Preprint
Sep 2026
SGA-Flow-GRPO: Spatial Gradient-Guided Credit Assignment for Flow-GRPO
This work reformulates the transition-level log-likelihood in Flow-GRPO into a token-wise representation natively aligned with DiT patch architectures, constructing spatially fine-grained importance sampling ratios and introduces a continuous spatial credit map derived from reward gradients.
Yun-Kai Yang, Yudong Zhang, Xin-Ying Chen et al.
· 0 citations