Research map / Image generation / Fast diffusion sampling
Bringing Stability to Diffusion: Decomposing and Reducing Variance of Training Masked Diffusion Models
Mengni Jia, Mengyu Zhou, Yihao Liu, xiaoxi jiang, guanjunjiang
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 P-POTS+MIRROR to standard MDM training | LLaDA-8B-Instruct · GSM8K | accuracy (%): 52.51 -> 60.53 | Improved |
| add P-POTS+MIRROR to standard training | LLaDA-8B-Instruct · HiTab | accuracy (%): 58.68 -> 67.10 | Improved |
| add P-POTS+MIRROR to standard training | LLaDA-8B-Instruct · OpenScience | accuracy (%): 45.52 -> 52.53 | Improved |
| add P-POTS+MIRROR to standard training | MMaDA-8B-MixCoT · text-to-image-2M | CLIP score: 31.36 -> 34.59 | Improved |
| add MIRROR on top of P-POTS (ablation) | P-POTS · GSM8K | accuracy (%): 58.58 -> 60.53 | Improved |
| add MIRROR on top of P-POTS (ablation) | P-POTS · OpenScience | accuracy (%): 46.80 -> 52.53 | Improved |
| add P-POTS on top of MIRROR (ablation) | MIRROR · HiTab | accuracy (%): 64.48 -> 67.10 | Improved |
| add MIRROR alone to standard training (ablation) | LLaDA-8B-Instruct · GSM8K | accuracy (%): 52.51 -> 53.70 | Improved |
| add P-POTS alone to standard training (ablation) | MMaDA-8B-MixCoT · text-to-image-2M | CLIP score: 31.36 -> 31.65 | No clear effect |
| add MIRROR alone to standard training (ablation) | MMaDA-8B-MixCoT · text-to-image-2M | CLIP score: 31.36 -> 34.07 | Improved |
| use P-POTS+MIRROR instead of clipped noise schedule | LLaDA-8B-Instruct · GSM8K | accuracy (%): 56.18 -> 60.53 | Improved |
| add stratified t-sampling (StraTS) to standard training (ablation) | LLaDA-8B-Instruct · GSM8K (seed 42) | accuracy (%): 50.64 -> 55.57 | Improved |
Closest papers by meaning
- Learn from Your Mistakes: Self-Correcting Masked Diffusion Models (NeurIPS 2026 · similarity 0.92)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.92)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.92)
- Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions (ICML 2025 · similarity 0.91)
- Remasking Discrete Diffusion Models with Inference-Time Scaling (NeurIPS 2025 · similarity 0.91)
- Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking (NeurIPS 2025 · 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.
