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

2026-06-18 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Contagion Networks, a new formal framework, measures how evaluator biases propagate through interacting LLM agents in multi-agent systems. A controlled 3-agent experiment using DeepSeek-chat, with structured, balanced, and evidence-based evaluator bias profiles, revealed that biases consistently propagate between agents, with gamma values ranging from 0.157 to 0.352. The study identified three propagation regimes and found that homogeneous-model agents produce contagion coefficients 3-5x weaker than cross-model coefficients observed in prior work (MM-EPC: gamma approx 0.85-1.3), placing them in a suppression regime. Critically, increasing the evaluator committee size from k=1 to k=3 reduced effective contagion by 72.4%, offering a practical mitigation strategy. The open-source Contagion Network experimental framework has been released.

Key takeaway

For machine learning engineers designing multi-agent LLM systems, recognize that evaluator biases propagate significantly, even within homogeneous models. You should consider increasing your evaluator committee size from one to three agents, as this demonstrably reduces effective bias contagion by 72.4%, enhancing system reliability. Utilize the Contagion Network framework to measure and manage these biases, ensuring more robust and fair multi-agent deployments.

Key insights

Evaluator biases propagate in multi-agent LLM systems, measurable via Contagion Networks.

Principles

• Evaluator biases consistently propagate across interacting LLM agents.
• Homogeneous LLM agents exhibit weaker bias contagion than cross-model interactions.
• Increasing evaluator committee size significantly mitigates bias propagation.

Method

Contagion Networks provide a formal framework to measure evaluator bias spread in multi-agent LLM systems, using a Cross-Agent Contagion Matrix Gamma_N.

In practice

• Utilize the open-source Contagion Network experimental framework.
• Increase evaluator committee size (k=1 to k=3) to reduce contagion by 72.4%.
• Consider DeepSeek-chat for multi-agent LLM experiments.

Topics

• Large Language Models
• Multi-Agent Systems
• Evaluator Bias
• Bias Propagation
• Contagion Networks
• AI Safety

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

Related on AIssential

• Simulating Hate Speech Cascades with Multi-LLM Agents: Empirical Grounding, Modeling Fidelity, and Intervention Strategies — Multi-LLM agents can faithfully simulate implicit hate speech cascades by conditioning on user profiles, community, and content.
• Language Models Exhibit Inconsistent Biases Towards Algorithmic Agents and Human Experts — LLMs exhibit inconsistent biases, preferring human experts in stated trust but algorithms in revealed preference tasks.
• 🔮 Is AI immune to groupthink? — LLM councils, like human committees, can suffer from groupthink, losing unique valuable ideas despite improving average answers.
• Misinformation Propagation in Benign Multi-Agent Systems — Multi-agent LLM systems can mitigate misinformation impact, but robustness depends on design and group dynamics.
• 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.
• 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.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.