Research map / Image generation / Fast diffusion sampling
Stepwise Credit Assignment for GRPO on Flow-Matching Models
Yash Savani, Branislav Kveton, Yuchen Liu, Yilin Wang, Jing Shi, Subhojyoti Mukherjee, Nikos Vlassis, Krishna Kumar Singh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace uniform final-image advantage with stepwise gain-based advantages | Flow-GRPO · GenEval (cfg=4.5) | Overall accuracy: 0.68 -> 0.71 | Improved |
| Replace uniform final-image advantage with stepwise gain-based advantages | Flow-GRPO · GenEval (cfg=4.5) | Counting accuracy: 0.64 -> 0.70 | Improved |
| Replace uniform final-image advantage with stepwise gain-based advantages | Flow-GRPO · GenEval (cfg=1.0) | Overall accuracy: 0.60 -> 0.60 | No clear effect |
| Replace uniform final-image advantage with stepwise gain-based advantages | Flow-GRPO · Reward per training step (4 reward settings) | Reward | Improved |
| Add stepwise credit assignment on top of DDIM-inspired SDE (ablation) | Flow-GRPO with DDIM SDE · Reward per training step | Reward | Improved |
| Use DDIM-inspired SDE instead of Flow-GRPO SDE (ablation) | Flow-GRPO and Stepwise-Flow-GRPO · Reward per training step | Reward | Improved |
| Use 5 ODE substeps for intermediate reward estimates (vs 1 to 3) (ablation) | Stepwise-Flow-GRPO · Reward estimate quality (appendix) | Training signal | Improved |
| Use more than 5 ODE substeps (6 to 10) (ablation) | Stepwise-Flow-GRPO · Reward estimate quality (appendix) | Training signal | No clear effect |
| Use joint normalization of gains instead of per-step normalization (ablation) | Stepwise-Flow-GRPO · Convergence (appendix) | Convergence speed | Improved |
Closest papers by meaning
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.91)
- Flow-DPPO: Divergence Proximal Policy Optimization for Flow Matching Models (NeurIPS 2026 · similarity 0.89)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.88)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.88)
- DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models (NeurIPS 2026 · similarity 0.87)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
