跳转至

从多智能体到单智能体：什么时候技能蒸馏才真正有效？

文章背景与核心概要

本文探讨了将多智能体系统（MAS）蒸馏为结构化数据科学任务的单智能体技能时的权衡问题。虽然单智能体蒸馏可以减少调度开销，但决定哪些工作流组件应该跨越控制边界长期以来一直是一个未解的挑战。为此，作者将工作流组件明确区分为“能力资源”（扩展智能体能力）与“pipeline引导”（约束解的探索）。为了解决性能差异问题，他们引入了“行为-结果自由度（\(F\)）”作为综合前诊断指标，并提出了 AdaSkill 框架——该框架在保留能力资源的同时，根据 \(F\) 有条件地继承 pipeline 引导。横跨四个任务族的 11 个数据集的实验表明，AdaSkill 成功实现了强劲的任务性能以及显著降低的部署开销。

Executive Summary

This paper investigates the trade-offs of distilling multi-agent systems (MAS) into single-agent skills for structured data-science tasks. While single-agent distillation reduces orchestration overhead, determining which workflow components should cross the control boundary has remained an open challenge. The authors distinguish capability resources (which expand agent capabilities) from pipeline guidance (which constrains solution exploration). To resolve performance discrepancies, they introduce Behavior-Outcome Freedom (\(F\)) as a pre-synthesis diagnostic and propose AdaSkill, a framework that preserves capability resources while conditionally inheriting pipeline guidance based on \(F\). Across 11 datasets spanning four task families, AdaSkill successfully achieves strong task performance alongside significantly lower deployment overhead.

元数据与出版信息

  • arXiv ID: arXiv:2604.01608 [cs.AI]
  • 学科分类: 人工智能 (cs.AI)
  • 作者: Binyan Xu, Dong Fang, Haitao Li, Kehuan Zhang
  • 提交时间: 2026年4月2日
  • 最后修订: 2026年8月19日（版本 v5）
  • 引用格式: Xu et al., 2026. From Multi-Agent to Single-Agent: When Is Skill Distillation Beneficial?. arXiv:2604.01608 [cs.AI].

Metadata & Publication Information

  • arXiv ID: arXiv:2604.01608 [cs.AI]
  • Subjects: Artificial Intelligence (cs.AI)
  • Authors: Binyan Xu, Dong Fang, Haitao Li, Kehuan Zhang
  • Submitted: April 2, 2026
  • Last Revised: August 19, 2026 (Version v5)
  • Citation: Xu et al., 2026. From Multi-Agent to Single-Agent: When Is Skill Distillation Beneficial?. arXiv:2604.01608 [cs.AI].

摘要

针对结构化数据科学任务的多智能体系统（MAS）通过跨越不同阶段、工具、共享状态、验证和修复的工作流外部化了分析控制。将这些工作流蒸馏为单智能体技能可以减少调度开销，但目前尚不清楚哪些工作流组件应该跨越控制边界。

我们区分了“能力资源”（扩展智能体可执行的操作）与“pipeline引导”（约束智能体探索哪些解决方案）。在相同的因果估计实例中，将在方法选择准确率下表现合格的源 pipeline 引导添加至能力匹配的技能中，会使归一化效用发生变化：方法选择准确率提高 +19.6 点，而数值误差则恶化 -10.3 点。

为了解释这种性能反转，我们引入了行为-结果自由度（Behavior-Outcome Freedom, \(F\)），这是一种用于预综合诊断的有符号行为-结果排序不匹配指标，并通过有符号锚点-排序迁移（Signed Anchor-Rank Transfer）将其候选条件角色形式化。受此机制启发，我们提出了 AdaSkill 框架，该框架能够保留经过验证的能力资源、消除运行时调度，并利用基于 \(F\) 的校准规则有条件地继承 pipeline 引导。

在 16 项能力匹配的干预实验中，原尺度下的“全量减去丢弃（Full-minus-Discard）”效应随着连续 \(F\) 值的增加而减小（\(r = -0.80, p < 0.001\)），而一项包含 15 种处理的原子消融实验将这种反转定位到了 pipeline 引导上。在横跨四个结构化数据科学任务族的 11 个数据集上，AdaSkill 成功将强大的任务性能与大幅降低的部署开销结合在了一起。

Abstract

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.

核心概念与贡献

  1. 资源与引导的区分（Resource vs. Guidance Distinction）： 将工作流组件分类为能力资源（智能体可以执行的操作）和pipeline引导（如何约束/指导解决方案）。
  2. 行为-结果自由度（\(F\)）： 一种新型的预综合诊断指标，旨在捕捉有符号的行为-结果排序不匹配。
  3. 有符号锚点-排序迁移（Signed Anchor-Rank Transfer）： 一项形式化机制，解释了为什么 pipeline 引导能够同时提高方法选择准确率并降低数值误差。
  4. AdaSkill 框架： 一种精简的蒸馏方法，既能保留有价值的能力，又能利用基于 \(F\) 的校准规则动态、有条件地应用 pipeline 引导。

Key Concepts & Contributions

  1. Resource vs. Guidance Distinction: Categorizes workflow components into capability resources (what an agent can execute) and pipeline guidance (how solutions are constrained/directed).
  2. Behavior-Outcome Freedom (\(F\)): A novel pre-synthesis diagnostic metric designed to capture signed behavior-outcome rank mismatch.
  3. Signed Anchor-Rank Transfer: A formal mechanism explaining why pipeline guidance can simultaneously improve method-selection accuracy while degrading numerical errors.
  4. AdaSkill Framework: A streamlined distillation approach that preserves valuable capabilities while dynamically and conditionally applying pipeline guidance using a calibrated rule over \(F\).

全文与访问链接

Full-Text & Access Links