Recoleta
GitHub
Source note

CODESKILL: Learning Self-Evolving Skills for Coding Agents

Authors
Yanzhou Li, Yiran Zhang, Xiaoyu Zhang, Xiaoxia Liu, Yang Liu
Source
arXiv
Published
Collection
Software Intelligence
Coding AgentsSkill LearningProcedural MemoryReinforcement LearningSoftware Engineering Agents
Open arXivMarkdown source

Summary

CODESKILL trains a small LLM to turn coding-agent trajectories into reusable skills and keep a compact skill bank. The claimed gain is higher pass rate for frozen coding agents without changing the coding agent itself.

Problem

  • Coding agents leave long action traces while fixing repositories or working in terminals, but raw traces are too large and task-specific to reuse directly.
  • Existing skill and memory methods often use fixed prompts or hand-written update rules, so they can keep weak, redundant, or over-specific knowledge.
  • This matters because a better skill bank can improve future software-engineering tasks while leaving the main coding model frozen.

Approach

  • CODESKILL learns a skill-management policy from trajectories. It outputs operations that generate a new skill, revise an existing skill, add a candidate, merge it with another skill, or drop it.
  • Each skill is a markdown instruction with a title, trigger condition, and actionable steps for the coding agent.
  • The skill bank has task-level skills for broad repair workflows and event-driven skills for local events such as command failures, test errors, or repeated failure modes.
  • Training starts with supervised targets from teacher models, then uses GRPO reinforcement learning.
  • The RL reward combines rubric-based skill quality, verifier-based execution improvement, and an alignment score that checks whether the coding agent actually used the skill.

Results

  • With Qwen3.5-35B-A3B as the frozen coding policy, CODESKILL reached 39.26 average success across EnvBench-Python, EnvBench-Java, SWE-Bench Verified, and Terminal-Bench 2, compared with 29.57 for no-skill and 35.25 for the strongest prompt or memory baseline.
  • The same setting reduced average solved-instance steps to 35.15, compared with 44.12 for no-skill and 36.99 for the strongest prompt or memory baseline.
  • Per-benchmark success with Qwen3.5-35B-A3B was 18.60 on EnvBench-Python versus 6.98 no-skill, 38.32 on EnvBench-Java versus 27.10, 66.00 on SWE-Bench Verified versus 57.33, and 34.12 on Terminal-Bench 2 versus 25.88.
  • With GPT-5.4-mini as the frozen coding policy, CODESKILL reached 30.73 average success, compared with 21.80 for no-skill and 27.86 for the strongest prompt or memory baseline.
  • The abstract reports an average pass-rate gain of 9.69 points over no-skill and 4.01 points over the strongest prompt-based or memory baseline, equal to about 33% and 11% relative gains.
  • In the lifecycle ablation, the full system kept 676 skills, compared with 1252 skills for extraction-only and extraction-plus-evolution variants, while preserving 39.26 average success in the Qwen3.5-35B-A3B setting.

Link