Google Research has introduced Retrieve-for-Train, or R4T, for one of search and recommendation’s hardest scaling problems: returning a strong set of results from one broad prompt. Its central move is bold—run reinforcement learning once offline, then deploy a compact diffusion retriever that generates every retrieval direction in one pass.
That design gives R4T a reported 12× to 20× speedup over autoregressive fan-out. The result is a retrieval system that aims to preserve variety and relevance without forcing an inference-time language model to generate each sub-query one after another.
One Prompt, Ten Retrieval Directions
Query fan-out splits one broad prompt into several sub-queries, giving a retrieval system more paths toward useful results. Generic LLMs can struggle with this job at inference time because paraphrastic collapse pushes their outputs toward similar meanings, while sequential generation adds latency.
R4T attacks both problems by learning the behavior before deployment. Its reinforcement learning process finds good fan-out behavior offline, then distills that behavior into a small diffusion model that produces all retrieval directions in one pass.
The core system uses a 53.9M-parameter diffusion transformer. It maps a query embedding to a full set of target embeddings, with the encoder using 128 dimensions. At inference, R4T generates all embeddings in a single non-autoregressive pass rather than producing one sub-query at a time.
The diffusion transformer uses a variance-exploding formulation within the EDM framework. Every fan-out method in the experiments produced 10 sub-queries, setting k = 10, while Best-of-N used N = 5.
Rewards That Balance Relevance and Variety
R4T’s training targets depend on the retrieval task. For open-ended abstract retrieval, or OAR, the reward combines three weighted terms: Groundedness with λg = 0.6, Diversity with λd = 0.2, and Alignment with λa = 0.2.
For weakly supervised compositional retrieval, or WSCR, the reward equals the fraction of reference-set items retrieved by the fan-out. That gives the system a direct measure of how much of the expected result set its sub-queries recover.
The experiments used the Polyvore fashion outfit dataset and a proprietary expert-playlist music dataset. They also used MuLan embeddings, with Gemma3-4B serving as the model in the reported experiments.
On Polyvore, Gemma3-4B R4T-FOLM averaged 49.1, ahead of Best-of-N at 40.9 and zero-shot at 38.5. The diversity score rose from 56.0 for zero-shot to 76.8, while R4T-Diffusion reached 74.3.
The music results followed the same direction. Gemma3-4B R4T-FOLM averaged 58.1, compared with 49.2 for Best-of-N, showing how the learned fan-out can cover a broader target space without surrendering retrieval performance.
The Speed Advantage Appears at Scale
R4T-FOLM also posted 20.9 Recall@5K and 64.6 Hit@5K, compared with 15.7 Recall@5K and 52.1 Hit@5K for Gemini-2.5-Flash. R4T-Diffusion kept a higher Vendi Score of 34.7 with 16.5 Recall@5K, adding another signal that its outputs retain variety.
The latency results reveal why the architecture matters for large retrieval workloads. At batch size 8, autoregressive fan-out took about 1.46 seconds, while diffusion took 0.07 seconds. At batch size 1024, the gap widened to nearly 50 seconds for autoregressive fan-out versus 4.21 seconds for diffusion.
That difference comes from the generation pattern. Autoregressive fan-out must step through its outputs, but R4T creates the full embedding set together. As batch sizes grow, the single-pass approach gives the retriever room to handle more queries without allowing fan-out latency to dominate the pipeline.
R4T runs RL once offline, then deploys a 53.9M-parameter diffusion retriever. That separation turns expensive behavior discovery into a preparation step and leaves inference with a model built for parallel output.
The larger idea reaches beyond one benchmark: retrieval systems can use learning to discover what a strong fan-out looks like, then compile that behavior into a faster generator. Google Research’s results put that approach at 12× to 20× faster than autoregressive fan-out, opening a new path for search and recommendation systems that need both coverage and speed.
