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

The paper introduces Contagion Networks, a framework for quantifying the propagation of evaluator biases in multi-agent systems using large language models (LLMs). In experiments with a three-agent setup employing DeepSeek-chat, biases were measured through the Cross-Agent Contagion Matrix, revealing propagation coefficients between 0.157 and 0.352, with homogeneous models exhibiting 3-5x weaker contagion than cross-model interactions. The study also presents a mitigation strategy by increasing evaluator committee size, which can reduce effective contagion by 72.4%, and releases the Contagion Network framework as open-source, providing valuable insights for practitioners managing bias in LLMs.