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

2026-06-19 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, long

Summary

A new framework, Contagion Networks, quantifies evaluator bias propagation in multi-agent LLM systems. Experiments using DeepSeek-chat agents with distinct bias profiles (structured, balanced, evidence-based) revealed that evaluator biases consistently propagate, with mean contagion coefficients γ ∈ [0.143, 0.304]. Homogeneous-model agents typically operate in a "suppression regime," where bias attenuates rapidly across hops (cumulative factor β₃=0.0055). This contrasts sharply with prior cross-model work (MM-EPC), which observed 3–5× stronger contagion (γ ≈ 0.85–1.3) leading to a "cascade regime." Crucially, increasing the evaluator committee size from k=1 to k=3 reduced effective contagion by 72.4%, demonstrating a practical mitigation strategy. The open-source framework is released for community use.

Key takeaway

For AI Architects designing multi-agent LLM systems where agents evaluate each other, you must account for evaluator bias propagation. Measure the Cross-Agent Contagion Matrix Γₚ pre-deployment to predict system stability. Prioritize homogeneous model families for evaluators, as they offer natural bias suppression. Additionally, implement evaluator committees of at least three agents and monitor strategy entropy H(ω) as a real-time health indicator to maintain cognitive diversity.

Key insights

Evaluator biases propagate in multi-agent LLM systems, but homogeneous models suppress it, while diverse evaluator committees mitigate it.

Principles

• Bias propagation magnitude depends on evaluator diversity.
• Homogeneous model pools provide implicit regularization.
• Spectral radius ρ(Γₚ) governs propagation regimes.

Method

The Contagion Networks framework measures bias propagation using a Cross-Agent Contagion Matrix Γₚ and Test-Time Reinforcement Learning (TTRL) for strategy updates.

In practice

• Measure Γ before system deployment.
• Prefer homogeneous evaluator pools.
• Use committees of ≥ 3 evaluators.

Topics

• Multi-Agent Systems
• LLM Evaluation
• Bias Propagation
• Contagion Networks
• Evaluator Diversity
• DeepSeek-chat

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Contagion Networks: Evaluator Bias Propagation in Multi-Agent LLM Systems — Evaluator biases propagate in multi-agent LLM systems, measurable via Contagion Networks.
• Agentic CLEAR: Automating Multi-Level Evaluation of LLM Agents — Agentic CLEAR automates multi-level evaluation for LLM agents, providing dynamic, data-driven insights into their behavior.
• When Agents Disagree: The Selection Bottleneck in Multi-Agent LLM Pipelines — Selector quality, not just team diversity, dictates multi-agent LLM pipeline performance.
• Network Effects and Agreement Drift in LLM Debates — LLM agent opinion dynamics are shaped by network structure, group imbalance, and a systematic "agreement drift" bias.
• DIY #19 - Evaluator-Optimiser LLM Agent with LangChain — The Evaluator-Optimiser pattern uses iterative feedback between two LLMs to achieve higher quality outputs for complex tasks.
• Beyond Static Responses: Multi-Agent LLM Systems as a New Paradigm for Social Science Research — A six-tier framework maps LLM agent evolution from static tools to complex adaptive systems for social science research.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.