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

Researchers demonstrate that evaluation biases in large language models systematically spread through multi-agent systems, with a new framework showing biases propagate at rates of 15.7-35.2% between same-model agents. Deploying evaluation committees of three agents reduces contagion by 72.4%, offering a practical mitigation strategy for AI systems relying on LLM evaluators.

This research addresses a critical vulnerability in autonomous multi-agent AI systems: the propagation of systematic biases through evaluator networks. When LLMs evaluate other agents' outputs, their inherent biases don't remain isolated—they influence subsequent agent behavior, creating feedback loops that amplify distortions. The Contagion Networks framework quantifies this phenomenon mathematically, revealing that even identical model architectures exhibit measurable bias propagation between 15.7% and 35.2%, fundamentally challenging assumptions about homogeneous system reliability.

The findings carry significant implications for AI system design. Current deployments often assume that using the same model repeatedly provides consistency and safety. This research exposes that consistency can mean consistent propagation of systematic errors. The distinction between same-model contagion (relatively suppressed) and cross-model contagion (3-5x stronger) suggests architectural diversity may be beneficial, though this introduces complexity tradeoffs. The practical mitigation—increasing evaluator committee size from one to three—reduces effective contagion by nearly three-quarters, providing an immediately implementable solution.

For developers building multi-agent systems in autonomous trading, content moderation, or decision-making pipelines, these findings necessitate rethinking evaluation architectures. Single-evaluator systems now appear riskier than previously understood. The research validates that bias mitigation requires intentional redundancy and diversity, not merely larger models or better training data. Organizations implementing LLM-based evaluation chains should prioritize committee structures and heterogeneous evaluators.

Future research should examine whether these contagion patterns persist across different domains (financial analysis, content moderation, code review) and whether strategic model combinations could further suppress propagation. Real-world deployment testing remains essential before assuming laboratory results transfer to production systems.

• →Evaluator biases in LLM agents propagate between 15.7-35.2% across same-model networks, creating systematic error feedback loops.
• →Three-agent evaluation committees reduce bias contagion by 72.4% compared to single evaluators, offering a practical mitigation strategy.
• →Same-model agent contagion coefficients are 3-5x weaker than cross-model systems, suggesting architectural diversity impacts bias propagation.
• →The Contagion Networks framework enables quantitative measurement of bias spread through agent networks using spectral analysis.
• →Current multi-agent system designs may underestimate risks from systematic evaluator bias when using single-model evaluation architectures.