A Comprehensive Study of LLM-Based Code Clone Detection

Large language models (LLMs) have strong code-understanding skills, so they are well-suited for code clone detection. This is a key code analysis task that shortens development cycles, improves code quality and reduces security vulnerabilities. Recent studies use prompt engineering, parameter fine-tuning and retrieval-augmented generation (RAG) to raise detection accuracy and stability, but systematic surveys on LLM-based clone detection are still not many. To fill this gap, we do a full survey and empirical analysis of LLMs in clone detection, with a focus on Java programs. First, we review existing methods from multiple dimensions, classifying them by methodological techniques and task requirements. To make the study reliable, we collect datasets of typical clone fragments from different perspectives and conduct unified evaluations on ten models of three types: traditional code analysis, specialized pre-trained deep learning models and general-purpose LLMs. We also use various prompting strategies to ensure fair and comparable assessments. We sum up the strengths and limitations of current methods and point out future research directions. Experiments show that for current clone detection tasks, general-purpose LLMs achieve the best overall performance, specialized deep models perform inconsistently in different settings, and traditional analysis methods do worse. However, the effectiveness of LLMs is still affected by task configuration and prompt design, which means there is much room to improve their robustness and consistency.