Research map / Reinforcement learning / RL from feedback
Consolidating Reinforcement Learning for Multimodal Discrete Diffusion Models
Tianren Ma, Mu Zhang, Yibing Wang, Qixiang Ye
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 |
|---|---|---|---|
| apply MaskGRPO instead of diffu-GRPO | LLaDA-8B-Instruct · GSM8K (len 256) | pass@1 accuracy: 79.8 -> 84.2 | Improved |
| apply MaskGRPO instead of TraceRL | LLaDA-8B-Instruct · MATH500 (len 512) | pass@1 accuracy: 39.5 -> 41.5 | Improved |
| apply MaskGRPO instead of UniGRPO | LLaDA-8B-Instruct · MBPP (len 256) | pass@1 accuracy: 43.1 -> 45.4 | Improved |
| apply MaskGRPO instead of UniGRPO | MMaDA-8B-Base · GenEval | overall score: 0.63 -> 0.80 | Improved |
| add SFT on BLIP3-o data before MaskGRPO | MMaDA-8B-Base · DPG-Bench | overall score: 75.81 -> 81.76 | Improved |
| add SFT on BLIP3-o data before MaskGRPO | MMaDA-8B-Base · GenEval | overall score: 0.80 -> 0.84 | Improved |
| use emerge sampler instead of MaskGIT sampler before RL (ablation) | MMaDA-8B-Base · GenEval (before RL) | overall score: 0.56 -> 0.51 | Hurt |
| use emerge sampler instead of MaskGIT sampler during RL (ablation) | MMaDA-8B-Base · GenEval (after RL) | overall score: 0.77 -> 0.84 | Improved |
| add MaskGRPO RL to MMaDA | MMaDA-8B-Base · HPSv3 (DPG-Bench prompts) | HPSv3: 8.81 -> 9.40 | Improved |
| truncate timestep range (gamma 0.6) for language tasks (ablation) | MaskGRPO · GSM8K | training stability | Mixed |
| use high truncation gamma=0.8 instead of 0.4 for image RL (ablation) | MaskGRPO · image generation (HPSv3 reward) | training stability | Improved |
| use small clip epsilon for visual tasks (ablation) | MaskGRPO · image generation (HPSv3 reward) | training dynamics | Improved |
Closest papers by meaning
- Learning Multimodal Behaviors from Scratch with Diffusion Policy Gradient (NeurIPS 2024 · similarity 0.90)
- Reinforcement Learning Meets Masked Generative Models: Mask-GRPO for Text-to-Image Generation (NeurIPS 2025 · similarity 0.89)
- Training Diffusion Models with Reinforcement Learning (ICLR 2024 · similarity 0.89)
- Sketch Then Paint: Hierarchical Reinforcement Learning for Diffusion Multi-Modal Large Language Models (NeurIPS 2026 · similarity 0.89)
- Score Regularized Policy Optimization through Diffusion Behavior (ICLR 2024 · similarity 0.88)
- Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
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.
