Not All Flips Are Conformity: Decomposing Stance Convergence in Multi-Agent LLM Debate
This study decomposes answer changes in multi-agent debate into three sources: spontaneous instability, stance-induced conformity, and reasoning-induced persuasion. It found that spontaneous change is a large baseline, while strict conformity is harmful. The research suggests that reasoning presentation alone can cause error adoption without valid logic, and provides methods to predict and mitigate harmful peer influence.