BoostAPR: Enhancing Automated Program Repair with Dual Reward Reinforcement Learning
Researchers have introduced BoostAPR, a novel three-stage framework designed to improve automated program repair (APR) using execution-grounded reinforcement learning. Addressing the limitations of sparse feedback and coarse rewards in existing methods, BoostAPR employs supervised fine-tuning on verified demonstrations, trains dual reward models for sequence-level assessment and line-level credit allocation, and utilizes PPO optimization to focus on critical code edits. This approach allows for precise reward distribution at an intermediate granularity suited for code changes. Evaluated across four major benchmarks, the model demonstrates significant performance improvements, achieving a 40.7% success rate on SWE-bench Verified, which is a 22.9 percentage point increase over the base model. It also recorded strong results on Defects4J, HumanEval-Java, and QuixBugs, highlighting its competitive standing among open-source models and its robust cross-language generalization capabilities. The study, published on arXiv, represents a significant advancement in applying AI to software engineering tasks, specifically in bug fixing and code maintenance.
Wire timeline
BoostAPR: Enhancing Automated Program Repair with Dual Reward Reinforcement Learning
Researchers have introduced BoostAPR, a novel three-stage framework designed to improve automated program repair (APR) using execution-grounded reinforcement learning. Addressing the limitations of sparse feedback and coarse rewards in existing methods, BoostAPR employs supervised fine-tuning on verified demonstrations, trains dual reward models for sequence-level assessment and line-level credit allocation, and utilizes PPO optimization to focus on critical code edits. This approach allows for precise reward distribution at an intermediate granularity suited for code changes. Evaluated across four major benchmarks, the model demonstrates significant performance improvements, achieving a 40.7% success rate on SWE-bench Verified, which is a 22.9 percentage point increase over the base model. It also recorded strong results on Defects4J, HumanEval-Java, and QuixBugs, highlighting its competitive standing among open-source models and its robust cross-language generalization capabilities. The study, published on arXiv, represents a significant advancement in applying AI to software engineering tasks, specifically in bug fixing and code maintenance.
cs.AI updates on arXiv.org