Research map / Reinforcement learning / RL from feedback
Enhancing Reasoning for Diffusion LLMs via Distribution Matching Policy Optimization
Yuchen Zhu, Wei Guo, Jaemoo Choi, Petr Molodyk, Bo Yuan, Molei Tao, Yongxin Chen
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 DMPO instead of the strongest non-DMPO RL baseline | LLaDA-8B-Instruct · Sudoku (length 256) | accuracy (%): +39.63% | Improved |
| apply DMPO instead of the strongest non-DMPO RL baseline | LLaDA-8B-Instruct · Countdown (length 256) | accuracy (%): +13.67% | Improved |
| apply DMPO instead of the strongest non-DMPO RL baseline | LLaDA-8B-Instruct · Sudoku (length 512) | accuracy (%): +31.57% | Improved |
| apply DMPO to LLaDA-Instruct | LLaDA-8B-Instruct · Countdown | accuracy (%): +59.38% | Improved |
| apply DMPO to LLaDA-Instruct | LLaDA-8B-Instruct · GSM8K | accuracy (%): +2.40% | Improved |
| apply DMPO instead of d1 | Dream-Instruct-7B · Countdown | accuracy (%): +22.32% | Improved |
| apply DMPO instead of d1 | Dream-Instruct-7B · MATH500 | accuracy (%): +3.90% | Improved |
| apply DMPO to SFT-tuned LLaDA instead of d1 | LLaDA-SFT · GSM8K (length 128) | accuracy (%): +4.78% | Improved |
| apply DMPO to LLaDA-1.5 (DPO-tuned) | LLaDA-1.5 · GSM8K (length 256) | accuracy (%): 80.97 -> 82.71 | Improved |
| add weight baseline subtraction (ablation) | DMPO with small batch · Sudoku (small batch) | reward | Improved |
| use WDDO loss instead of WDCE (ablation) | DMPO · Sudoku (small batch) | reward | Mixed |
| use DMPO off-policy reuse instead of diffu-GRPO | dLLM RL · Countdown | sample efficiency: 2~3x | Improved |
Closest papers by meaning
- d1: Scaling Reasoning in Diffusion Large Language Models via Reinforcement Learning (NeurIPS 2025 · similarity 0.94)
- Diversity-Aware Policy Optimization for Large Language Model Reasoning (NeurIPS 2025 · similarity 0.91)
- Break the Block: Dynamic-size Reasoning Blocks for Diffusion Large Language Models via Monotonic Entropy Descent with Reinforcement Learning (ICML 2026 · similarity 0.91)
- Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages (NeurIPS 2026 · similarity 0.91)
- Accelerating RL for LLM Reasoning with Optimal Advantage Regression (NeurIPS 2025 · similarity 0.91)
- The Flexibility Trap: Rethinking the Value of Arbitrary Order in Diffusion Language Models (ICML 2026 · similarity 0.91)
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.
