AutoStructor: A Generative AI-Based Framework for Automated Program Repair with Deep Learning-Guided Fault Localization

Current automated program repair (APR) approaches still suffer from two critical limitations: inaccurate fault localization and ineffective patch generation. This paper presents a novel hybrid framework, AutoStructor, that proposes a Quality-Aware Graph-Based Machine Learning Fault Localization (GBMFL) and a Large Language Model (LLM)-based patch generation technique to effectively address these challenges. Our key innovation lies in integrating 21 comprehensive code quality metrics and sophisticated multi-modal attention mechanisms, including quality-aware, graph-aware, sparse, and multiscale attention into the GBMFL architecture, coupled with intelligent test failure analysis for superior code understanding and patch generation. AutoStructor bridges the gap between fault localization and patch generation through an end-to-end pipeline that leverages generative AI for intelligent patch synthesis. We evaluate our approach on the Defects4J benchmark, demonstrating significant improvements in fault localization accuracy (81.9% Top-1, 93.4% Top-3) with a Mean First Rank of 1.13, and successful patch generation for 48 out of 61 active Lang project bugs. Notably, our structured test analysis approach enables cost-effective deployment using compact LLMs, achieving effective repair.