An Empirical Study of Group Conformity in Multi-Agent Systems
Prior research lacks systematic investigation into social bias propagation and conformity behavior in large language model (LLM) agents, particularly regarding emergent group-level conformity in socially contentious debates. Method: We design a multi-agent debate framework simulating over 2,500 debates on controversial topics, integrating logistic regression, ANOVA, and dynamic stance tracking to quantify agent capabilities and model neutral agent stance evolution. Contribution/Results: We首次 demonstrate that LLM agents spontaneously exhibit human-like conformity without explicit training; agent intelligence—not agent count—dominates stance convergence; 72% of neutral agents align with the majority stance; bias propagation speed reaches 3.8× that observed in human experiments. We propose “agent influence” as a novel metric for quantifying behavioral impact, providing both theoretical grounding and empirical evidence for modeling LLM social behavior and advancing controllable, governance-oriented AI systems.