Research map / Reinforcement learning / RL from feedback
GRPO-Guard: Mitigating Implicit Over-Optimization in Flow Matching via Regulated Clipping
Jing Wang, Jiajun Liang, Jie Liu, Henglin Liu, Gongye Liu, Jun Zheng, Wanyuan Pang, Ao Ma, Zhenyu Xie, Xintao Wang, Meng Wang, Pengfei Wan and 1 more
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 FlowGRPO with GRPO-Guard (RatioNorm + gradient reweighting) | FlowGRPO, SD3.5-M · GenEval proxy, 1860 steps | average gold score: 0.84 -> 0.89 | Improved |
| replace FlowGRPO with GRPO-Guard | FlowGRPO, SD3.5-M · TextRender proxy, 480 steps | average gold score: 0.88 -> 0.99 | Improved |
| replace FlowGRPO with GRPO-Guard | FlowGRPO, SD3.5-M · PickScore proxy, 1020 steps | average gold score: 1.16 -> 1.20 | Improved |
| replace DanceGRPO with GRPO-Guard | DanceGRPO, Flux.1-dev · GenEval proxy, 1260 steps | average gold score: 0.88 -> 1.02 | Improved |
| replace DanceGRPO with GRPO-Guard | DanceGRPO, Flux.1-dev · TextRender proxy, 540 steps | average gold score: 0.96 -> 1.02 | Improved |
| add RatioNorm mean correction to FlowGRPO (ablation) | FlowGRPO, SD3.5-M · OCR task, 480 steps | gold score | Improved |
| add inter-step variance alignment (full RatioNorm) on top of mean correction (ablation) | FlowGRPO, SD3.5-M · OCR task, 480 steps | gold score / proxy growth | Mixed |
| use TempFlowGRPO reweighting instead of 1/dt gradient balancing (ablation) | FlowGRPO, SD3.5-M · OCR task, 480 steps | optimization speed / gold score | Mixed |
| add 1/dt gradient reweighting on top of RatioNorm (ablation) | GRPO-Guard, SD3.5-M · OCR task, 480 steps | gold score | Improved |
Closest papers by meaning
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.91)
- TIC-GRPO: Provable and Efficient Optimization for Reinforcement Learning from Human Feedback (NeurIPS 2026 · similarity 0.90)
- Flow-DPPO: Divergence Proximal Policy Optimization for Flow Matching Models (NeurIPS 2026 · similarity 0.90)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (NeurIPS 2026 · similarity 0.88)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (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.
