The ongoing evolution of software defect detection methodologies leveraging large language models is rapid; however, the current research landscape has not been adequately investigated. Existing reviews inclusively categorize smaller language models as LLMs, failing to concentrate specifically on the domain of software defect detection and omitting recent advancements in the field. This highlights a gap in the detailed analysis and assessment of cutting-edge applications of large language models in detecting software flaws.
To solve the problems, a research team led by Zulie PAN published their new research on 15 June 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
The team collected high-quality research papers focusing on LLM-based software defect detection, study the progress of related technology development, and predict the development prospects.
In the research, they categorize and summarize existing research based on the distinct applications of LLMs in dynamic and static detection scenarios. Firstly, dynamic detection methods are categorized based on the different phases in which they employ LLMs, such as using them for test case generation, providing feedback guidance, and conducting output assessment. Secondly, static detection methods are classified according to whether they analyze the source code or the binary of the software under test.
They investigated the prompt engineering and model fine-tuning strategies adopted within collected studies. Furthermore, they analyzed the LLMs utilized in the collected papers, the datasets employed, the target software addressed, and the technical performance outcomes. Finally, they summarize the overall concepts and trends of existing research and propose potential directions for future LLM-based software defect detection studies. As the technology evolves, LLMs promise to not only streamline the defect detection process but also to usher in a new era of efficient and error-free software development, thereby elevating industry standards and safeguarding user experiences. This study marks a significant milestone in leveraging the power of language intelligence to mitigate software risks, setting the stage for future innovations in the realm of software engineering.

DOI:10.1007/s11704-025-40672-2