Google Research's retriever spreads a search query in one pass, 12-20x faster

single source· 1 articles · confidence: medium · first seen 2026-09-17 06:19 UTC

What this means for you

Nothing to act on yet: no code, no weights, no released model. If you build retrieval, the claim worth tracking is single-pass fan-out — one 53.9M-parameter model instead of generating variants token by token — but the 12–20× figure is the authors' own, with no harness or evaluation date published.

Google Research has described Retrieve-for-Train (R4T), a search framework that returns result sets which are both on-topic and varied. It trains a fan-out model — one query expanded into several retrieval directions — with reinforcement learning, a reward signal rather than labelled examples. That model then synthesises the training data for a 53.9M-parameter diffusion retriever, which produces every direction in a single pass. The authors report that as 12× to 20× faster than autoregressive fan-out, which emits directions token by token, with no evaluation date given for the figures. No code or weights have been released.

Key facts

  • ·Retrieve-for-Train trains a fan-out language model with reinforcement learning, rewarding groundedness, diversity and alignment. source
  • ·That fan-out model synthesises the training data for a 53.9M-parameter diffusion retriever. source
  • ·The diffusion retriever generates all retrieval directions in a single pass, reported as 12× to 20× faster than autoregressive fan-out. source
  • ·No code or model weights have been released. source
  • ·The speed comparison is reported without an evaluation date. source

What the sources say

  • MarkTechPostWalks through the two-stage training pipeline and the single-pass retriever, and notes no release.

Sources

The original reporting. Follow these — they did the work.

← the wire