End-to-End Contextualized URL Extraction from Scholarly Papers

Uniform resource locators (URLs) have become ubiquitous in scholarly papers, and extracting them is an upstream task for downstream analyses such as link rot analyses, meta-analyses, and reproducibility studies. Most existing works extract URLs with a monolithic method, such as a regular expression, which assumes URLs appear in a perfect form throughout the text. In practice, layout variety and conversion errors leave many URLs imperfect, without a scheme, or truncated across lines and columns. Although context around a URL is needed to understand its role in a paper, a URL in a footnote or reference sits apart from its citing sentence. We present VisURLX, a pipeline built on a multimodal large language model (MLLM) that reads each rendered page image directly. It preserves page layout such as superscript markers and column structure, and extracts each URL with its target sentence and surrounding context across the body, footnotes, and references. We evaluate VisURLX on sURL-3K, our benchmark of 3,376 manually annotated target sentences with their URLs, using a proprietary backend (GPT-5-mini) and an open-weight backend (Qwen3-VL-32B). VisURLX on GPT-5-mini achieves an end-to-end F1 score of 86.86%, about 20 points above the strongest heuristic pipeline built on GROBID, PyMuPDF-Layout, and olmOCR. As one application, we classify the extracted URLs into open access datasets and software (OADS), reaching a weighted F1 score of 83.1%. The code, data, and the full text of every prompt are available at https://github.com/lamps-lab/sURL3k-VisURLX.