Skip to content
arXiv· arXiv:2608.22275· Han Zheng, Rafaila Galanopoulou, Ilia Shumailov, Tianqi Fan, Aiden Hall, Dominik Maier, Mathias Payer·Indexed by this site · Originally published

CodeMechanic: Defect-Property-Guided Program Mitigation

CodeMechanic: Bug-Property-Guided Program Mitigation

AI summary

On 101 real ARVO defects, CodeMechanic's first attempt generates 47.6% more plausible patches than the best baseline, with 91% fewer tokens. It uses defect properties to guide an LLM to reconstruct spatial memory safety properties from crashes, validate pointer and buffer bounds, and insert fail-stop guardrails before dangerous accesses, trading availability for safety. Manual audit shows that its patches semantically equivalent to developer fixes are 3.4–4.3 times those of the baseline.

Read originalarxiv.org