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.

Original post →

More from Research

Research channel →