Large Language Models in Software Network Analysis: A Survey of Applications, Challenges, and Future Trends

Abstract

Software Network Analysis is a fundamental tool for comprehending the structure, dependencies, complexity and evolution of today's software systems. Traditional analysis techniques are however, not suitable for managing large-scale software repositories and offer limited semantic understanding. Large Language Models (LLMs) have contributed much to software analysis with their newfound abilities in code understanding, reasoning, and automated knowledge generation. This survey provides detailed overview of the integration of LLM with Software Network Analysis including its core principles, recent advancements, key applications, challenges, and research directions. It explores how LLMs can be harnessed for software dependency analysis, code comprehension, vulnerability identification, software architecture analysis, graph reasoning and graph representation learning. It delves further into hybrid models that combine LLMs with GNNs to improve analytical skills and scalability. In addition, the survey identifies several problems: computational complexity, hallucination, explainability, privacy protection, security, and limitations of benchmarks. Lastly, it presents the future trends like lightweight domain-specific LLMs, multimodal graph-language learning, explainable AI, and autonomous software engineering systems that benefit researchers and practitioners.