Context
A research preprint introduces CoSearch, jointly training a multi-step reasoning agent and a generative document ranker while keeping the first-stage dense retriever fixed. The authors report improvements over Search-R1 across seven QA benchmarks.
What changed
CoSearch applies GRPO to the search agent and a trainable generative reranker, using semantically grouped intermediate queries and composite ranking/answer rewards; the first-stage retriever is fixed.
Why it matters
Intermediate agent-generated subqueries and the document admission/reranking stage can shape which evidence reaches a reasoning agent. This suggests retrieval and reranking should be evaluated separately from final citations.
Visibility OS interpretation
Visibility OS interpretation: an AI search system may adapt its evidence-selection step to downstream answer utility, not only initial query-document similarity. This is a mechanism illustrated by experimental research, not a demonstrated commercial search ranking factor.
What remains uncertain
Will jointly optimized rankers generalize to live, open-web search and source citations? How stable are the effects across corpora, models, and retrieval latency constraints?