Specification-Guided Program Repair: Intermediate Behavioral Signals Outperform End-to-End Test Feedback

arXiv·medium signal

Le-Anh et al. show that LLM-based automated program repair improves significantly when guided by intermediate behavioral signals (localized assertions about program state) rather than coarse pass/fail test outcomes. The approach mirrors human debugging: reasoning about where internal logic deviates from intent. For builders using LLMs for code repair and debugging, this suggests that feeding intermediate state information to the model is more effective than just showing test results.