Google Research introduced Retrieve-for-Train to speed up complex AI search. It replaces heavy inference-time reasoning with a one-time offline reinforcement learning step. The framework trains a lightweight diffusion model to generate diverse, expert-level search results instantly. It avoids paraphrastic collapse and autoregressive latency bottlenecks. The system uses a composite reward balancing groundedness, diversity, and alignment. It achieves 12 to 20 times faster inference than autoregressive approaches. The diffusion model bypasses the need for chain-of-thought reasoning tokens. It was tested on open-ended abstract retrieval and weakly supervised compositional retrieval tasks. The method outperformed single-query search, zero-shot expansion, and Best-of-N baselines. It maintains sub-second to a few seconds response times even under large context batches.
What changed
Complex AI search relied on expensive inference-time reasoning with autoregressive models.
What it unlocks
Instantly generating a diverse, expert-level set of search results without heavy processing at query time.
- 12 to 20 times faster inference
- 53.9M parameters for diffusion model
- 10 sub-queries per main search prompt
- 4B parameter open-source models used
What you need to act on it
- offline reinforcement learning training
- a pre-trained diffusion model
Sources