Contagion Networks: Evaluator Preference Propagation in Multi-Agent LLM Systems

The Silent Spread of Bias: How Evaluator Preferences Infect Multi-Agent LLM Systems

The latest preprint from arXiv (2606.20493v2) introduces a phenomenon the authors call "Contagion Networks" — a formal study of how evaluator preferences propagate through multi-agent LLM systems. The core finding is deceptively simple yet profound: when LLMs act as judges or evaluators within agent networks, their strategic biases don't stay isolated. They spread like an infection through the system's evaluation pathways, whether those biases come from explicit prompt instructions or from shared architectural priors baked into the models themselves.

This is not merely a theoretical curiosity. The paper demonstrates that even subtle preference signals — a tendency to favor longer responses, a bias toward certain reasoning patterns, or an affinity for particular stylistic choices — can cascade through multi-agent feedback loops. An agent evaluated favorably for a certain behavior becomes more likely to exhibit that behavior, which then gets reinforced by other evaluators, creating self-reinforcing feedback cycles that can systematically distort the entire system's outputs.

Why This Matters

For AI practitioners, this research exposes a critical blind spot in current multi-agent system design. Most teams focus on individual agent capabilities — how well each model performs its specific task. But the Contagion framework reveals that the evaluation architecture itself can become the dominant factor in system behavior. A well-intentioned evaluation prompt like "prefer concise answers" can, through propagation, create a system-wide collapse toward brevity that sacrifices accuracy or nuance.

The implications are particularly acute for:

• Automated quality assurance pipelines where LLM evaluators judge other LLMs
• Multi-agent debate and consensus systems where agents evaluate each other's reasoning
• Iterative refinement loops where outputs are repeatedly judged and improved

Implications for AI Practitioners

First, evaluator diversity is not optional — it's a systemic requirement. Using the same model family (e.g., all GPT-4 variants) for both generation and evaluation creates a closed loop where shared architectural biases amplify without correction. Practitioners should deliberately mix evaluator models with different training distributions, architectures, and prompting strategies.

Second, audit evaluation cascades, not just final outputs. The paper suggests tracing preference propagation paths through the network to identify where biases amplify. This means instrumenting multi-agent systems to track not just what decisions were made, but why evaluators favored certain outputs over others.

Third, consider external grounding points. The most robust multi-agent systems will include periodic "reality checks" — human evaluations, ground-truth benchmarks, or rule-based verifiers that are immune to the contagion dynamics. These serve as immune systems against runaway preference propagation.

Key Takeaways

• Evaluator preferences in multi-agent LLM systems propagate and amplify through feedback loops, creating systemic biases that can overwhelm individual agent capabilities
• Using homogeneous model families for both generation and evaluation creates dangerous closed loops of shared bias amplification
• Practitioners must instrument their systems to trace evaluation cascades, not just final outputs
• Robust multi-agent architectures require diverse evaluator models and external grounding points to break contagion dynamics