Research map / Reinforcement learning / RL from feedback
GR-GRPO: Graph-Diffused Credit for Autoregressive Image RL Alignment
Zheyu Zhang, Peng-Tao Jiang, Tianyi Zheng, Jian Zhang, Jinwei Chen, Bo Li
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 graph-regularized credit (GR-GRPO) to GRPO (ablation) | Janus-Pro-1B GRPO · GenEval | overall success rate: 0.84 -> 0.86 | Improved |
| add difficulty-aware budget allocation to GRPO alone (ablation) | Janus-Pro-1B GRPO · GenEval | overall success rate: 0.84 -> 0.84 | No clear effect |
| add difficulty-aware budget allocation on top of graph regularization (ablation) | GR-GRPO · GenEval | overall success rate: 0.86 -> 0.87 | Improved |
| use confidence-adaptive gate instead of constant gate (ablation) | GR-GRPO · GenEval | overall success rate: 0.83 -> 0.87 | Improved |
| propagate negative evidence in addition to positive (ablation) | GR-GRPO · GenEval | overall success rate: 0.87 -> 0.84 | Hurt |
| use K=64 instead of K=32 or 128 (ablation) | GR-GRPO · GenEval | overall success rate: 0.84 (K=32), 0.86 (K=128) -> 0.87 (K=64) | Improved |
| use intermediate diffusion strength 0.5 instead of 0.3 or 0.7 (ablation) | GR-GRPO · GenEval | overall success rate: 0.84 (0.3), 0.84 (0.7) -> 0.87 (0.5) | Improved |
| GR-GRPO vs vanilla GRPO on Janus-Pro-1B | Janus-Pro-1B · GenEval | overall success rate: 0.84 -> 0.87 | Improved |
| GR-GRPO vs vanilla GRPO on Janus-Pro-7B | Janus-Pro-7B · GenEval | overall success rate: 0.87 -> 0.92 | Improved |
| GR-GRPO vs STAGE on Color Attr. | Janus-Pro-7B · GenEval (Color Attr.) | success rate: 0.79 -> 0.74 | Hurt |
| GR-GRPO vs STAGE on DrawBench UnifiedReward | Janus-Pro-7B · DrawBench | UnifiedReward: 3.10 -> 3.07 | Hurt |
| GR-GRPO vs vanilla GRPO on DrawBench | Janus-Pro-1B · DrawBench | HPSv2.1: 27.22 -> 27.55 | Improved |
Closest papers by meaning
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.89)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.89)
- Delving into RL for Image Generation with CoT: A Study on DPO vs. GRPO (NeurIPS 2025 · similarity 0.88)
- GARDO: Reinforcing Diffusion Models without Reward Hacking (NeurIPS 2026 · similarity 0.87)
- Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO (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.
