ScholarWorks@경상국립대학교

초록

During software maintenance, developers spend a considerable amount of time locating the source of bugs. Despite the advent of LLMs and vibe coding, accurately identifying the locations for bug fixes in large-scale codebases remains a persistent challenge. In this paper, we propose a new bug localization approach that counts the modification frequencies of similar bug reports for identifying source locations that cause bugs. Given a new bug report, our approach extracts the similar bug reports that include completed bug fixes. Then, our approach counts the number of times that each method was modified in those similar reports. Finally, our approach identifies the methods that were frequently modified in those similar reports and recommends the methods to be fixed. We also implement the approach as an MCP tool that can be used in LLM-based development environments. We evaluated our approach against reimplemented versions of FineLocator as well as LLMs on seven open-source projects. Across all projects our approach improves Precision@10, Recall@10, MAP, and MRR over the baseline FineLocator. The average absolute gains are 0.19, 0.11, 0.09, and 0.12, respectively. The corresponding improvements are 2.12, 2.22, 2.50, and 2.71 times. When deployed as an MCP tool, MCP+LLM yields 0.52, 0.35, 0.36, and 0.34, compared to LLM-only at 0.17, 0.09, 0.11, and 0.10 across the same metrics. These gains hold for both Java and Python codebases, demonstrating that combining large-language-model semantics with modification-frequency evidence yields substantially more accurate and language-agnostic bug localization than conventional methods. Our implementation, datasets, and replication scripts are publicly released to foster future research and industrial adoption of high-precision method-level bug localization.

키워드