Research map / Image generation / Fast diffusion sampling
Is Your Diffusion Sampler Actually Correct? A Sampler-Centric Evaluation of Discrete Diffusion Language Models
Luhan Tang, Longxuan Yu, Shaorong Zhang, Greg Ver Steeg
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 |
|---|---|---|---|
| reduce number of sampling steps below sequence length (ablation) | SEDD, MDLM, LLaDA, ReMDM with oracle denoiser · OWT oracle | transition KL, NLL, entropy | Hurt |
| increase SEDD score sharpening beta above 1 (ablation) | SEDD oracle sampler · OWT oracle | transition KL and 3-gram diversity | Hurt |
| increase SEDD score sharpening beta above 1 (ablation) | SEDD oracle sampler · OWT oracle | GenPPL and sequence NLL | Improved |
| increase SEDD score sharpening (lower temperature) (ablation) | SEDD oracle sampler · OWT oracle | MAUVE | Hurt |
| increase local sharpening beta from 1 to 25 (ablation) | AR bigram generator · OWT | GenPPL: 688.4 -> 1.27 | Improved |
| use ReMDM-conf instead of ReMDM-loop (ablation) | ReMDM oracle · OWT oracle | transition-level error | Improved |
| add nucleus sampling (p=0.9) (ablation) | ReMDM oracle · OWT oracle | transition KL, NLL, entropy | Improved |
| add nucleus sampling (p=0.9) (ablation) | ReMDM oracle · OWT oracle | MAUVE | No clear effect |
| vary ReMDM variant or step count (ablation) | ReMDM oracle · OWT oracle | MAUVE | No clear effect |
| increase diffusion steps for LLaDA with oracle posterior (ablation) | LLaDA oracle · OWT oracle | transition KL, entropy, n-gram diversity | No clear effect |
Closest papers by meaning
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.90)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.90)
- Think while You Generate: Discrete Diffusion with Planned Denoising (ICLR 2025 · similarity 0.90)
- Informed Correctors for Discrete Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.89)
- Fast Sampling via Discrete Non-Markov Diffusion Models with Predetermined Transition Time (NeurIPS 2024 · similarity 0.89)
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.
