Research map / Image generation / Fast diffusion sampling
Simple Denoising Diffusion Language Models
Huaisheng Zhu, Zhengyu Chen, Shijie Zhou, Zhihui Xie, Yige Yuan, Shiqi Chen, Zhimeng Guo, Siyuan Xu, Hangfan Zhang, Vasant G. Honavar, Teng Xiao
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 anti-uniform sharpening regularization (negative gradient with random tokens) (ablation) | SDDLM · OWT (170M) | Gen PPL: 77.07 -> 45.18 | Improved |
| add anti-uniform sharpening regularization (ablation) | SDDLM · LM1B (170M) | Gen PPL: 173.04 -> 116.84 | Improved |
| add anti-uniform sharpening regularization (ablation) | SDDLM · OWT (170M) | sample entropy: 5.53 -> 5.31 | Hurt |
| replace random negatives with the noisy sequence itself (V2 vs V1) (ablation) | SDDLM-V1 · LM1B (170M) | Gen PPL: 116.84 -> 101.32 | Improved |
| replace random negatives with the noisy sequence itself (V2 vs V1) (ablation) | SDDLM-V1 · OWT (170M) | Gen PPL: 45.18 -> 50.05 | Hurt |
| use simplified denoising loss on corrupted positions instead of Duo ELBO loss | SDDLM vs Duo · OWT (170M) | Gen PPL: 80.43 -> 77.07 | No clear effect |
| use simplified denoising loss instead of Duo ELBO loss | SDDLM vs Duo · Zero-shot PPL, Wikitext (trained on OWT) | perplexity: 33.91 -> 35.02 | Hurt |
| use simplified denoising loss instead of Duo ELBO loss | SDDLM vs Duo · Zero-shot PPL, Pubmed (trained on OWT) | perplexity: 45.34 -> 44.75 | Improved |
| use simplified denoising loss instead of Duo ELBO loss | SDDLM vs Duo · LAMBADA (1.1B, zero-shot) | accuracy: 17.02 -> 26.53 | Improved |
| add anti-uniform sharpening regularization (ablation) | SDDLM · LAMBADA (1.1B, zero-shot) | accuracy: 26.53 -> 31.75 | Improved |
| add anti-uniform sharpening regularization (ablation) | SDDLM · BoolQ (1.1B, zero-shot) | accuracy: 60.34 -> 59.66 | No clear effect |
| add unsupervised classifier-free guidance with random-token unconditional branch | SDDLM · LAMBADA (1.1B, zero-shot) | accuracy: 26.53 -> 32.60 | Improved |
Closest papers by meaning
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.91)
- Support Before Frequency in Discrete Diffusion (NeurIPS 2026 · similarity 0.91)
- The Diffusion Duality (ICML 2025 · similarity 0.91)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
- TEncDM: Understanding the Properties of the Diffusion Model in the Space of Language Model Encodings (AAAI 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.
