Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 26·5 min read
CODESKILL: Learning Self-Evolving Skills for Coding Agents
TL;DR
Existing methods for coding agents struggle with skill construction and maintenance, limiting their adaptability. CODESKILL is a framework that uses large language models (LLMs) to extract and evolve procedural skills from coding trajectories.
✦ Why It Matters
Engineers can leverage CODESKILL to enhance the adaptability and efficiency of coding agents in software development tasks.
Key Takeaways
- Implement CODESKILL for dynamic skill management in coding agents.
- Utilize reinforcement learning to evolve agent skills effectively.
- Focus on maintaining a compact skill bank for efficiency.
- Explore hybrid feedback mechanisms to improve skill quality.
Full Summary
Coding agents generate complex behavior while solving software engineering tasks, but current methods for managing their skills often rely on static rules and prompts. CODESKILL addresses this by introducing a learnable management policy that extracts multi-granularity procedural skills from the agents' experiences.
It evolves these skills using reinforcement learning, combining feedback from skill quality assessments and execution results. In experiments conducted on EnvBench, SWE-Bench Verified, and Terminal-Bench 2, CODESKILL demonstrated a 9.69% increase in average pass rates over baseline methods without skills, and a 4.01% improvement over the best existing prompt-based approaches.
Importantly, CODESKILL maintains a compact skill bank, ensuring efficiency in skill management. This advancement suggests a more dynamic and effective approach to skill evolution in coding agents.