LLM Fault Localisation within Evolutionary Computation Based Automated Program Repair
S. Murtaza, Aidan Mccoy, Zhiyuan Ren, Aidan Murphy, Wolfgang Banzhaf · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Repairing bugs can be a daunting task for even a human experienced in debugging, so naturally, attempting to automatically repair programs with a computer system is quite challenging. The existing methods of automated program repair leave a lot of room for improvement. Fault localization, which aims to find lines of code that are potentially buggy, minimises the search space of an automated program repair system. Recent work has shown improvement in these fault localization methods, with the use of Large Language Models. Here, we propose a system where a LLM-based fault localization tool, which we call SemiAutoFL, is used within a fully automatic program repair program, ARJA-e. We show that utilising LLM-based fault localization with ARJA-e can significantly improve its performance on real world bugs. ARJA-e with SemiAutoFL can repair 10 bugs that ARJA-e was previously unable to so do. This finding adds to our understanding of how to improve fault localization and automated program repair, highlighting the potential for more efficient and accurate fault localisation methods being applied to automated program repair.