What is it about?
Large language models are increasingly used to detect duplicated or similar source code, but it is unclear which models work best and whether combining several models improves detection. This paper compares a range of suitable models across public and industrial datasets. We find that no single model is best in every setting, although several perform consistently well. We also show that combining models can improve detection, particularly on large-scale codebases, but the way their results are combined matters.
Featured Image
Photo by Google DeepMind on Unsplash
Why is it important?
Our findings show that there is no single large language model that performs best for scalable code clone detection across all datasets. Instead, several model characteristics are associated with performance, and combining models can significantly improve detection on large-scale code. This suggests that effective clone detection depends not only on choosing a strong model, but also on selecting and combining models for the specific deployment setting.
Perspectives
This paper brings together several years of work on scalable code clone detection and reflects many of the questions that arose when trying to move from promising models to methods that work reliably on large, real-world codebases. The collaboration between the University of Limerick, Trinity College Dublin, and Huawei was instrumental in making this work possible. For me, one of the most interesting outcomes was that there was no simple answer to which model is “best”: performance depended on the setting, and combining models was beneficial only under particular conditions. I hope the paper provides a useful starting point for researchers and practitioners facing similar model-selection decisions.
Muslim Chochlov
University of Limerick
Read the Original
This page is a summary of: Selecting and Combining Large Language Models in Scalable Code Clone Detection, ACM Transactions on Software Engineering and Methodology, August 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3838728.
You can read the full text:
Resources
Replication package for the paper
The package includes supplementary data such as scripts required to execute the experiments, raw experimental outputs, publicly available datasets and intermediate artifacts required to reproduce the results.
Industrial C/C++ dataset
The dataset of industrial C/C++ clones that was used in the evaluation and that is openly available.
Large language model fine-tuned for clone detection
CodeBERT based large language model used in the evaluation
Contributors
The following have contributed to this page
