Google DeepMindAnalysisAI ModelsSeptember 15, 2026

Google Research's Retrieve-for-Train speeds query fan-out 12-20x

Retrieve-for-Train (R4T) uses offline reinforcement learning to compile reward-aligned fan-outs into supervision, then distills them into a lightweight diffusion retriever, cutting inference-time thinking budget. Google Research reports 12x to 20x faster query fan-out versus autoregressive decomposition.

2 sources

More stories today

Open the live feed