Librarian Agents preprint reports gains from preparing document workspaces

Researchers report that Librarian Agents improved answering accuracy by 33 percentage points over the strongest retrieval and precomputation baselines on BrowseComp-Plus. The preprint describes a system that prepares documents for future questions rather than retrieving information only when a question arrives. A builder agent organizes raw documents into a hierarchy of synthetic files and reference links. Reinforcement learning trains it to improve a separate answering agent's accuracy. The researchers also report a 31-point average accuracy improvement on BRIGHT in unseen domains without additional domain training. Upfront workspace construction costs are spread across repeated queries.