Reproduce “LLM-Driven Kernel Evolution: Automating Driver Updates in Linux” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Arina Kharlamova, Jiawen Liu, Tianyi Zhang, Xinrui Yang, Humaid Alqasimi, Youcheng Sun, Chun Jason Xue
AUTODRIVER is an LLM-driven system designed to automate Linux kernel driver updates by synthesizing compatible patches, evaluated with DRIVEBENCH, an executable corpus for reproducible co-evolution research. The system achieved a 56.4% compilation success rate for patches, with the DeepSeek LLM backend generating higher quality and more accurate patches than ChatGPT.
Key insights
- Large Language Models can effectively automate tasks in safety-critical, low-level systems when grounded in structured, executable data and augmented with robust, closed-loop validation mechanisms.
- An Abstract Syntax Tree (AST)-based static scoring metric, derived using Tree-sitter, provides a precise, language-aware measure of patch quality that correlates strongly with human judgment and functional stability.
- Iterative refinement, incorporating compiler diagnostics, significantly improves LLM patch generation, with most recoverable errors being resolved within the first five feedback loops.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.