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

Evaluator biases in large language models can spread rapidly through multi-agent systems, compromising their overall performance. Research has shown that when these models serve as evaluators, their systematic biases propagate across interacting agents, creating a network effect that amplifies existing flaws. The Contagion Networks framework provides a formal method for measuring the spread of evaluator biases, allowing for a deeper understanding of this phenomenon. In an experiment using DeepSeek-chat, three distinct evaluator bias profiles were used to demonstrate the propagation of biases across a 3-agent network. The results highlight the importance of addressing evaluator biases in multi-agent systems to prevent the amplification of errors. This matters to practitioners because unmitigated evaluator biases can lead to significant performance degradation in AI systems, making it essential to develop strategies for mitigating these biases¹.