AtomSD: Procedural Atom-Guided Self-Distillation for LLM Agents

Computer Science > Computation and Language arXiv:2610.26926 (cs) [Submitted on 17 Oct 2026] Title: AtomSD: Procedural Atom-Guided Self-Distillation for LLM Agents Authors: Bohan Yu, Pengfei Cao, Kai Tian, Chen Han, Shi-Yang Li, Chenxi Zhou, Junlin Liu, Cao Liu, Ke Zeng, Jun Zhao, Kang Liu.

计算机科学 > 计算与语言 arXiv:2610.26926 (cs) [提交于 2026 年 10 月 17 日] 标题：AtomSD：面向大语言模型智能体的程序化原子引导自蒸馏。作者：Bohan Yu, Pengfei Cao, Kai Tian, Chen Han, Shi-Yang Li, Chenxi Zhou, Junlin Liu, Cao Liu, Ke Zeng, Jun Zhao, Kang Liu。

Abstract: Large language models (LLMs) possess broad factual knowledge and strong reasoning capabilities, but solving complex tasks also requires procedural knowledge about how to act. Existing approaches often represent procedural knowledge through trajectories or bundled skills, limiting flexible reuse across interaction histories.

摘要：大语言模型（LLMs）拥有广泛的事实知识和强大的推理能力，但解决复杂任务还需要关于如何行动的程序性知识。现有的方法通常通过轨迹或捆绑技能来表示程序性知识，这限制了在交互历史中进行灵活重用的能力。

This paper introduces Procedural Atoms, fine-grained and independently applicable units that represent Intent, When, Do, and Avoid guidance. Building on this representation, we propose AtomSD, an Atom-guided Self-Distillation framework in which the learner continually extracts Procedural Atoms from its own successful trajectories and distills them into the evolving policy through Atom-conditioned supervision and task-reward optimization.

本文引入了“程序化原子”（Procedural Atoms），这是一种细粒度且可独立应用的单元，用于表示意图（Intent）、时机（When）、执行（Do）和规避（Avoid）指导。基于这种表示，我们提出了 AtomSD，这是一个原子引导的自蒸馏框架，学习者通过原子条件监督和任务奖励优化，不断从自身的成功轨迹中提取程序化原子，并将其蒸馏到不断演进的策略中。

Comprehensive experiments show that AtomSD consistently outperforms strong reward-based and distillation methods, exceeding the state-of-the-art baseline by up to 9.31 points in average performance. Further analyses demonstrate that the proposed Procedural Atoms provide effective guidance, highlighting the value of atom-level procedural knowledge for policy learning.

综合实验表明，AtomSD 在性能上持续优于强大的基于奖励和蒸馏的方法，平均性能比当前最先进的基准高出 9.31 个百分点。进一步分析证明，所提出的程序化原子提供了有效的指导，突显了原子级程序性知识对于策略学习的价值。