PATCHAGENT: A Practical Program Repair Agent Mimicking Human Expertise

Zheng Yu, Ziyi Guo, Yuhang Wu, Jiahao Yu, Meng Xu, Dongliang Mu, Yan Chen, Xinyu Xing

34th USENIX Security Symposium · Day 1

The increasing complexity of modern software has led to a proliferation of vulnerabilities, posing significant challenges for developers to identify and fix them promptly. Automated Program Repair (APR) techniques aim to alleviate this burden by autonomously triaging and patching software bugs. While recent advancements in **Large Language Models (LLMs)** have shown considerable promise in generating code patches, most existing LLM-based APR tools focus primarily on patch generation, often assuming perfect fault localization or lacking robust patch validation mechanisms. This fragmented approach prevents them from providing a truly end-to-end solution for real-world vulnerability repair.

AI review

Solid engineering work that actually moves the needle on automated vuln repair. The 92% success rate on real CVEs is impressive, and the ablation study proves the interaction optimizations aren't just window dressing. Not revolutionary—it's LLM orchestration, not a new exploitation primitive—but it's the kind of rigorous systems work that practitioners can actually use.