Research map / Reinforcement learning / RL from feedback
Fine-Grained GRPO for Precise Preference Alignment in Flow Models
Yujie Zhou, Pengyang Ling, Jiazi Bu, Yibin Wang, Yuhang Zang, Jiaqi Wang, Li Niu, Guangtao Zhai
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 Singular Stochastic Sampling (single-step SDE) | DanceGRPO · HPSv2 test prompts (trained on HPS reward only) | HPS-v2.1: 0.353 -> 0.376 | Improved |
| add Singular Stochastic Sampling | DanceGRPO · HPSv2 test prompts (trained on HPS reward only) | CLIP Score: 0.375 -> 0.351 | Hurt |
| add Singular Stochastic Sampling | DanceGRPO · HPSv2 test prompts (trained on HPS+CLIP reward) | UnifiedReward: 3.569 -> 3.688 | Improved |
| add Multi-Granularity Advantage Integration (ablation) | G2RPO w/o MGAI · HPSv2 test prompts (trained on HPS+CLIP reward) | UnifiedReward: 3.688 -> 3.783 | Improved |
| add Multi-Granularity Advantage Integration (ablation) | G2RPO w/o MGAI · HPSv2 test prompts (trained on HPS+CLIP reward) | ImageReward: 1.421 -> 1.483 | Improved |
| add Multi-Granularity Advantage Integration (ablation) | G2RPO w/o MGAI · HPSv2 test prompts (trained on HPS reward only) | HPS-v2.1: 0.376 -> 0.385 | Improved |
| add Multi-Granularity Advantage Integration (ablation) | G2RPO w/o MGAI · HPSv2 test prompts (trained on HPS reward only) | CLIP Score: 0.351 -> 0.355 | No clear effect |
| use granularities {1,2,3} instead of {1,2} (ablation) | G2RPO · HPSv2 test prompts (trained on HPS+CLIP reward) | HPS-v2.1: 0.375 -> 0.376 | No clear effect |
| use granularities {1,2,3} instead of single granularity {1} (ablation) | G2RPO · HPSv2 test prompts (trained on HPS+CLIP reward) | ImageReward: 1.421 -> 1.483 | Improved |
| use granularity {1,3} instead of {1,2,3} (ablation) | G2RPO · HPSv2 test prompts (trained on HPS+CLIP reward) | HPS-v2.1: 0.378 vs 0.376 | No clear effect |
| use G2RPO instead of MixGRPO | MixGRPO · HPSv2 test prompts (trained on HPS+CLIP reward) | UnifiedReward: 3.661 -> 3.783 | Improved |
| use G2RPO instead of MixGRPO | MixGRPO · HPSv2 test prompts (trained on HPS reward only) | CLIP Score: 0.358 -> 0.355 | Hurt |
Closest papers by meaning
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.91)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.90)
- Using Human Feedback to Fine-tune Diffusion Models without Any Reward Model (CVPR 2024 · similarity 0.89)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.89)
- DACA-GRPO: Denoising-Aware Credit Assignment for Reinforcement Learning in Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.88)
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.
