Stanford's Librarian Agents use RL to curate workspaces, beating retrieval by 33 points
StanfordUniversity · hf · 2026-10-20
A Stanford/CMU team (Chelsea Finn's group) introduces Librarian Agents: a builder agent trained with reinforcement learning to reorganize raw document corpora into reusable hierarchical workspaces of synthetic files and reference links, tailored for a downstream answering agent. The insight: LLM agents are good at navigating existing filesystems for immediate tasks but rarely trained to construct workspaces for future queries. Workspace construction is framed as an MDP optimized to directly maximize downstream answering accuracy. Results: +33 percentage points over the strongest retrieval and precomputation baselines on BrowseComp-Plus, amortized construction costs across queries, and zero-shot generalization to unseen domains on BRIGHT with a 31-point average accuracy gain.
More from coding & agent
- Code-Graph-RAG maps real code dependencies with Tree-sitter and Memgraph, hits 5.3k GitHub stars — JeremyCMorgan · 2026-10-20
- Open-source agent workspace Busabase adds team roles and work auditing — PrajwalTomar_ · 2026-10-20
- Claude Code's --worktree --tmux keeps agents alive through dropped SSH and closed terminals — petrusenko_max · 2026-10-20
- PSMcpServer turns any PowerShell script into a live MCP tool just by dropping it in a folder — dfinke · 2026-10-20
- How to Become a Model PM: A 6-Level Roadmap — Only 7 of 316 Have a PhD — aakashgupta · 2026-10-20
- Wabi 2.0 Launches: A 'Multiplayer' Personal Agent Now Open to All — soleio · 2026-10-20