Cross-Level Requirements Tracing Based on Large Language Models
Chuyan Ge, Tiantian Wang, Xiaotian Yang, Christoph Treude
Abstract
Cross-level requirements traceability, linking <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">high-level requirements</b> (<bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">HLRs</b>) and <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">low-level requirements</b> (<bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LLRs</b>), is essential for maintaining relationships and consistency in software development. However, the manual creation of requirements links necessitates a profound understanding of the project and entails a complex and laborious process. Existing machine learning and deep learning methods often fail to fully understand semantic information, leading to low accuracy and unstable performance. This paper presents the first approach for cross-level requirements tracing based on large language models (LLMs) and introduces a data augmentation strategy (such as synonym replacement, machine translation, and noise introduction) to enhance model robustness. We compare three fine-tuning strategies—LoRA, P-Tuning, and Prompt-Tuning—on different scales of LLaMA models (1.1B, 7B, and 13B). The fine-tuned LLMs exhibit superior performance across various datasets, including six single-project datasets, three cross-project datasets within the same domain, and one cross-domain dataset. Experimental results show that fine-tuned LLMs outperform traditional information retrieval, machine learning, and deep learning methods on various datasets. Furthermore, we compare the performance of GPT and DeepSeek LLMs under different prompt templates, revealing their high sensitivity to prompt design and relatively poor result stability. Our approach achieves superior performance, outperforming GPT-4o and DeepSeek-r1 by 16.27% and 16.8% in F1 score on cross-domain datasets. Compared to the baseline method that relies on prompt engineering, it achieves a maximum improvement of 13.8%.