一句话：探索用大语言模型自动化开源软件故障分析，显著提升效率但故障分类仍受限。
软件故障理解是软件开发与维护实证研究的基础。传统故障分析依赖专家完成故障信息收集、过滤、人工调查与分类，耗时耗力，难以支撑大规模、快速演化系统下的实证研究。作者将实证软件故障研究分解为三个关键阶段，并探索利用大语言模型（LLM）支持开源软件故障分析。结果显示，LLM 可将处理时间从数周缩短至约两小时，显著提升分析效率；在识别与缺陷相关的问题上表现良好，但在准确分类故障类别方面仍有限。研究揭示了 LLM 自动化实证软件故障研究的潜力与当前挑战。
Understanding software faults is fundamental to empirical research on software development and maintenance. Traditional fault analysis, although effective, typically relies on multiple human expert-driven steps, including software fault information collection, filtering, manual investigation, and classification. These processes are labor-intensive and time-consuming, creating significant bottlenecks for large-scale fault studies in complex and critical software systems with abundant faults and limiting iterative empirical research amid rapid software evolution. In this paper, we decompose empirical software fault studies into three key phases and conduct an exploratory study on applying large language models (LLMs) to support fault analysis in open-source software. Our results show that LLMs can substantially improve analysis efficiency, reducing processing time to approximately two hours compared to weeks of manual effort. While LLMs demonstrate promising performance in identifying bug-related issues, they remain limited in accurately classifying fault categories. These findings highlight both the potential and current challenges of using LLMs to automate empirical software fault studies.