CODESKILL: An LLM-Based Framework for Self-Evolving Coding Agents
A recent study presents CODESKILL, a framework designed for coding agents to develop and enhance skills through their own task-solving experiences. In contrast to traditional approaches that depend on static prompts and heuristic guidelines, CODESKILL employs a management policy based on LLM and trained via reinforcement learning. It derives multi-granularity procedural skills from agent trajectories, refines them with fresh experiences, and keeps a concise skill repository for upcoming tasks. The hybrid reward system integrates both dense rubric-based evaluations of skill quality and sparse, verifiable execution signals. This research can be accessed on arXiv with the identifier 2605.25430.
Key facts
- CODESKILL is an LLM-based framework for coding agent self-evolution.
- It extracts multi-granularity procedural skills from coding-agent trajectories.
- Skills are evolved with new experience and maintained in a compact skill bank.
- The framework is trained with reinforcement learning.
- Reward uses dense rubric-based skill-quality feedback and sparse verifiable execution signals.
- Existing methods rely on fixed prompts and heuristic update rules.
- The paper is published on arXiv with ID 2605.25430.
- The approach reformulates skill extraction and maintenance as a learnable management policy.
Entities
Institutions
- arXiv