When large language models (LLMs) are used as evaluators in multi-agent systems — for example, in automated quality assurance, supply chain validation, or content moderation — their systematic evaluation biases can spread through the agent network, a new research paper warns. The study, titled "Contagion Networks: Evaluator Bias Propagation in Multi-Agent LLM Systems" and authored by Zewen Liu, introduces a formal framework for measuring this phenomenon.
The research, published on arXiv under a Creative Commons license, is directly relevant to enterprise technology leaders deploying LLM-based agents for decision-making tasks where consistency and objectivity are critical.
What Are Contagion Networks?
According to the paper, Contagion Networks provide a mathematical structure to quantify how evaluator biases — systematic preferences or tendencies in scoring or judging — propagate when multiple LLM agents interact. The framework measures this via a Cross-Agent Contagion Matrix, denoted Gamma_N, where N is the number of agents.
The paper defines three propagation regimes determined by the spectral radius rho(Gamma_N): a suppression regime where biases decay, a critical regime, and an amplification regime where biases strengthen.
Experiment and Key Findings
The researchers conducted a controlled 3-agent experiment using DeepSeek-chat with three distinct evaluator bias profiles: structured, balanced, and evidence-based. They measured the Cross-Agent Contagion Matrix Gamma_3 and found that evaluator biases consistently propagated between agents, with contagion coefficients gamma in the range [0.157, 0.352]. This occurred even though all agents used the same underlying model.
| Metric | Value |
|---|---|
| Contagion coefficient range (homogeneous model) | [0.157, 0.352] |
| Contagion coefficient from prior cross-model work (MM-EPC) | approx 0.85-1.3 |
| Homogeneous vs. cross-model strength | 3-5x weaker |
| Regime for homogeneous agents | Suppression |
| Contagion reduction from committee size k=1 to k=3 | 72.4% |
Notably, the contagion coefficients for homogeneous-model agents were 3-5x weaker than cross-model coefficients reported in prior work (the MM-EPC benchmark, where gamma ≈ 0.85-1.3). This places homogeneous-model systems in the suppression regime, meaning biases are less likely to amplify but still propagate measurably.
Actionable Mitigation: Committee Sizing
The paper offers a concrete, actionable mitigation strategy: increasing the evaluator committee size. Specifically, raising the committee from k=1 to k=3 reduces effective contagion by 72.4%. This finding is critical for enterprise architects designing multi-agent LLM systems, where evaluation tasks — such as code review, contract analysis, or logistics plan scoring — often rely on a single evaluator agent.
Implications for Enterprise Deployments
For CTOs and digital transformation leaders, the research underscores that using multiple, diverse LLM-based evaluators can significantly dampen bias propagation. The open-source Contagion Network experimental framework released by the authors allows organizations to test bias dynamics in their own agent topologies.
While the experiment focused on a single model (DeepSeek-chat), the framework is model-agnostic. Enterprises integrating multiple LLMs — from different vendors or fine-tuned for specific domains — may see stronger contagion effects, as suggested by prior cross-model coefficients.
The study provides a quantitative basis for designing robust multi-agent evaluation systems, a growing area in supply chain automation, trade compliance checks, and other high-stakes enterprise applications where bias can cascade into costly errors.
