Code Language Models Fine-Tuned to Detect Cross-Language Programming Bugs

A new research paper explores using pre-trained code language models to identify cross-language bugs, which occur when multiple programming languages interact within a single project. The study fine-tuned 13 different CodeLMs on a specially constructed dataset involving three programming language combinations: Python-C/C++, Java-C/C++, and Python-Java. Researchers developed CLCFinder, a tool for identifying cross-language code, and created a dataset containing nine distinct interaction types between languages. After fine-tuning, all models showed performance improvements, with UniXcoder-base achieving the highest F1 score of 0.7407. The investigation analyzed how factors like dataset size, token sequence length, and code comments affect detection capabilities. Multilingual programming has become increasingly common due to its advantages, but it introduces bugs that traditional single-language tools struggle to detect. The paper was announced as arXiv:2507.21954v2 with a replace-cross announcement type.