Kuaishou proposes GBPO, fixing GRPO's dead gradients and unbounded updates for generative recommenders
_reachsumit · x · 2026-11-18
A Kuaishou arXiv paper revisits GRPO-style objectives for generative recommendation via a unified logit-gradient view, identifying dead-gradient regions from clipping, unbounded negative-advantage updates, and insensitivity to policy confidence. It proposes GBPO with stop-gradient normalization — fully gradient-active, sign-symmetrically bounded, confidence-weighted — achieving competitive accuracy with more stable training.
More from Research
- UCLA hosts California AI Amplitudes school applying ML to quantum scattering amplitudes — KyleCranmer · 2026-11-18
- Looped layers don't help multimodal embeddings unless recurrence is pretrained (+4.23 points) — _reachsumit · 2026-11-18
- Ant Group's Repo-Premium shows BM25 matches dense retrieval in fair code-search tests — _reachsumit · 2026-11-18
- GainReason rewards each reasoning sentence by its predictive gain for generative recommendation — _reachsumit · 2026-11-18
- Training-free IN-Retriever uses embedding attention to navigate multi-hop RAG queries — _reachsumit · 2026-11-18
- Tencent's TokenFormer V2 matches top unified recommenders with 5.1× fewer FLOPs — _reachsumit · 2026-11-18