Research map / Image generation / Fast diffusion sampling
Generalized Discrete Diffusion with Self-Correction
Linxuan Wang, Ziyi Wang, Yikun Bai, Wei Deng, Guang Lin, Qifan Song
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 explicit uniform transitions without remasking (SCDD) instead of GIDD+ | masked diffusion LM, p_u=0.2 · OWT, 32 steps | generative perplexity: 90.5 -> 74.5 | Improved |
| use SCDD instead of GIDD+ | masked diffusion LM, p_u=0.2 · LM1B, 16 steps | generative perplexity: 192.7 -> 159.2 | Improved |
| use SCDD instead of GIDD+ at 1024 steps | masked diffusion LM, p_u=0.2 · OWT, 1024 steps | generative perplexity: 71.2 -> 55.7 | Improved |
| use SCDD instead of ReMDM-cap at 32-step generation | masked diffusion LM · 32-step generation | generative perplexity: 55% decrease | Improved |
| add uniform noise during training (SCDD p_u=0.1 vs MDLM) | MDLM · LM1B | validation perplexity: 32.98 -> 39.16 | Hurt |
| add uniform noise during training (SCDD p_u=0.1 vs MDLM) | MDLM · OWT | validation perplexity: 24.72 -> 28.41 | Hurt |
| use SCDD instead of GIDD+ for likelihood | best variants · LM1B and OWT | validation perplexity: 3.7% and 9.9% decrease | Improved |
| increase p_u from 0.1 to 0.2 in SCDD (ablation) | SCDD · LM1B | validation perplexity: 39.16 -> 46.54 | Hurt |
| use SCDD instead of GIDD+ for correction | p_u=0.2, OWT, 1024 steps · OWT | correction rate: 0.40 -> 0.75 | Improved |
| increase maximum uniform noise ratio p_u (ablation) | SCDD · Wikitext-103 | total corrections | Improved |
| shift peak uniform noise later (t_peak 0.75 instead of 0.25) (ablation) | SCDD · Wikitext-103 | timing of correction: nearly 40% of corrections in first 200 steps | Improved |
| apply one SCDD denoising step to corrupted sequences (K=50) | SCDD p_u=0.2 · OWT validation | generative perplexity: 154.8 -> 25.5 | Improved |
Closest papers by meaning
- Generalized Interpolating Discrete Diffusion (ICML 2025 · similarity 0.90)
- Learn from Your Mistakes: Self-Correcting Masked Diffusion Models (NeurIPS 2026 · similarity 0.88)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.88)
- Diffusion Model is Effectively Its Own Teacher (CVPR 2025 · similarity 0.87)
- Remasking Discrete Diffusion Models with Inference-Time Scaling (NeurIPS 2025 · similarity 0.87)
- Informed Correctors for Discrete Diffusion Models (NeurIPS 2025 · similarity 0.87)
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.
