Evaluator bias spreads between LLM agents like a contagion, study finds
Zewen Liu introduces Contagion Networks, a formal framework showing that evaluator biases in LLM-based multi-agent systems propagate measurably between agents, with committee-based mitigation cutting effective contagion by 72.4%.
Score breakdown
The paper provides a measurable, formal account of how evaluation bias spreads in multi-agent LLM pipelines and identifies a concrete structural intervention — expanding evaluator committee size — that reduces that spread by 72.4%.
- 01Zewen Liu introduces Contagion Networks, a formal framework for measuring evaluator bias propagation in multi-agent LLM systems.
- 02A controlled 3-agent experiment used DeepSeek-chat with three evaluator bias profiles: structured, balanced, and evidence-based.
- 03Measured contagion coefficients gamma ranged from 0.157 to 0.352, confirming bias spreads even within the same underlying model.
Zewen Liu's paper introduces Contagion Networks, a formal framework for measuring how evaluator biases propagate through multi-agent LLM systems when those models serve as evaluators. The core contribution is the Cross-Agent Contagion Matrix Gamma_N, whose spectral radius rho(Gamma_N) governs which of three propagation regimes a system falls into. In a controlled experiment with 3 agents running DeepSeek-chat under three distinct evaluator bias profiles (structured, balanced, and evidence-based), the measured contagion coefficients gamma fall in the range [0.157, 0.352] — confirming that bias propagation occurs even when all agents share the same underlying model.
The paper also demonstrates a practical mitigation: increasing evaluator committee size from k=1 to k=3 reduces effective contagion by 72.4%.
A key comparative finding is that homogeneous-model agents produce contagion coefficients 3–5x weaker than the cross-model coefficients reported in prior work (MM-EPC: gamma ≈ 0.85–1.3), placing same-model multi-agent systems in the suppression regime rather than an amplification regime. The paper also demonstrates a practical mitigation: increasing evaluator committee size from k=1 to k=3 reduces effective contagion by 72.4%. The open-source Contagion Network experimental framework accompanies the paper.
Key facts
- 01Zewen Liu introduces Contagion Networks, a formal framework for measuring evaluator bias propagation in multi-agent LLM systems.
- 02A controlled 3-agent experiment used DeepSeek-chat with three evaluator bias profiles: structured, balanced, and evidence-based.
- 03Measured contagion coefficients gamma ranged from 0.157 to 0.352, confirming bias spreads even within the same underlying model.
- 04Three propagation regimes are identified, governed by the spectral radius rho(Gamma_N) of the Cross-Agent Contagion Matrix.
- 05Homogeneous-model agents produced contagion coefficients 3–5x weaker than cross-model coefficients in prior work (MM-EPC: gamma ≈ 0.85–1.3), placing them in the suppression regime.
- 06Increasing evaluator committee size from k=1 to k=3 reduces effective contagion by 72.4%.
- 07The open-source Contagion Network experimental framework is released alongside the paper.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 19, 2026 · 10:25 UTC. How this works →