等推理成本下多智能体结构未超越单一冻结智能体
原标题：At Equal Inference Cost, Multi-Agent Structure Does Not Beat a Single Frozen Agent
AI 摘要
该研究在固定总推理调用数的条件下，比较了多智能体团队与单一智能体的性能。作者提出MA-Evolve方法，在冻结的7B模型上进化Planner-Executor-Critic团队的角色提示词。实验表明，在ALFWorld任务中，团队虽达到最高平均成功率，但与单一智能体无显著差异，且额外调用未带来收益。价值分解显示所有提升均来自执行者角色，规划者和批评者提示词退化为空。结论是，在等推理成本下，多智能体结构并未超越单一冻结智能体。
正文节选
At Equal Inference Cost, Multi-Agent Structure Does Not Beat a Single Frozen Agent Abstract Multi-agent pipelines of large language models, in which a planner, an executor, and a critic collaborate, routinely report gains over a single agent and have become a dominant design pattern for complex task completion. These gains are almost always obtained at greater inference, since a team issues several model calls per environment step where a single agent issues one. Existing automated methods searc