Research map / Image generation / Fast diffusion sampling
Alleviating Sparse Rewards by Modeling Step-Wise and Long-Term Sampling Effects in Flow-Based GRPO
Yunze Tong, Mushui Liu, Canyu Zhao, Wanggui He, Shiyi Zhang, Peng Zhang, Hongwei Zhang, Jinlong Liu, Hao Jiang
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 |
|---|---|---|---|
| add TP-GRPO (turning-point aggregated rewards, no constraint) to Flow-GRPO | Flow-GRPO on SD3.5-M · GenEval | GenEval score: 0.9673 -> 0.9714 | Improved |
| add the first-step consistency constraint to TP-GRPO (ablation) | TP-GRPO on SD3.5-M · GenEval | GenEval score: 0.9714 -> 0.9725 | Improved |
| add the first-step consistency constraint to TP-GRPO (ablation) | TP-GRPO on SD3.5-M · Visual Text Rendering (OCR) | OCR accuracy: 0.9718 -> 0.9651 | Hurt |
| add the first-step consistency constraint to TP-GRPO (ablation) | TP-GRPO on SD3.5-M · Human Preference Alignment (test prompts) | PickScore: 24.73 -> 24.67 | No clear effect |
| add TP-GRPO (w constraint) to Flow-GRPO | Flow-GRPO on SD3.5-M · Human Preference Alignment (test prompts) | PickScore: 24.02 -> 24.67 | Improved |
| add TP-GRPO (w/o constraint) to Flow-GRPO | Flow-GRPO on SD3.5-M · Visual Text Rendering (OCR) | OCR accuracy: 0.9579 -> 0.9718 | Improved |
| replace TempFlow-GRPO branching process reward with TP-GRPO (w constraint) | SD3.5-M on DrawBench ImageReward (text rendering policy) · DrawBench (OCR-trained policy) | ImageReward: 0.6316 -> 0.7378 | Improved |
| add TP-GRPO (w/o constraint) to Flow-GRPO | Flow-GRPO on SD3.5-M · DrawBench (GenEval-trained policy) | ImageReward: 0.8792 -> 0.9267 | Improved |
| add TP-GRPO (w/o constraint) to Flow-GRPO | Flow-GRPO on SD3.5-M · DrawBench (PickScore-trained policy) | ImageReward: 1.3875 -> 1.3714 | No clear effect |
| shrink SDE sampling window moderately (to size 8) from all steps (ablation) | TP-GRPO on SD3.5-M · Training curve over 2400 steps | reward | Improved |
| shrink SDE sampling window too far (to 4 steps) (ablation) | TP-GRPO on SD3.5-M · Training curve over 2400 steps | reward | Hurt |
| lower noise scale alpha to 0.4 from 0.7 (ablation) | TP-GRPO on SD3.5-M · Training curve | reward: 0.7 -> 0.4 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TempFlow-GRPO — TEMPFLOW-GRPO: WHEN TIMING MATTERS FOR GRPO IN FLOW MODELS (ICLR 2026) · better on 9 of 9 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DenseGRPO: From Sparse to Dense Reward for Flow Matching Model Alignment (ICLR 2026 · similarity 0.91)
- TEMPFLOW-GRPO: WHEN TIMING MATTERS FOR GRPO IN FLOW MODELS (ICLR 2026 · similarity 0.90)
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.90)
- GARDO: Reinforcing Diffusion Models without Reward Hacking (NeurIPS 2026 · similarity 0.88)
- Towards Better Alignment: Training Diffusion Models with Reinforcement Learning Against Sparse Rewards (CVPR 2025 · similarity 0.88)
- Diff-Instruct with Diffused Reward: Towards Principled One-step Generator RL (NeurIPS 2026 · 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.
