AI Driven Software Defect Prediction and Automated Code Optimization in Large-Scale Software Systems

Abstract

The engineering, maintenance, and verification of large-scale software systems present substantial financial and operational challenges within the modern software development lifecycle (SDLC). Traditional Software Defect Prediction (SDP) frameworks historically relied on static, coarse-grained code metrics, which frequently overlooked early-stage design context and struggled to align with continuous integration pipelines. However, the integration of artificial intelligence (AI), machine learning (ML), and large language models (LLMs) has fundamentally transformed this paradigm, shifting software engineering from heuristic static analysis toward predictive, generative, and neuro-symbolic pipelines. This review article provides a comprehensive evaluation of the state of the art in AI-driven defect prediction, effort-aware resource allocation, and automated program repair (APR). It synthesizes diverse advancements, ranging from early requirement-driven predictive models like FORESIGHT to change-level Just-In-Time (JIT) predictors designed to mitigate cold-start issues in data-scarce environments. Furthermore, the paper examines machine learning-guided compiler optimization frameworks, such as Google’s MLGO, which successfully embed reinforcement learning policies to achieve automated code size reductions and performance speedups. Finally, the article investigates the mitigation of static analysis false positives through neuro-symbolic hybrid frameworks like MemHint and ZeroFalse. Despite these rapid technological leaps, an empirical investigation underscores the "94% rule," which highlights a stark disconnect regarding enterprise-level utility where distributed architectures, verification overheads, and security risks limit optimization efficacy. Ultimately, this review maps the ongoing industry-wide progression toward closed-loop architectures where flexible AI generation is safely bounded by formal mathematical verification.