A new framework quantifies bias in multi-agent LLM systems by simulating small echo chambers. Research shows significant stance shifts during discussions, particularly when agents are initialized with aligned conservative viewpoints, demonstrating that multi-agent interactions can exacerbate polarization.
HOW THIS AFFECTS YOU
●
researcherYou should consider multi-agent dynamics rather than isolated model evaluations when studying systemic bias.