# CODESKILL: Learning Self-Evolving Skills for Coding Agents

Source: [CODESKILL: Learning Self-Evolving Skills for Coding Agents](<https://www.sota2.com/research/paper/6a152968dbd7150506f7e373>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-27T05:00:13.467Z

Authors: Yanzhou Li, Yiran Zhang, Xiaoyu Zhang, Xiaoxia Liu, Yang Liu

Published: 2026-05-25T05:12:49.000Z

## Abstract

Coding agents produce rich trajectories while solving software-engineering tasks. To enable agent self-evolution, these trajectories can be distilled into reusable procedural skills that compactly encode experience to guide future behavior. However, existing skill construction and maintenance methods often rely on fixed prompts and heuristic update rules, leaving it unclear how knowledge should be selected, abstracted, and maintained to best serve downstream agents. We propose CODESKILL, an LLM-based framework that reformulates skill extraction and skill-bank maintenance as a learnable management policy. CODESKILL extracts multi-granularity procedural skills from coding-agent trajectories, evolves skills with new experience, and maintains a compact skill bank for future task solving. We train CODESKILL with reinforcement learning, using a hybrid reward that combines dense rubric-based skill-quality feedback with sparse verifiable execution feedback from the frozen downstream agent. Experiments on EnvBench, SWE-Bench Verified, and Terminal-Bench 2 show that CODESKILL improves average pass rate by 9.69 over the no-skill baseline and by 4.01 over the strongest prompt-based or memory baseline, while maintaining the skill bank at a stable size during iterative construction.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Environment setup and dependency repair on EnvBench-Python (held-out)](<https://www.sota2.com/research/sota/environment-setup-and-dependency-repair-on-envbench-python-held-out.md>): Environment setup and dependency repair; EnvBench-Python (held-out); Success Rate: 18.6.

- [Environment setup and dependency repair on EnvBench Java (held-out)](<https://www.sota2.com/research/sota/environment-setup-and-dependency-repair-on-envbench-java-held-out.md>): Environment setup and dependency repair; EnvBench Java (held-out); Task Pass Rate: 38.32.

- [Repository-level issue resolution on SWE-Bench Verified (held-out (sample of 150))](<https://www.sota2.com/research/sota/repository-level-issue-resolution-on-swe-bench-verified-held-out-sample-of-150.md>): Repository-level issue resolution; SWE-Bench Verified (held-out (sample of 150)); Task Pass Rate: 66.

- [Terminal-based problem solving on Terminal-Bench 2 (out-of-distribution)](<https://www.sota2.com/research/sota/terminal-based-problem-solving-on-terminal-bench-2-out-of-distribution.md>): Terminal-based problem solving; Terminal-Bench 2 (out-of-distribution); Task Success Rate: 34.12.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.25430>)

[PDF](<https://arxiv.org/pdf/2605.25430.pdf>)