Don’t ask one LLM to read the literature. Turn literature mining into a pipeline. Scientific papers were written for humans, not for machine learning. A single experiment can have its composition in one table, curing conditions in the text, acronyms defined somewhere else, and measurements reported in different units. Li and coauthors handle this by breaking literature extraction into specialized tasks. One LLM agent reads tables, another finds experimental conditions, others expand acronyms, normalize units, reconcile names and merge the pieces into complete records. That modular choice turns a messy reading task into something repeatable. From more than 27,000 papers, they build a curated database of 8,979 experimental records with over 100 attributes, processing the selected corpus in under an hour. The useful lesson is broader: When scientific knowledge is scattered across papers, don’t treat extraction as one giant prompting problem. Decompose it into a reproducible data pipeline. That is how LLMs can help turn the scientific literature from something we read into infrastructure we can compute on. Li et al., npj Computational Materials (2026), CC BY-NC-ND 4.0. nature.com/articles/s41524-0…