From Empirical Evaluation to Context-Aware Enhancement: Repairing Regression Errors with LLMs

The study presents RegressionBug4APR, a benchmark comprising 200 regression bugs from Java and Python, aimed at evaluating automated program repair (APR) techniques, particularly those utilizing large language models (LLMs). Traditional APR methods failed to address these bugs, while LLM-based approaches demonstrated a 1.6x improvement in repair success when enhanced with context from bug-inducing changes. This research underscores the importance of context-aware strategies in improving the efficacy of LLMs for fixing regression errors, providing valuable insights for practitioners in software debugging and APR development.