The unprecedented growth of scientific literature makes comprehensive literature analysis increasingly challenging. One of important stages of literature analysis is the extraction of references, necessary for various purposes, such as backward snowballing or citation network analysis. Its manual execution in large scale is prohibitively labor-intensive and its automation is not trivial due to non-typical references and minor errors that are very often encountered in bibliographic data. This paper investigates the use of Large Language Models (LLMs) for this purpose. We present a novel, domain-agnostic pipeline that processes research paper PDFs to extract and parse reference lists into structured data. Evaluating the pipeline on a corpus of 20 papers containing 1,070 references across four diverse research domains, we demonstrate high accuracy (95% success rate) and robust performance, quantified by a new Bibliographic Information Completeness (BIC) metric (average score of 0.79). Our results confirm that LLMs are highly effective for this task, with performance primarily constrained by upstream PDF text extraction quality rather than semantic understanding. This work contributes to the progress in AI-supported literature analysis, significantly reducing the manual burden of bibliographic data collection.
Jakub et al. (2025) studied this question.