Research map / Reinforcement learning / RL from feedback
Embedding-perturbed Exploration Preference Optimization for Flow Models
Sujie Hu, Chubin Chen, Jiashu Zhu, Jiahong Wu, Xiangxiang Chu, Xiu 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 embedding-level perturbation (E2PO) to DiffusionNFT-style training | SD3.5-M with DiffusionNFT · GenEval (G=4, K=12) | GenEval score: 0.868 -> 0.932 | Improved |
| add embedding-level perturbation (E2PO) at small group size | SD3.5-M with DiffusionNFT · GenEval (G=2, K=4) | GenEval score: 0.822 -> 0.917 | Improved |
| add E2PO versus DiffusionNFT with large group size G=24 | SD3.5-M · GenEval | GenEval score: 0.922 -> 0.932 | Improved |
| increase latent group size G from 4 to 48 (ablation) | baseline group-based RL · intra-group reward variance | zero-std group ratio / reward std | No clear effect |
| add embedding perturbations instead of larger latent group | group-based RL · training dynamics | intra-group reward std | Improved |
| use a single noise seed (G=1) with semantic perturbations only (ablation) | E2PO · GenEval / PickScore | reward | Hurt |
| balanced split G=4, K=12 versus extreme G=1 or K=1 at budget N=48 (ablation) | E2PO · GenEval | reward | Improved |
| add noise-aware sampling schedule instead of static perturbed conditions (ablation) | E2PO · qualitative at step 150 | visual fidelity | Improved |
| perturb both CLIP-L and OpenCLIP-G instead of CLIP-L only (ablation) | E2PO · PickScore reward | reward stability | Hurt |
| add reference-anchored batching (keep original prompt as anchor) (ablation) | E2PO · training reward | convergence and asymptotic reward | Improved |
| add E2PO to DiffusionNFT on PickScore training | SD3.5-M · PickScore (G=8) | PickScore: 23.31 -> 23.38 | Improved |
| add E2PO to DiffusionNFT on DivGenBench diversity | SD3.5-M · DivGenBench PVS | PVS: 0.282 -> 0.301 | 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 3 of 3 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
- Flow-GRPO: Training Flow Matching Models via Online RL (NeurIPS 2025 · similarity 0.86)
- On the Effect of Negative Gradient in Group Relative Deep Reinforcement Optimization (NeurIPS 2025 · similarity 0.86)
- Beyond Bradley-Terry Models: A General Preference Model for Language Model Alignment (ICML 2025 · similarity 0.86)
- TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment (NeurIPS 2026 · similarity 0.86)
- Random Policy Evaluation Uncovers Policies of Generative Flow Networks (ICML 2025 · similarity 0.85)
- PRECISE: SDE-Consistent Stochastic Sampling for RL Post-Training of Flow-Matching Models (NeurIPS 2026 · similarity 0.85)
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.
