Google Research presents faster diffusion-based query fan-out
Google Research has introduced Retrieve-for-Train (R4T), a framework designed to produce coherent and varied result sets for broad searches. Instead of generating similar sub-queries sequentially, R4T uses reinforcement learning offline to train a fan-out language model. The model is optimized for groundedness, diversity and alignment, then generates synthetic training pairs without human labels.
A 53.9M-parameter diffusion transformer learns to convert one query embedding into a complete set of target embeddings in a single non-autoregressive pass. Nearest-neighbor search maps those embeddings to database items. The approach is intended to reduce paraphrastic collapse, in which multiple sub-queries retrieve nearly identical results, and to avoid the latency of repeated autoregressive generation and retrieval.
In Polyvore experiments, Gemma3-4B R4T-FOLM achieved an average OAR score of 49.1, compared with 40.9 for Best-of-N and 38.5 for zero-shot generation. Its diversity score was 76.8, while R4T-Diffusion reached 74.3. On the music dataset, R4T-FOLM scored 58.1 versus 49.2 for Best-of-N. At batch size 8, diffusion took 0.07 seconds compared with 1.46 seconds for autoregressive fan-out; at batch size 1024, the times were 4.21 and nearly 50 seconds, respectively. The researchers report a 12× to 20× speedup.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.