Retrieve-for-Train

A Google Research framework that uses offline reinforcement learning to compile a set-level retrieval reward into supervision and distils it into a diffusion retriever performing single-pass query fan-out.