An empirical study on the capability of LLMs in decomposing bug reports

Bug reports are essential to the software development life cycle. They help developers track and resolve issues, but are often difficult to process due to their complexity, which can delay resolution and affect software quality. This study investigates whether large language models (LLMs) can assist developers in automatically decomposing complex bug reports into smaller, self-contained units, making them easier to understand and address. We first conducted an empirical study using 127 resolved privacy-related bug reports collected from Apache Jira. We evaluated ChatGPT-4o and DeepSeek-V1 using zero-shot and few-shot prompting. Under zero-shot prompting, the models correctly decomposed only 10 and 11 reports, respectively. Few-shot prompting increased their true decomposition rates by 140% (10 vs. 24) and 163.64% (11 vs. 29), respectively, although the overall accuracy remained limited. To examine whether the findings generalize beyond privacy-related reports and to expand the dataset size, we further analyzed 127 security-related bug reports using GPT-5.5 and DeepSeek-V4-Pro. We compared zero-shot, few-shot, chain-of-thought, and persona prompting. Few-shot prompting achieved the best performance for both models, correctly decomposing 89 reports with GPT-5.5 and 51 reports with DeepSeek-V4-Pro. Across the experiments, we identify a concrete failure taxonomy for LLM-based bug report decomposition, including over-decomposition, over-analysis, incorrect interpretation of solutions, lacking key information, and empty output. Over-decomposition is the most common failure mode, suggesting that LLMs often struggle to distinguish independent bug units from supporting information such as logs, code snippets, background context, and proposed fixes. Overall, LLMs show potential for automated bug report decomposition, but their reliability, prompt sensitivity, and adherence to decomposition instructions still require improvement.