AI-Driven Software Debugging and Automatic Repair Framework: Research on Machine Learning and Large Language Models in Program Repair

Software debugging is one of the most time-consuming and challenging stages in the software development process; developers spend an average of over 50% of their time debugging and fixing code errors. The AI-in-Software-Debugging-Research project on GitHub integrates machine learning (ML) and large language model (LLM) technologies to explore cutting-edge methods for automated program error detection and repair, representing the latest research direction at the intersection of software engineering and artificial intelligence. This article will analyze from multiple dimensions including background, technical framework, and application practices.

Software debugging faces three core challenges: difficulty in error localization (complex causal chains, hard-to-track dependencies, concurrent timing issues), high repair costs (prone to introducing regression bugs, lack of automated verification), and strong knowledge dependency (relying on personal experience, low efficiency for novices). Automated Program Repair (APR) has gone through four generations of evolution: search-based (GenProg) → semantics-based (SemFix) → learning-based (SequenceR) → LLM-based (ChatRepair), among which LLMs have the advantages of no need for specialized training and strong generalization capabilities.

This framework includes four core components: 1. Error detection and localization (static analysis, dynamic analysis, anomaly detection); 2. Error understanding and classification (NLP analysis of error types, contextual semantic understanding); 3. Automatic patch generation (LLM repair, retrieval-based repair, hybrid strategies); 4. Repair verification and evaluation (test verification, semantic verification, quality assessment).

Machine learning is applied to error localization (learning-based LEL, CNN/RNN for code structure processing) and patch generation (sequence-to-sequence learning, NMT, GNN). The advantages of LLMs lie in pre-trained knowledge, contextual understanding, and code generation capabilities; their APR process includes context construction, prompt engineering (Zero-shot/Few-shot/Chain-of-Thought), candidate generation, screening, and application.

Current challenges: repair quality (prone to introducing new errors), verification difficulties (incomplete test cases), weak generalization (poor cross-project/language performance), high computational cost. Future directions: multi-modal APR, interactive repair, continuous learning, causal reasoning.

Application scenarios include: 1. Development assistance (IDE integration, code review); 2. Automated testing (CI/CD integration, regression testing); 3. Legacy code maintenance (modernization, technical debt management); 4. Education and training (programming teaching, code review training).

Academic tools: classic APR (GenProg, Prophet, Angelix), deep learning APR (SequenceR, CURE), LLM-based APR (ChatRepair, RepairLLaMA). Commercial tools: GitHub Copilot, Amazon CodeWhisperer, Tabnine, Snyk.

Developers should embrace AI-assisted tools (use critically), focus on code quality (improve test coverage), and continue learning (adapt to AI trends). This project represents an important direction in software engineering; AI-driven debugging and repair will become a standard configuration, and although there are limitations, the trend is irreversible.