Contagion Networks: evaluator bias propagates through multi-agent LLM systems

arXiv·medium signal

This arXiv paper introduces 'Contagion Networks,' showing that when LLMs serve as evaluators inside multi-agent systems, their systematic evaluation biases propagate through the agent network rather than staying local. The work formalizes how a single biased judge can contaminate downstream agent decisions, a direct risk for LLM-as-judge pipelines and agent self-improvement loops. For builders running multi-agent orchestration, it argues for diversifying or auditing evaluator agents to avoid systemic bias amplification.