谁说服了谁？追踪多智能体系统中LLM如何改变想法
原标题：Who Convinced Whom? Tracing How LLMs Change Their Minds in Multi-Agent Systems
AI 摘要
该论文研究多智能体辩论（MAD）中同伴交流如何影响LLM个体决策与集体结果，提出MAD Bench基准（2400题，覆盖知识推理、价值观道德判断、安全拒答），并用轮次匹配的独立反思作为对照评估三个开放权重模型。研究发现辩论中的答案改变常在生成理由前已确定，同伴推理比单纯答案影响更大，来源可靠性线索强于专业身份，注意力掩码显示影响路径依赖架构；辩论相对独立投票仅带来约1个百分点的小幅准确率提升，但一致错误率增加3.1-7.4个百分点。结论是更强共识可能与更多共享错误并存，需在准确率之外评估依赖性与一致错误。
正文节选
Who Convinced Whom? Tracing How LLMs Change Their Minds in Multi-Agent Systems Abstract Multi-agent debate (MAD) aims to improve language-model decisions through the exchange of arguments, yet agreement after debate need not reflect independent corroboration. We investigate how peer communication influences individual decisions and collective outcomes through behavioral and internal interventions. We introduce MAD Bench, a 2,400-question benchmark spanning knowledge and reasoning, values and mor