Code Researcher: AI Agent for Linux Kernel Crash Patching
A new LLM-based agent called Code Researcher has been developed to patch crashes in systems code, specifically targeting the Linux kernel. The agent performs multi-step reasoning over code semantics, patterns, and commit history to retrieve relevant context before generating patches. Evaluated on the kBenchSyz benchmark, Code Researcher significantly outperforms strong baselines in fixing Linux kernel crashes. The research, published on arXiv under ID 2506.11060, highlights the underexplored application of coding agents to complex systems codebases.
Key facts
- Code Researcher is an LLM-based deep research agent for code.
- It is designed to generate patches for crashes in systems code.
- The agent uses multi-step reasoning over code semantics, patterns, and commit history.
- It was evaluated on kBenchSyz, a benchmark of Linux kernel crashes.
- Code Researcher significantly outperforms strong baselines.
- The research is published on arXiv (ID 2506.11060).
- The work addresses the challenge of making changes to large systems codebases.
- It is inspired by recent progress on deep research agents.
Entities
Institutions
- arXiv