文章背景与核心概要
大语言模型（LLM）智能体在软件工程任务中展现出了巨大潜力，但其能力的自主演进过程长期面临着局部更新带来的过拟合和泛化能力不足问题。现有的方法通常将技能更新视为孤立的事件，忽视了技能之间的关联与系统的一致性，从而导致在复杂编程任务中表现不稳定。
为了解决这一技术瓶颈，本文提出了 GSE（Globalized Skill Evolution，全局化技能演进） 框架。该框架通过构建技能关系图（SRG）来维护整个技能库的连贯性，利用基于聚类的技能整合技术从单次更新中提取可重用能力，并采用回放驱动的验证机制来防止行为退避和过拟合。在 OpenHands 和 mini-SWE-agent 上的实验表明，GSE 在诸如暴露缺陷的测试用例生成以及误报缺陷报告过滤等软件工程任务中，显著超越了现有的演进技术。
Learning Globally Reusable Skills for Coding Agents
arXiv: 2608.06153
Date: August 6, 2026
Authors: Chen Yang, Jiashuo Tian, Ziqi Wang, Xinyin Liu, Meiru Ye, Junjie Chen
Learning Globally Reusable Skills for Coding Agents
arXiv: 2608.06153
Date: August 6, 2026
Authors: Chen Yang, Jiashuo Tian, Ziqi Wang, Xinyin Liu, Meiru Ye, Junjie Chen
Summary
The paper introduces GSE (Globalized Skill Evolution), a novel framework designed to improve how Large Language Model (LLM) agents evolve their capabilities. Current methods often treat skill updates as isolated, local changes, which frequently leads to overfitting and poor generalization. GSE addresses these limitations by:
- Modeling Inter-skill Relationships: Utilizing a Skill Relation Graph (SRG) to maintain consistency and coherence across the entire skill bank.
- Enhancing Generalization: Implementing cluster-based skill consolidation to extract reusable capabilities from individual updates.
- Ensuring Reliability: Employing replay-driven verification to mitigate behavioral regressions and prevent overfitting.
Empirical evaluations on OpenHands and mini-SWE-agent demonstrate that GSE significantly outperforms existing evolution techniques in software engineering tasks, including bug-revealing test generation and false-positive bug report filtering.
Summary
The paper introduces GSE (Globalized Skill Evolution), a novel framework designed to improve how Large Language Model (LLM) agents evolve their capabilities. Current methods often treat skill updates as isolated, local changes, which frequently leads to overfitting and poor generalization. GSE addresses these limitations by:
- Modeling Inter-skill Relationships: Utilizing a Skill Relation Graph (SRG) to maintain consistency and coherence across the entire skill bank.
- Enhancing Generalization: Implementing cluster-based skill consolidation to extract reusable capabilities from individual updates.
- Ensuring Reliability: Employing replay-driven verification to mitigate behavioral regressions and prevent overfitting.
Empirical evaluations on OpenHands and mini-SWE-agent demonstrate that GSE significantly outperforms existing evolution techniques in software engineering tasks, including bug-revealing test generation and false-positive bug report filtering.
Key Performance Highlights
- Test Generation: Improved precision by 6.1%–34.1% and recall by 31.8%–180.0%.
- False-Positive Filtering: Improved precision by 15.4%–96.4% and recall by 13.1%–19.8%.
- Industrial Impact: A 61.4% improvement in F1-score when deployed on an internal industrial agent.
Key Performance Highlights
- Test Generation: Improved precision by 6.1%–34.1% and recall by 31.8%–180.0%.
- False-Positive Filtering: Improved precision by 15.4%–96.4% and recall by 13.1%–19.8%.
- Industrial Impact: A 61.4% improvement in F1-score when deployed on an internal industrial agent.
Access & Resources
Access & Resources
Metadata
- Primary Subject: Software Engineering (cs.SE)
- Secondary Subject: Artificial Intelligence (cs.AI)
- DOI: https://doi.org/10.48550/arXiv.2608.06153
Metadata
- Primary Subject: Software Engineering (cs.SE)
- Secondary Subject: Artificial Intelligence (cs.AI)
- DOI: https://doi.org/10.48550/arXiv.2608.06153