Research map / Image generation / Fast diffusion sampling
Improving Discrete Diffusion Unmasking Policies Beyond Explicit Reference Policies
Chunsan Hong, Seonho An, Min-Soo Kim, Jong Chul 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 |
|---|---|---|---|
| use learned unmasking policy instead of max-confidence | LLaDA-8B-Instruct · Sudoku | accuracy: 70.5% -> 81.7% | Improved |
| use learned unmasking policy instead of max-confidence | LLaDA-8B-Instruct · Zebra | accuracy: 33.7% -> 36.2% | Improved |
| use learned Top-K unmasking policy instead of max-confidence | LLaDA-8B-Instruct · GSM8K | accuracy: 68.4% -> 70.3% | Improved |
| use learned Top-K unmasking policy instead of max-confidence | LLaDA-8B-Instruct · Math500 | accuracy: 27.2% -> 28.4% | Improved |
| use learned policy instead of margin baseline | LLaDA-8B-Instruct · Math500 | accuracy: 0.284 -> 0.284 | No clear effect |
| add learned unmasking policy on top of diffu-GRPO model (vs random) | diffu-GRPO LLaDA · GSM8K | accuracy: 0.638 -> 0.764 | Improved |
| add learned unmasking policy on top of diffu-GRPO model (vs max-confidence) | diffu-GRPO LLaDA · GSM8K | accuracy: 0.751 -> 0.764 | Improved |
| add divergence regularization to reference policy (vs random init without it, DColT) (ablation) | UPO policy · Sudoku | accuracy: at least 2% | Improved |
| add divergence regularization to reference policy (vs random init without it, DColT) (ablation) | UPO policy · GSM8K | accuracy: at least 3% | Improved |
| use max-confidence realization instead of Top-K (ablation) | UPO policy · Sudoku | accuracy | Improved |
| use Top-K realization instead of max-confidence (ablation) | UPO policy · GSM8K | accuracy | Improved |
| add divergence term (CE to g_conf / KL to Top-K) to loss (ablation) | UPO policy · Sudoku and GSM8K | accuracy | Improved |
Closest papers by meaning
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.93)
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · similarity 0.91)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.91)
- Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions (ICML 2025 · similarity 0.91)
- Learn from Your Mistakes: Self-Correcting Masked Diffusion Models (NeurIPS 2026 · similarity 0.90)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.90)
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.
