Coding agents gain reusable repair skills while their model stays fixed
Hierarchical Skill Evolution for Coding Agent · arXiv:2610.35616
In plain words
Coding agents solve programming problems by searching code, making edits, and testing them. Those steps can be useful again, but each attempt also contains details that fit only one issue or codebase.
Hierarchical Skill Evolution collects repeated repair steps across tasks and repositories, while keeping the coding model unchanged.
On two sets of repository repair tasks, agents using its skills solved more tasks on average than agents given no extra skills. The result covers the tested models and agent tools; it does not mean every task improved.
The research
A coding trajectory contains both reusable repair steps and details tied to one issue and repository. How can an agent learn from those attempts without carrying a solution over to a codebase where it may not apply?
Hierarchical Skill Evolution turns task outcomes into shared repair and recovery guidance by grouping evidence within repositories and across them.
To test whether repair guidance can transfer beyond the codebase where it was learned, the authors gave GLM-5.2 agents a frozen library of ten skills evolved from SWE-Gym trajectories.
They compared it with no extra skills and other skill methods on SWE-bench Verified and DeepSWE, where a task counts as resolved when its patch passes the test set’s verifier.
Across both test sets and the Pi and Claude Code harnesses, HSE had the highest mean resolved rate among the compared methods. This supports the claim that its SWE-Gym skills transfer to these test sets without target-set adaptation: agents using them resolved a larger share of tasks on average.
The static comparison tested GLM-5.2 with Pi and Claude Code, averaging three runs; aggregate gains do not mean every task improved. Target-task adaptation used trajectories from the same complete task sets later evaluated, so it does not establish performance on unseen target tasks.
What the evidence shows
- From cases to procedures
A coding agent’s trajectory records how it searched, edited code, and tested a fix. But the paths, files, and patch may fit only the original issue and repository.
The research asks how to retain useful repair procedures without treating one task’s solution as reusable guidance. This distinction matters because a record of what happened is not automatically a procedure that applies to another codebase.
- How HSE builds skills
Hierarchical Skill Evolution (HSE) is a method that turns coding-agent experience into reusable skills without changing the coding model. It first extracts evidence from individual task trajectories, then looks for recurring patterns within each repository, and finally combines patterns supported across repositories.
An external writer proposes skill candidates and an external evaluator checks their support, scope, and safety. Successful patterns can become repair procedures; recurring failures can become recovery guidance.
- Testing transfer
The authors tested transfer to SWE-bench Verified and DeepSWE, two sets of repository repair tasks. A task is resolved when its patch passes the set’s verifier, and resolved rate is the percentage of tasks that pass.
With GLM-5.2 under Pi and Claude Code, the frozen library learned on SWE-Gym had the highest mean resolved rate among compared methods in all four settings, and outperformed giving the agent no extra skills.
- Different skills matter
Removing either procedural or recovery skills lowered the mean resolved rate in every tested setting, but their relative contribution differed by task set. Removing procedural skills caused the larger loss on SWE-bench Verified; removing recovery skills caused the larger loss on DeepSWE.
On DeepSWE, the pack without recovery skills was only slightly above No Skills. The comparison suggests recovery guidance accounts for much of HSE’s measured advantage there.
- What the evidence supports
Controls also tested how HSE constructs the library. Shuffling the same training trajectories across repositories lowered rates, as did keeping the evidence hierarchy but freezing the initial writer and evaluator rules.
These comparisons support the role of both hierarchical grouping and rule updates within the tested procedure. Updating the transferred library with target-task trajectories raised measured rates further, but evaluation used the same complete task sets that supplied those trajectories, leaving performance on unseen target tasks unestablished.
Deep
- Original paper authors
- Jun Gao, Yun Peng, Changhai Zhou, Yuhan Zhu, Xiangyu Zeng
- Submitted to arXiv
- License
- CC BY 4.0