Retrieve-for-Train: offline RL and diffusion retriever for faster fan-out search
Google Research's Retrieve-for-Train uses offline RL to train fan-out LMs, synthesize supervision, and distill behavior into a 53.9M-parameter diffusion retriever for single-pass, non-autoregressive fan-out.
What changed
Introduced Retrieve-for-Train: offline RL trains fan-out LMs, synthesizes (query→set) supervision, and distills into a 53.9M-parameter diffusion retriever for single-pass fan-out.
Why it matters
Developers and enterprises can cut inference latency for set-valued search, reducing autoregressive fan-out from nearly 50 seconds to sub-second to a few seconds and getting a 12 to 20 speedup.
Who should care
DevelopersResearchersEnterprises
Worth trying
Yes — Compact 53.9M diffusion model gives 12 to 20 speedup and sub-second to few-second latency.