Token-Level Bug Localization on Large-Scale Codebases

Abstract

Bug Localization (BL) is important for software development, yet existing approaches primarily work at file-level, while developers often need more fine-grained bug locations. This study proposes the first token-level BL approach to help developers localize buggy tokens more precisely. To address the absence of token-level BL datasets, we constructed a large-scale token-level dataset containing bugs across multiple programming languages from open-source repositories. Leveraging this dataset, we fine-tuned existing Large Language Models to frame token-level BL as a sequence-labeling problem. The model takes source code and bug report as input and outputs the probability of each token being buggy. Additionally, we implemented an IDE plugin for visualizing BL results to developers, facilitating seamless integration into workflows. We conducted an evaluation comparing our token-level approach with simple token-level adaptations of existing file-level approach.

BibTeX

@mastersthesis{haoming-mthesis,
    author = {Haoming Huang},
    title = {Token-Level Bug Localization on Large-Scale Codebases},
    school = {東京科学大学 情報理工学院 情報工学系 情報工学コース},
    year = 2025,
}