SEED-APR: A closed-loop self-evolving framework for automated program repair

Current automated program repair (APR) methods that employ large language models frequently encounter issues such as patch overfitting, constrained semantic generalization, and stagnant evolution due to static test-case evaluation and single-path generation. In order to address these issues, a self-evolving patch generation framework, termed SEED-APR, is proposed. The integration of runtime feedback with semantic diffusion is a key feature of this framework. The approach employed in this study involves the utilization of a Code Semantic Diffusion Model (CodeSDM) to generate diverse, semantically equivalent patches via denoising diffusion in latent space. The Runtime Feedback Score (RF-score) is derived from lightweight sandbox execution and is used to evaluate patches against real-world behavioral constraints. The model autonomously refines its repair strategy through LoRA-based self-evolution, using contrastive learning to convert runtime feedback into gradient updates. This process establishes a closed loop of generation, evaluation and optimization. Experiments on Defects4J v2.0 and Bugs.jar demonstrate that SEED-APR attains repair accuracy of 58.7% and 55.2%, respectively, accompanied by 56.8% zero-shot cross-project accuracy and 0.78 semantic diversity. This performance has been demonstrated to be significantly superior to that of existing methods, including GPT-4-Coder. A comparison of SEED-APR with state-of-the-art methods such as GPT-4-Coder and CodeT5+ reveals significant improvements in terms of repair accuracy, cross-project generalization, and semantic diversity.