Improving File-Level Bug Localization Using Pre-Trained Code Models: A Comprehensive Review

Bug localization, the automatic detection of source code files that contain a given defect, is a fundamental problem in software maintenance. Pre-trained models, such as CodeBERT, GraphCodeBERT, UniXcoder, and CodeT5, can effectively bridge the semantic gap between bug reports in natural language and source code. However, existing studies use inconsistent datasets and evaluation protocols, leading to non-comparable and non-reproducible results. This review focuses on file-level bug localization using pre-trained models, going beyond prior surveys by identifying cross-study inconsistencies, highlighting a structural gap in LLM-based file-level evaluation, and providing a critical analysis of existing approaches. The key contributions are: a five-category taxonomy combining IR, ML, deep learning, pre-trained language model, and LLM-based approaches (the first taxonomy targeting file-level granularity among all paradigms); a cross-study analysis suggests that identical models on identical benchmarks report Top-10 accuracy values differing by up to 20 percentage points (as a result of undisclosed experimental differences); and the mapping into a structured framework of eight open research gaps to seven evidence-supported directions, followed by the most recent advances in the field, including studies published up to 2026.