A systematic mapping study on the use of llms for requirements traceability link discovery
Rizvi, Syed Babar Ali (2025)
2025
Tietojenkäsittelyopin maisteriohjelma - Master's Programme in Computer Science
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-07-31
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202507317979
https://urn.fi/URN:NBN:fi:tuni-202507317979
Tiivistelmä
Requirements traceability helps developers and stakeholders understand how different software artifacts including requirements, design models, source code, and test cases are connected. It is considered important for change impact analysis, test coverage verification, and quality assurance, especially in safety-critical domains. However, creating and maintaining this trace links is often time-consuming and incomplete in practice, which has led to long-standing interest in automating the process through techniques known as Traceability Link Recovery (TLR).
Earlier approaches to TLR have relied on information retrieval, rule-based heuristics, and supervised learning. While these methods can identify basic relationships, they often struggle to handle language variability and project-specific semantics. In recent years, research has begun to explore whether Large Language Models (LLMs) can improve traceability by understanding natural language more deeply. Some studies report encouraging results, suggesting that LLMs may identify trace links even when textual overlap is minimal. Hybrid methods, such as retrieval-augmented generation, have also been investigated to address the limitations of LLM input size and generalization.
This thesis presents a systematic mapping study of 132 peer-reviewed studies (2015–2025) that investigate the use of LLMs for requirements traceability link recovery (TLR). The study classifies techniques into prompting, fine-tuning, and retrieval-augmented generation (RAG), and analyzes their use across various traceability tasks (e.g., requirements-to-code, requirements-to-test). Key findings reveal that 44% of studies use prompting, 29% fine-tuning, and 13% RAG, with requirements-to-code being the most common traceability direction.
While LLM-based methods outperform traditional approaches in many cases, challenges such as input length constraints, hallucinations, and explainability remain. This mapping highlights emerging tools, datasets, and industrial applications, offering actionable insights for researchers and practitioners. The results underscore LLMs’ growing potential to augment traceability in complex, evolving software systems.
Earlier approaches to TLR have relied on information retrieval, rule-based heuristics, and supervised learning. While these methods can identify basic relationships, they often struggle to handle language variability and project-specific semantics. In recent years, research has begun to explore whether Large Language Models (LLMs) can improve traceability by understanding natural language more deeply. Some studies report encouraging results, suggesting that LLMs may identify trace links even when textual overlap is minimal. Hybrid methods, such as retrieval-augmented generation, have also been investigated to address the limitations of LLM input size and generalization.
This thesis presents a systematic mapping study of 132 peer-reviewed studies (2015–2025) that investigate the use of LLMs for requirements traceability link recovery (TLR). The study classifies techniques into prompting, fine-tuning, and retrieval-augmented generation (RAG), and analyzes their use across various traceability tasks (e.g., requirements-to-code, requirements-to-test). Key findings reveal that 44% of studies use prompting, 29% fine-tuning, and 13% RAG, with requirements-to-code being the most common traceability direction.
While LLM-based methods outperform traditional approaches in many cases, challenges such as input length constraints, hallucinations, and explainability remain. This mapping highlights emerging tools, datasets, and industrial applications, offering actionable insights for researchers and practitioners. The results underscore LLMs’ growing potential to augment traceability in complex, evolving software systems.