Librarian Agents: Training LLMs to Curate Long-Context Workspaces that Anticipate Future Queries

What it claims

A builder LLM is trained with RL to rewrite a raw document corpus into a filesystem workspace before any queries arrive. The workspace holds synthetic summary files, hierarchy and reference links back to the originals. A frozen answering agent then navigates it with ordinary list and read tools. The pitch is that agents already handle filesystems well (the paper cites Claude Code-style harnesses), so the learned structure is exposed through an interface that needs no new retriever. The headline results are an average 33-point gain over the best retrieval and precomputation baseline on BrowseComp-Plus, and an average 31-point gain on BRIGHT, which the builder was never trained on.

How it works

• State: the current filesystem, meaning the raw documents plus any synthetic files, folders and links created so far.
• Action space: deliberately restricted to four builder actions, to shorten the optimisation horizon:
• CreateFile consolidates raw and/or synthetic files and links back to its sources.
• MergeFiles builds a higher-level summary or directory over child files.
• Delete removes synthetic files only. Raw documents are never deleted, so the answering agent can always fall back to them.
• Stop ends construction.

The authors note this could be widened to raw touch/mkdir/mv/rm but did not do so.

• Answering agent: frozen. It has file-listing and file-reading actions and is scored with 0/1 correctness.
• Objective:
• The general utility is accuracy minus token-cost and file-read penalties, where the penalties apply only to correct answers, minus a construction-cost term.
• The dense per-step reward is the change in utility. It telescopes to the sparse objective, which makes it potential-based shaping.
• Practical surrogate:
• The answering agent is probed every 8 builder turns, and the reward combines the accuracy deltas between probes with a terminal answer reward.
• Advantages are computed per step in the style of multi-turn GRPO.
• Data hygiene:
• Reward is estimated on minibatches of training queries that the builder never sees.
• Evaluation queries are never used to build workspaces or to select checkpoints.

flowchart LR
  D[Raw corpus] --> B[Builder policy]
  B -->|Create/Merge/Delete| F[Workspace]
  F --> B
  F -->|every 8 turns| P[Probe frozen answerer<br/>on hidden train queries]
  P -->|accuracy delta| R[Per-step advantage]
  B -->|Stop| T[Terminal answer reward]
  T --> R
  R -->|GRPO update| B

Evidence

The fetched text contains only the headline figures:

• BrowseComp-Plus: +33 points averaged over the best retrieval and precomputation baseline.
• BRIGHT (zero-shot): +31 points average.

The text breaks off before the experiments section. Baseline identities, backbone models, the number of training queries and the size of the construction-cost amortisation are therefore not available here.

How to use it

• Code: https://github.com/teresaz05/librarian-agents
• Project site: https://librarian-agents.github.io/
• Training configuration: Appendix A.7.
• Filesystem instantiation of the actions: Appendix A.8.

The page gives no install commands, model names, sizes or licence.

Limits

• Cost penalties are off in the main run. The primary experiment sets the token, read and construction weights to zero (λ_tok = λ_read = α = 0) and optimises accuracy only. The "amortizes construction cost" claim therefore does not come from the cost-penalised objective. Check how construction cost was actually measured before relying on it.
• Narrow queries may not survive compression. The authors flag that synthesised artifacts can drop details that turn out to matter, especially for needle-in-a-haystack queries. Their mitigation is the raw-document fallback, not the learned structure.
• Training needs a query distribution and a live answerer. Training requires sampled training queries and repeated rollouts of the answering agent. A corpus with no representative queries has no reward signal.
• The learned structure is tied to one answerer. The answering agent is frozen during training. Whether a workspace built for one answerer transfers to a different one is not addressed in the fetched text.
• Partial view of the paper. This analysis covers only the introduction and method. Experimental details, ablations and the authors' own limitations section were not available.