Too Polite to Disagree

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

Research Question

The paper asks whether sycophancy, already observed in single LLMs, propagates and amplifies in multi-agent systems where models interact. It also investigates whether giving agents information about the sycophancy tendencies of their peers can improve collective decision-making and reduce error cascades. 

Methodology

The authors construct a multi-agent discussion system with six LLMs that iteratively debate a user’s claim across several rounds, starting with independent judgments and then updating based on peer responses. They introduce quantitative “sycophancy scores” (static and dynamic) for each agent and provide these as credibility signals to guide interaction. 

Findings

Providing agents with peer sycophancy rankings reduces the influence of highly sycophantic agents and limits error propagation during discussions. This intervention improves final decision accuracy by about 10.5 percentage points, showing that simple inference-time signals can significantly improve group reasoning. 

Limitations

The study uses controlled multi-agent setups with binary stance decisions, which may not capture the full complexity of real-world agent tasks or open-ended reasoning. It also focuses on synthetic or benchmark-style tasks, leaving questions about generalization to more realistic, tool-using or enterprise environments. 

Why It’s Important

The paper highlights a new failure mode where sycophancy is not just an individual bias but a collective phenomenon that can amplify across interacting agents. It also shows that lightweight, model-agnostic interventions can improve reliability without retraining, which is highly relevant for practical deployment of multi-agent AI systems.