Research map / Reinforcement learning / RL from feedback
UDM-GRPO: Stable and Efficient Group Relative Policy Optimization for Uniform Discrete Diffusion Models
Jiaqi Wang, Haoge Deng, Ting Pan, Yang Liu, Chengyuan Wang, Fan Zhang, Yonggang Qi, Xinlong Wang
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 |
|---|---|---|---|
| use final clean sample as action instead of intermediate prediction x1^t (ablation) | URSA + GRPO (backward trajectory) · GenEval | overall score: 0.84 -> 0.89 | Improved |
| use final clean sample as action instead of intermediate prediction x1^t (ablation) | URSA + GRPO (backward trajectory) · PickScore | PickScore: 21.99 -> 23.10 | Improved |
| use final clean sample as action instead of intermediate prediction x1^t (ablation) | URSA + GRPO (backward trajectory) · OCR | accuracy: 0.23 -> 0.23 | No clear effect |
| train on forward-process trajectory instead of backward trajectory (ablation) | URSA + GRPO (final clean sample action) · GenEval | overall score: 0.89 -> 0.94 | Improved |
| train on forward-process trajectory instead of backward trajectory (ablation) | URSA + GRPO (final clean sample action) · PickScore | PickScore: 23.10 -> 23.51 | Improved |
| train on forward-process trajectory instead of backward trajectory (ablation) | URSA + GRPO (final clean sample action) · OCR | accuracy: 0.23 -> 0.34 | Improved |
| remove classifier-free guidance during training (CFG-Free) (ablation) | UDM-GRPO · GenEval | overall score: 0.94 -> 0.96 | Improved |
| remove classifier-free guidance during training (CFG-Free) (ablation) | UDM-GRPO · PickScore | PickScore: 23.51 -> 23.81 | Improved |
| remove classifier-free guidance during training (CFG-Free) (ablation) | UDM-GRPO · OCR | accuracy: 0.34 -> 0.57 | Improved |
| optimize only three consecutive early high-noise timesteps instead of all or random timesteps (Reduced-Step) (ablation) | UDM-GRPO · GenEval and PickScore | reward / convergence speed | Improved |
| optimize only three consecutive early high-noise timesteps instead of all or random timesteps (Reduced-Step) (ablation) | UDM-GRPO · OCR | accuracy | No clear effect |
| apply naive Flow-GRPO adaptation (intermediate action, backward trajectory) to URSA (ablation) | URSA · long training | reward stability | Hurt |
Closest papers by meaning
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.89)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Co-GRPO: Co-Optimized Group Relative Policy Optimization for Masked Image Generation (NeurIPS 2026 · similarity 0.88)
- Delving into RL for Image Generation with CoT: A Study on DPO vs. GRPO (NeurIPS 2025 · similarity 0.87)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.87)
- Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model (CVPR 2024 · 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.
