Context
Zhou and colleagues propose Learning to Retrieve from Agent Trajectories (LRAT), using multi-step agent search, browsing and reasoning traces as retrieval supervision instead of relying solely on human click behavior.
What changed
Introduced an agent-centric retriever training method that mines candidate browsing, unbrowsed rejections and post-browse reasoning signals from agent search trajectories.
Why it matters
Documents that help an agent resolve an intermediate information gap may differ from those most relevant to the initial user query; retriever training can therefore change which sources are surfaced to AI agents.
Visibility OS interpretation
Visibility OS interpretation: model agent-generated subqueries, candidate retrieval, document opening and reasoning contribution as distinct discovery stages. No evidence any named commercial platform has adopted LRAT.
What remains uncertain
How transferable are the results beyond the training models, 10,000 questions and tested retrieval environments? Does reasoning length reliably indicate evidence utility? Can these gains be reproduced in deployed commercial systems?
Evidence claims
The LRAT authors report constructing 26,482 agent search trajectories and 91,713 training query-document pairs from 10,000 InfoSeekQA questions using four retrievers.
In the authors' evaluated deep-research benchmarks, retrievers trained with LRAT improved evidence recall and end-to-end task success across tested agent architectures compared with their baselines.
Related historical records
Agentic-R trains retrievers for multi-step search using local relevance and final-answer utility
That record related to this event.