Abstract
Reconstructing bioinformatics workflows from the literature is the foundation of scientific analysis. However, the required details-processing steps, software tools, versions, and parameter settings-are dispersed across narrative text, tables, figure captions, and supplemental files. Manual reconstruction typically takes hours per paper and is error-prone, while existing question-answering (QA) and retrieval systems focus on local passages and lack the full-text, multimodal capabilities needed to automatically rebuild complete workflows. We introduce BioWorkflow, a large language model (LLM)-based, retrieval-augmented framework that automates end-to-end workflow extraction from publications by (i) parsing PDFs and building a unified index over text, tables, and figures with chunk-level summaries and embeddings; (ii) hierarchically decomposing queries with dynamic reformulation when new entities or ambiguities emerge; (iii) performing iterative, context-aware retrieval and assembling a directed workflow that captures steps, tools, versions, and parameters; and (iv) linking each predicted element to its cited evidence and running automated consistency checks to suppress hallucinations and ensure traceability. Evaluated on 100 expert-annotated papers, BioWorkflow recovers ~80% of workflow steps (versus ~20% for existing tools), improves reproducibility, completeness, and accuracy by >20% over strong LLM baselines, and reduces curation time to 3-5 minutes per paper, enabling rapid and reliable reuse of published pipelines.