreement, which often lowers accuracy and can make debate systems fail to outperform single-agent baselines. High debater sycophancy is especially harmful, while the best results often come from balanced role mixes rather than all-agreeable agents; by contrast, judge sycophancy seems to matter less, suggesting centralized systems are more robust to that specific effect. 

Limitations

The authors note that their evaluation is limited to specific model families and multi-agent frameworks, so the results may not generalize to other LLMs, scales, or collaboration architectures. They also acknowledge that their metrics may miss other forms of sycophancy across domains and cultures, and that their mitigation ideas still need larger real-world deployment studies to test long-term robustness and unintended effects. 

Why it’s important

This paper matters because it shows that sycophancy can be a collective systems problem, not just an individual model trait. It also offers a practical design lesson for multi-agent systems: reliability depends not only on model quality, but on structuring agent roles and interaction dynamics so that productive disagreement is preserved instead of collapsed.