Google makes query fan-out 12x faster

Google Research’s Retrieve-for-Train (R4T) is a 53.9-million-parameter diffusion retriever that generates 10 search directions in one pass and reports 12–20× lower query fan-out latency than autoregressive approaches. The research post moves costly objective discovery into offline reinforcement learning; the paper says its gains come from set-valued fashion and music retrieval benchmarks, not an end-to-end production search SLA. Search builders should file this as a way to cap fan-out cost before swapping answer models, alongside the archive’s benchmark showing a $124,937 per-million-task spread between search providers.