Research map / Reinforcement learning / RL from feedback
Distillation Models are Good Samplers for Diffusion Reinforcement Learning
Zunxu Liu, Aiqiu Wu, Zhaofan Qiu, 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 |
|---|---|---|---|
| use late distillation sampling (S3) instead of static fully distilled sampling (S2) (ablation) | RL with distilled sampler · Image OCR | OCR accuracy: 0.79 -> 0.95 | Improved |
| use late distillation sampling (S3) instead of early distillation sampling (S4) (ablation) | RL with distilled sampler · Image OCR | OCR accuracy: 0.81 -> 0.95 | Improved |
| use late distillation sampling (S3) instead of static fully distilled sampling (S2) (ablation) | RL with distilled sampler · Video OCR | OCR accuracy: 0.26 -> 0.42 | Improved |
| replace vanilla 40-step sampling (S1) with late distillation sampling (S3) (ablation) | RL with distilled sampler · Image OCR | training time (GPU hours): 160 -> 16 | Improved |
| add reward-aware distillation (ablation) | DMSampler · Image OCR | OCR accuracy: 0.96 -> 0.97 | Improved |
| add reward-aware distillation (ablation) | DMSampler · Video OCR | OCR accuracy: 0.48 -> 0.50 | Improved |
| add iterative co-evolution phases (RL phase 1 to 3) (ablation) | DMSampler · Image OCR | policy OCR accuracy: 0.95 -> 0.97 | Improved |
| add iterative co-evolution phases (RL phase 1 to 3) (ablation) | DMSampler · Video OCR | policy OCR accuracy: 0.42 -> 0.50 | Improved |
| use distilled sampler instead of direct step reduction with CFG (DRS3, 10 steps) | DMSampler · Video OCR | OCR accuracy: 0.23 -> 0.50 | Improved |
| use DMSampler instead of DiffusionNFT with 50-step CFG sampling | DiffusionNFT · VBench | total score: 84.74 -> 85.00 | Improved |
| add DMSampler to FlowGRPO | FlowGRPO · GenEval | GenEval score: 0.95 -> 0.96 | Improved |
| add DMSampler to DGPO | DGPO · GenEval | GenEval score: 0.97 -> 0.97 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- DiffusionNFT — DiffusionNFT: Online Diffusion Reinforcement with Forward Process (ICLR 2026) · better on 3 of 3 shared results
- DGPO — DGPO: Discovering Multiple Strategies with Diversity-Guided Policy Optimization (AAAI 2024) · better on 0 of 1 shared result
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
- Score Regularized Policy Optimization through Diffusion Behavior (ICLR 2024 · similarity 0.91)
- Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View (NeurIPS 2026 · similarity 0.91)
- Diffusion Policies Creating a Trust Region for Offline Reinforcement Learning (NeurIPS 2024 · similarity 0.90)
- FASTER: Value-Guided Sampling for Fast RL (NeurIPS 2026 · similarity 0.90)
- Learning to Sample From Diffusion Models via Inverse Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · 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.
