arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2604.01608cs.AI
收藏

从多智能体到单智能体:技能蒸馏何时有益?

From Multi-Agent to Single-Agent: When Is Skill Distillation Beneficial?

The Chinese University of Hong Kong(香港中文大学);LIGHTSPEED

Binyan Xu, Dong Fang, Haitao Li, Kehuan Zhang

AI总结:

研究探讨了技能蒸馏在多智能体系统到单智能体系统转换中的有效性,提出通过评估指标的拓扑刚性预测技能效用,并引入AdaSkill框架实现自适应蒸馏。

摘要:

多智能体系统(MAS)通过分配专业知识来处理复杂任务,但往往导致协调开销大、上下文碎片化和脆弱的阶段顺序。将MAS转化为单智能体技能可以避免这些成本,但缺乏明确的指导来决定何时和什么进行蒸馏。本文揭示技能效用并非由任务决定,而是由评估指标决定。引入Metric Freedom(F),首个先验预测技能效用的指标。F通过Mantel测试量化指标评分景观的拓扑刚性。基于F,提出AdaSkill框架,分为两个阶段:第一阶段作为选择性提取机制,提取工具和知识并丢弃限制性结构以保留探索。第二阶段对自由指标进行迭代优化。在4个任务、11个数据集和6个指标上评估,F强烈预测技能效用(r=-0.85,p<0.0001)。惊人的是,相同智能体轨迹在刚性和自由指标下产生相反的技能提升,证明技能效用本质上是指标层面的属性。驱动这一信号,AdaSkill在减少成本达8倍和延迟达15倍的同时,匹配或超过原始MAS性能。

英文摘要:

Multi-agent systems (MAS) for structured data-science tasks externalize analytical control through workflows spanning stages, tools, shared state, verification, and repair. Distilling such workflows into a single-agent skill can reduce orchestration overhead, but it remains unclear which workflow components should cross the control boundary. We distinguish capability resources, which expand what an agent can do, from pipeline guidance, which constrains which solutions it explores. On the same causal-estimation instances, adding task-qualified source pipeline guidance to a capability-matched skill changes normalized utility by +19.6 points under method-selection accuracy but -10.3 points under numerical error. To explain this reversal, we introduce Behavior-Outcome Freedom (F), a pre-synthesis diagnostic of signed behavior-outcome rank mismatch, and formalize its candidate-conditional role through Signed Anchor-Rank Transfer. Motivated by this mechanism, we propose AdaSkill, which preserves validated capability resources, removes runtime orchestration, and conditionally inherits pipeline guidance using a calibrated rule over F. Across 16 capability-matched interventions, the native-scale Full-minus-Discard effect decreases across the continuous F scale (r = -0.80, p < 0.001), while a 15-treatment atomic sweep localizes the reversal to pipeline guidance. Across 11 datasets spanning four structured data-science task families, AdaSkill combines strong task performance with substantially lower deployment overhead.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑