Research map / Reinforcement learning / RL from feedback
Expand and Prune: Maximizing Trajectory Diversity for Effective GRPO in Generative Models
Shiran Ge, Chenyi Huang, Yuang Ai, Qihang Fan, Huaibo Huang, Ran He
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 Flow-GRPO with Pro-GRPO (expand-and-prune) | SD 3.5-M · DrawBench | In-domain PickScore: 23.322 -> 24.008 | Improved |
| Replace Flow-GRPO with Pro-GRPO | SD 3.5-M · HPSv2 | ImageReward (OOD): 1.463 -> 1.509 | Improved |
| Replace Flow-GRPO with Pro-GRPO | SD 3.5-M · GenEval Overall | Overall: 0.719 -> 0.726 | Improved |
| Replace Flow-GRPO with Pro-GRPO | SD 3.5-M · GenEval Position | accuracy: 0.288 -> 0.323 | Improved |
| Use Pro-GRPO-Flash (24->12 pruning, no expansion) | SD 3.5-M · DrawBench | In-domain PickScore and speedup: 23.322 -> 23.722; 1.41x speedup | Improved |
| Replace DanceGRPO with Pro-GRPO (HPS reward) | SD v1.4 · HPSv2.1 in-domain | HPSv2.1: 0.369 -> 0.391 | Improved |
| Replace DanceGRPO with Pro-GRPO (HPS reward) | SD v1.4 · HPSv2 | ImageReward (OOD): 1.079 -> 1.114 | Improved |
| Replace DanceGRPO with Pro-GRPO (HPS+CLIP reward) | SD v1.4 · DrawBench | ImageReward (OOD): 0.835 -> 0.974 | Improved |
| Pro-GRPO vs DanceGRPO (HPS reward) on PickScore | SD v1.4 · DrawBench | PickScore (OOD): 21.903 -> 21.896 | No clear effect |
| Increase initial group size Gmax from 32 to 48 (K=8) (ablation) | Pro-GRPO SD v1.4 · HPSv2.1 reward | in-domain HPSv2.1: 0.386 -> 0.391 | Improved |
| Increase Gmax from 48 to 64 (K=8) (ablation) | Pro-GRPO SD v1.4 · HPSv2.1 reward | in-domain HPSv2.1: 0.391 -> 0.393 | No clear effect |
| Move pruning checkpoints from {10,20} to {30,40} (ablation) | Pro-GRPO SD v1.4 (32->8) · HPSv2.1 reward | in-domain HPSv2.1: 0.373 -> 0.391 | Improved |
Closest papers by meaning
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.89)
- GVPO: Group Variance Policy Optimization for Large Language Model Post-Training (NeurIPS 2025 · similarity 0.88)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.88)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.88)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.87)
- Co-GRPO: Co-Optimized Group Relative Policy Optimization for Masked Image Generation (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.
