Enhancing APR for LLM Bugs

This study explores automated program repair (APR) for large language model (LLM)-generated defects, emphasizing Codex-edit's ability to outperform traditional APR tools like TBar and Recoder through diverse patch generation and multi-hunk repair strategies. Using a dataset of 113 Java tasks from LeetCode, it examines defect taxonomy, repair effectiveness, and fault-localization's role. Codex-edit demonstrates strengths in repairing algorithmic and multi-hunk errors, blending templates and...