Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries
The authors report a 33-percentage-point accuracy gain over the strongest retrieval and precomputation baselines on BrowseComp-Plus.
A builder agent turns raw documents into a hierarchical filesystem of synthetic files and reference links for a separate answering agent. Reinforcement learning trains the builder to maximize downstream answering accuracy. The authors also report a 31-point average accuracy improvement on BRIGHT in zero-shot tests on unseen domains. Upfront workspace construction costs are spread across repeated queries.
ArXiv · AI/CL/LG's note
A builder agent turns raw documents into a hierarchical filesystem of synthetic files and reference links for a separate answering agent. Reinforcement learning trains the builder to maximize downstream answering accuracy. The authors also report a 31-point average accuracy improvement on BRIGHT in zero-shot tests on unseen domains. Upfront workspace construction costs are spread across repeated queries.
ArXiv · AI/CL/LG's note
score 5