Code Clone Detection Techniques Based on Large Language Models

Code duplication, commonly known as code cloning, is a persistent challenge in software development. While reusing code fragments boosts productivity, excessive cloning poses challenges to maintenance and elevates the risk of bugs. Therefore, integrating code clone detection into the development process is crucial. The extensive code-related knowledge inherent in Large Language Models (LLMs) renders them high-potential candidates for addressing diverse software engineering challenges. However, the effectiveness of LLMs in the specific task of code clone detection requires precise evaluation. This paper proposes an innovative methodology leveraging few-shot instruction-tuned GPT-3.5 Turbo and GPT-4 to detect code clones across all types, focusing on complex clones (Type-3 and Type-4). Unlike conventional approaches confined to specific language pairs or tasks, our method employs versatile language models, showcases generalization strengths for semantic understanding, and leverages instruction tuning with few-shot inference for task-specific adaptability in code clone detection. A conversational dataset was crafted from BigCloneBench for instruction tuning, enhancing task alignment and performance. This study evaluates the proficiency of LLMs in identifying code clones, analyzing the impact of instruction tuning, and assessing the efficiency across various clone types. Experimental results demonstrate these models achieving competitive performance against existing tools for overall and complex clone detection. Integration into an Integrated Development Environment (IDE) enables real-time detection and automated refactoring, bridging the gap between theoretical advancements and practical usability. This work highlights the potential of generalized LLMs setting a new standard in a field traditionally dominated by specialized tools and demonstrates their adaptability for complex challenges in code analysis and maintainability.