EVENT DOSSIER
SkillRevise proposes an execution-based framework aimed at improving the agent skills generated by large language models (LLMs). This method uses execution evidence to diagnose skill deficiencies, retrieves repair principles from general memory, and applies execution-based editing for iterative revision. Evaluations on three main benchmarks, two domain-specific studies, and six LLMs show that SkillRevise significantly outperforms single-generation baselines, increasing the success rate of basic agents on SkillsBench from 36.05% to 61.63%. Moreover, the revised skills can be migrated across different executors and task environments, indicating that it captures reusable process knowledge beyond a single executor. The related code is available in the GitHub repository.
SkillRevise 提出了一种基于执行的框架，通过迭代修订初始技能以改进大语言模型（LLM）代理的技能。该方法利用执行证据诊断技能缺陷，从通用记忆中检索修复原则并应用执行锚定编辑，最终在修订预算内保留最佳观察到的技能。在三个主要基准、两个领域特定研究和六个 LLM 上的评估显示，SkillRevise 显著优于单次生成基线，将 SkillsBench 上基础代理的成功率从 36.05% 提升至 61.63%。此外，修订后的技能可在不同执行器和任务环境中迁移，表明其捕捉了超越单一执行器的可重用过程知识。相关代码已发布在 https://github.com/HKUST-KnowComp/skillrevise。