@article{LI2026, 
author = {HongBei LI and BoWen YU and JunXia GUO},
title = {A graph-based fault localization method utilizing multi-modal code representations},
year = {2026},
journal = {Journal of Beijing University of Chemical Technology (Natural Science Edition)},
volume = {53},
number = {5},
pages = {80-91},
keywords = {fault localization, combination of large and small models, graph neural network, multi-granularity localization},
url = {https://www.sciopen.com/article/10.13543/j.bhxbzr.2026.05.009},
doi = {10.13543/j.bhxbzr.2026.05.009},
abstract = {Automatic fault localization techniques play a critical role in software development, but existing techniques lack in-depth characterization of program flow and code syntactic structure, making it difficult to understand deep logic and contextual information. They also lack universality in dealing with code entities of different granularity, limiting the information interaction of different levels of error localization and the comprehensive understanding of error characteristics.To addres these issues, this study proposes a graph-based fault localization framework utilizing multi-modal code representations, aiming to enhance fault localization effectiveness in two ways: firstly, leveraging multi-source information such as syntactic structure, semantic information, execution logic, and test coverage details to construct multi-modal code representation graphs using both large and small models, enhancing the ability to identify faults in complex code structures; secondly, forming multi-granularity localization models by sharing parameters between method and statement-level localization models to improve fault localization method accuracy. Experimental results conducted on 395 faults from the Defects4J dataset demonstrate that compared to the best baseline technique, our framework results in an increase of 25 correctly localized faults at the method level with a 1.12% improvement in mean first rank (MFR), and an increase of 34 correctly localized faults at the statement level with a 5.34% improvement in MFR.}
}