Evaluating the Use of LLMs for Documentation to Code Traceability
Fuente:
arXiv
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866908480662667264 |
|---|---|
| author | Alor, Ebube Khatoonabadi, SayedHassan Shihab, Emad |
| author_facet | Alor, Ebube Khatoonabadi, SayedHassan Shihab, Emad |
| contents | Large Language Models (LLMs) offer new potential for automating documentation-to-code traceability, yet their capabilities remain underexplored. We present a comprehensive evaluation of LLMs (Claude 3.5 Sonnet, GPT-4o, and o3-mini) in establishing trace links between various software documentation (including API references and user guides) and source code. We create two novel datasets from two open-source projects (Unity Catalog and Crawl4AI). Through systematic experiments, we assess three key capabilities: (1) trace link identification accuracy, (2) relationship explanation quality, and (3) multi-step chain reconstruction. Results show that the best-performing LLM achieves F1-scores of 79.4% and 80.4% across the two datasets, substantially outperforming our baselines (TF-IDF, BM25, and CodeBERT). While fully correct relationship explanations range from 42.9% to 71.1%, partial accuracy exceeds 97%, indicating that fundamental connections are rarely missed. For multi-step chains, LLMs maintain high endpoint accuracy but vary in capturing precise intermediate links. Error analysis reveals that many false positives stem from naming-based assumptions, phantom links, or overgeneralization of architectural patterns. We demonstrate that task-framing, such as a one-to-many matching strategy, is critical for performance. These findings position LLMs as powerful assistants for trace discovery, but their limitations could necessitate human-in-the-loop tool design and highlight specific error patterns for future research. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2506_16440 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Evaluating the Use of LLMs for Documentation to Code Traceability Alor, Ebube Khatoonabadi, SayedHassan Shihab, Emad Software Engineering Artificial Intelligence Machine Learning Large Language Models (LLMs) offer new potential for automating documentation-to-code traceability, yet their capabilities remain underexplored. We present a comprehensive evaluation of LLMs (Claude 3.5 Sonnet, GPT-4o, and o3-mini) in establishing trace links between various software documentation (including API references and user guides) and source code. We create two novel datasets from two open-source projects (Unity Catalog and Crawl4AI). Through systematic experiments, we assess three key capabilities: (1) trace link identification accuracy, (2) relationship explanation quality, and (3) multi-step chain reconstruction. Results show that the best-performing LLM achieves F1-scores of 79.4% and 80.4% across the two datasets, substantially outperforming our baselines (TF-IDF, BM25, and CodeBERT). While fully correct relationship explanations range from 42.9% to 71.1%, partial accuracy exceeds 97%, indicating that fundamental connections are rarely missed. For multi-step chains, LLMs maintain high endpoint accuracy but vary in capturing precise intermediate links. Error analysis reveals that many false positives stem from naming-based assumptions, phantom links, or overgeneralization of architectural patterns. We demonstrate that task-framing, such as a one-to-many matching strategy, is critical for performance. These findings position LLMs as powerful assistants for trace discovery, but their limitations could necessitate human-in-the-loop tool design and highlight specific error patterns for future research. |
| title | Evaluating the Use of LLMs for Documentation to Code Traceability |
| topic | Software Engineering Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2506.16440 |