UDC 004.8
METHOD FOR DEFECT LOCALIZATION AT THE SYSTEM LEVEL USING LLM
Yu. L. Leohin, Dr.in technical sciences, full professor, Department Professor, MTUCI, Moscow, Russia;
orcid.org/0000-0003-3321-4497, e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
A. A. Syroezhko, postgraduate student, MTUCI, Moscow, Russia;
orcid.org/0009-0005-3790-5561, e-mail: This email address is being protected from spambots. You need JavaScript enabled to view it.
This paper formulates defect localization problem as a search for relevant semantic systems within a
source code, shifting the focus from individual files or code snippets to higher-level structural units. The
study proposes a method to address this problem using Large Language Models (LLMs). The proposed approach is based on clustering vector representations of source code and comprises the following stages:
AST-based code segmentation and vectorization, dimensionality reduction (UMAP) combined with clustering
(HDBSCAN), and vector search. The interpretability of the resulting systems is also examined. An integrated
metric is developed to evaluate the quality of proposed clustering technique, along with parameters for its
optimization. The results of numerical experiments are presented, demonstrating the effectiveness of the
method for the problem stated.
Key words: bug localization, semantic source code analysis, large language models, abstract syntax
tree, text vector representations, density-based clustering, text summarization, code segmentation, clustering
metrics