Google researchers have proposed a retrieval system that learns complex search planning during training instead of asking a large language model to reason through every query at runtime. The approach targets searches that need a useful set of complementary results—for example, assembling tents, sleeping bags, stoves and lights rather than returning several similar tents.
Retrieve-for-Train first uses reinforcement learning to teach a language model to generate grounded and diverse subqueries. That model produces synthetic query-and-target pairs, which train a 53.9-million-parameter diffusion retriever. At runtime, the smaller model maps a query directly to a set of result embeddings in one non-autoregressive pass, avoiding hundreds of sequential reasoning tokens.
The reward balances three properties: closeness to items that actually exist in the database, diversity across the whole set and alignment with the original request. Each element matters. Grounding alone can reward meaningless strings near database coordinates, while alignment without diversity can collapse into paraphrases. The work was presented at ICML 2026 and tested in open-ended and weakly supervised retrieval settings. It is a research framework rather than a Google Search product launch, but it shows a practical way to trade expensive per-query reasoning for a smaller, specialized model trained ahead of time.