QiMeng-PRepair: Edit-Aware Reward Optimization Solves LLM Over-Editing in Code Repair
arXiv·medium signal
LLMs achieve strong program repair performance but suffer from over-editing — excessive modifications that overwrite correct code and hinder bug localization. QiMeng-PRepair systematically quantifies this impact and introduces precise repair via edit-aware reward optimization, teaching models to make minimal, targeted fixes. Directly applicable to anyone using LLMs for automated code repair in CI/CD pipelines.