Google Research Introduces Retrieve-for-Train (R4T): RL-Compiled Diffusion Retriever for 12× to 20× Faster Query Fan-Out

Google Research introduced Retrieve-for-Train (R4T), a method that trains query fan-out via RL and distills it into a 53.9M-parameter diffusion model. The model generates all search directions in one pass, speeding up fan-out by 12–20× compared to autoregressive approaches.

• R4T-FOLM on Gemma3-4B raised average OAR on Polyvore from 40.9 to 49.1
• Diffusion retriever with 53.9M parameters outputs all embeddings in one pass
• At batch 1024, autoregression takes ~50 s, diffusion takes 4.21 s
• Three rewards — groundedness, diversity, alignment — prevent paraphrase collapse