Programs improve evolving ai agent skills beyond text descriptions
Program-Aided Patchable Skill Evolution
Computation and LanguageArtificial Intelligence
Summary
Language models can learn skills to solve complex tasks without changing their underlying code by using reusable knowledge. The authors found that representing these skills as editable computer programs, not just text descriptions, makes the skills easier to update and improve. Their method, called PASE, evolves skills by modifying programs associated with them, which leads to better stability and generalization. Tests show PASE performs better than previous methods by a few percentage points. This highlights that using programs helps robots or AI agents learn and adapt more effectively.
What this means in practice
- •For ai system developers: Create adaptive agent skills that are easier to maintain by updating executable programs rather than natural language text.
- •For software engineers building automation tools: Enhance automation by evolving reusable procedural knowledge through patchable code components reflecting agent experience.
Authors
Jiahao Wang, Yuqi Xiong, Qing Yang, Zhipeng Xu, Zhenghao Liu, Zulong Chen, Yubin Bao
Abstract
Agent skills have emerged as an effective means of enabling LLM agents to solve complex, interactive tasks by encoding reusable knowledge without modifying model parameters. Recent approaches further improve skill adaptation by distilling reusable experience from agent execution trajectories and evolving skills based on downstream task feedback. However, existing methods primarily represent skills as natural language documents, intertwining procedural logic with descriptive knowledge and making skills difficult to maintain and evolve. To address this limitation, we propose PASE (Program-Aided Patchable Skill Evolution), a program-aided framework that iteratively evolves skills by associating them with executable programs. Unlike prior approaches that encode trajectory-derived experience primarily in natural language, PASE separates reusable procedural logic into editable programs, enabling targeted updates to executable components while reducing the need to repeatedly revise natural language skills. Extensive experiments across diverse models and benchmarks demonstrate that PASE consistently outperforms all baselines, by achieving over 2.8% improvements. Further analyses show that PASE exhibits greater evolutionary stability and generalisation capability, highlighting the critical role of programmatic knowledge in skill evolution. We will release all data and code via GitHub.