Research map / Image generation / Fast diffusion sampling
Ultra-Fast Language Generation via Discrete Diffusion Divergence Instruct
Haoyang Zheng, Xinyang Liu, Xiangrui Kong, Nan Jiang, Zheyuan Hu, Weijian Luo, Wei Deng, Guang Lin
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 score decomposition (ablation) | DiDi-Instruct · OpenWebText (8 NFEs) | generative PPL: 33584 -> 62.236 | Improved |
| add coupled time t (ablation) | DiDi-Instruct · OpenWebText (8 NFEs) | generative PPL: 360.75 -> 62.236 | Improved |
| add omega(t) correction (ablation) | DiDi-Instruct · OpenWebText (8 NFEs) | generative PPL: 82.489 -> 62.236 | Improved |
| add pi(t) weighting (ablation) | DiDi-Instruct · OpenWebText (8 NFEs) | generative PPL: 69.656 -> 62.236 | Improved |
| add regularization (ablation) | DiDi-Instruct · OpenWebText (8 NFEs) | generative PPL: 84.594 -> 62.236 | Improved |
| add regularization (ablation) | DiDi-Instruct · OpenWebText (128 NFEs) | generative PPL: 18.446 vs 17.499 without it | Hurt |
| add guided inference (RGAS) (ablation) | DiDi-Instruct · OpenWebText (8 NFEs) | generative PPL: 88.274 -> 62.236 | Improved |
| add guided inference (RGAS) (ablation) | DiDi-Instruct · OpenWebText (128 NFEs) | generative entropy: 5.00 -> 5.15 | Improved |
| add score decomposition (cumulative) (ablation) | Baseline without tricks · OpenWebText (8 NFEs) | generative PPL: 803.922 -> 667.830 | Improved |
| add coupled time t (cumulative) (ablation) | Baseline + score decomposition · OpenWebText (8 NFEs) | generative PPL: 667.830 -> 101.019 | Improved |
| add guided inference (cumulative) (ablation) | DiDi-Instruct with regularization · OpenWebText (8 NFEs) | generative PPL: 88.274 -> 62.236 | Improved |
| distill with DiDi-Instruct instead of 1024-step teacher | MDLM 169M · OpenWebText (16 NFEs) | generative PPL: 38.53 -> 38.188 | Improved |
Closest papers by meaning
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.88)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.88)
- Pruning Large Language Models with Semi-Structural Adaptive Sparse Training (AAAI 2025 · similarity 0.88)
- Large Language Diffusion Models (NeurIPS 2025 · similarity 0.88)
- The Diffusion Duality (ICML 2025 · similarity 0.88)
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · 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.
