Study Reveals Bias Shifts in Multi-Agent LLM Conversations

A new framework from arXiv preprint 2501.14844 quantifies biases in multi-agent systems of conversational Large Language Models (LLMs), focusing on how biases emerge when models interact rather than in isolation. The researchers simulated small echo chambers where pairs of LLMs, initialized with aligned perspectives on polarizing topics, engaged in discussions. Contrary to expectations, significant shifts in stance were observed, particularly within echo chambers where all agents initially expressed conservative viewpoints. The study addresses a gap in bias detection methodologies, which typically consider models in isolation and neglect contextual applications. The framework aims to reduce potential risks associated with deploying generative models in critical settings by detecting biases in generated text. The findings highlight the need to study biases in multi-agent contexts, as they may differ from single-model outputs. The preprint, announced as a replace-cross, was published on arXiv with the identifier 2501.14844.