VCE-Skill: Enhancing Skill Self-Evolution with Version-Change Experience
- ID: 897b7984
- 原文链接: https://arxiv.org/abs/2608.16544
- PDF: https://arxiv.org/pdf/2608.16544
- 作者: Jianming Chen, Xuanbin Ye, Yawen Wang, Junjie Wang, Qing Wang, Fanjiang XU
- 日期: 2026-08-17
- 更新: 2026-08-17
- 分类: learning
- 来源类型: paper
- 标签: skill-evolution, self-improvement, agents, arxiv
- 质量评分: 4/5
- 抓取时间: 2026-08-19T04:47:40Z
中文导读
现有技能自演化方法只用当前任务的执行轨迹修订技能，公共技能版本历史中积累的演化知识基本未被利用作者试点研究揭示两类来源互补：公开技能变更提供可复用的演化先验，轨迹提供贴合当前任务的证据VCE-Skill 把噪声大实现相关的公开技能变更蒸馏为结构化可复用的'版本变更经验'，并与基础演化器的轨迹派生提案自适应融合实验显示技能自演化平均提升 3.20-4.98 分，迁移实验进一步表明所得技能跨模型迁移表现更强，把公开版本变更确立为一种此前被低估的先验知识来源
为什么值得关注
技能自演化新先验：蒸馏公开技能版本变更史，均值提升 3.2-5.0 分且跨模型迁移更强
关键信息
- 论文标题: VCE-Skill: Enhancing Skill Self-Evolution with Version-Change Experience
- 作者: Jianming Chen, Xuanbin Ye, Yawen Wang, Junjie Wang, Qing Wang, Fanjiang XU
- arXiv: https://arxiv.org/abs/2608.16544
- 发布时间: 2026-08-17
- arXiv 分类: cs.MA, cs.AI
- 关联标签: skill-evolution, self-improvement, agents, arxiv
English Abstract
Agents increasingly rely on reusable skills to encode task knowledge, tool-use procedures, and validation rules. Existing skill self-evolution methods primarily revise skills using execution trajectories collected from current tasks, leaving the evolution knowledge accumulated in public skill version histories largely untapped. Our pilot study reveals a clear complementarity between the two sources: public skill changes provide reusable evolution priors, whereas trajectories provide evidence grounded in the current task. Motivated by this, we propose VCE-Skill, which distills noisy and implementation-specific public skill changes into reusable, structured version-change experience and adaptively fuses it with trajectory-derived proposals from the base evolver, thereby exploiting external experience while retaining task-specific evidence. Extensive experiments demonstrate that VCE-Skill improves skill self-evolution, increasing mean scores by 3.20--4.98 points; transfer experiments further show that the resulting skills achieve stronger cross-model transfer performance. Our work highlights public skill version changes as a previously underexplored yet effective source of prior knowledge and advances trajectory-driven skill self-evolution.
English Summary
Agents increasingly rely on reusable skills to encode task knowledge, tool-use procedures, and validation rules, but existing skill self-evolution methods revise skills only from execution trajectories of current tasks, leaving the evolution knowledge accumulated in public skill version histories untapped. A pilot study reveals complementarity: public skill changes provide reusable evolution priors while trajectories provide evidence grounded in the current task. VCE-Skill distills noisy, implementation-specific public skill changes into structured, reusable version-change experience and adaptively fuses it with trajectory-derived proposals from the base evolver. Experiments show 3.20-4.98 point mean score improvements, and transfer experiments show the resulting skills achieve stronger cross-model transfer.
Obsidian Notes
- 内容由
opencli arxiv paper拉取 arXiv 元数据与摘要生成。 - 中文导读与价值判断均锚定在条目已有摘要、论文摘要、作者、日期与分类信息上；未补充论文摘要之外的实验细节。
- metadata source: opencli