Research map / Reinforcement learning / RL from feedback
On-Policy Diffusion Reinforcement Learning Meets Off-Policy Quality Anchoring
Zunxu Liu, Zhaofan Qiu, Yazhen Xie, Yingwei Pan, Ting Yao, Tao Mei
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 reward renormalization with off-policy samples (ablation) | DiffusionNFT on OCR · OCR | GPU hours: 64 -> 56 GPU hours, OCR 0.970 -> 0.969 | Improved |
| add perceptual anchored reward on top of renormalization (ablation) | DiffusionCompass · OCR (out-of-domain) | Aesthetic score: 4.48 -> 5.03 | Improved |
| add anchoring (off-policy) loss on top of renormalization and anchored reward (ablation) | DiffusionCompass · OCR | OCR accuracy: 0.963 -> 0.972, Aes. 5.03 -> 4.88 is not a gain | Improved |
| add distribution-aware filtering (ablation) | DiffusionCompass · OCR | OCR accuracy: 0.972 -> 0.975 | Improved |
| mix on-policy and off-policy RL instead of pure on-policy (ablation) | SD3.5-M · OCR | OCR accuracy: 0.970 -> 0.975 | Improved |
| mix on-policy and off-policy RL instead of pure off-policy (ablation) | SD3.5-M · OCR | OCR accuracy: 0.836 -> 0.975 | Improved |
| raise off-policy loss weight lambda2 from 0.1 to 0.5 (ablation) | DiffusionCompass · GenEval | GenEval overall: 0.97 -> 0.92 (OCR 0.975 -> 0.959) | Hurt |
| set filtering strength alpha to 1000 instead of 100 (ablation) | DiffusionCompass · GenEval | GenEval overall: 0.97 -> 0.94 (OCR 0.975 -> 0.961) | Hurt |
| use SD3.5-M own initial samples as anchors instead of Qwen-Image (ablation) | DiffusionCompass · OCR | OCR accuracy: 0.975 -> 0.973 | No clear effect |
| use Flux anchors instead of Qwen-Image (ablation) | DiffusionCompass · OCR | Aesthetic score: 5.00 -> 4.94 | No clear effect |
| add off-policy anchoring to FlowGRPO | FlowGRPO · OCR | OCR accuracy: 0.915 -> 0.939 | Improved |
| add off-policy anchoring to DiffusionNFT | DiffusionNFT · ImageReward | ImageReward: 1.64 -> 1.70 (VideoOCR 0.412 -> 0.433) | Improved |
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.
- DiffusionNFT — DiffusionNFT: Online Diffusion Reinforcement with Forward Process (ICLR 2026) · better on 8 of 8 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
- Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Score Regularized Policy Optimization through Diffusion Behavior (ICLR 2024 · similarity 0.90)
- Diffusion Policies Creating a Trust Region for Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Learning a Diffusion Model Policy from Rewards via Q-Score Matching (ICML 2024 · similarity 0.89)
- Diffusion Actor-Critic: Formulating Constrained Policy Iteration as Diffusion Noise Regression for Offline Reinforcement Learning (ICLR 2025 · similarity 0.89)
- Confronting Reward Overoptimization for Diffusion Models: A Perspective of Inductive and Primacy Biases (ICML 2024 · similarity 0.89)
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.
