Research map / Image generation / Fast diffusion sampling
Balancing Understanding and Generation in Discrete Diffusion Models
Yue Liu, Yuzhong Zhao, Zheyong Xie, Qixiang Ye, Jianbin Jiao, Yao Hu, Shaosheng Cao, Yunfan Liu
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 |
|---|---|---|---|
| mix uniform noise into masked diffusion (XDLM vs UDLM) on likelihood | XDLM · OpenWebText validation | perplexity: 25.937 -> 24.097 | Improved |
| XDLM vs MDLM on in-domain likelihood | XDLM · OpenWebText validation | perplexity: 23.321 -> 24.097 | Hurt |
| XDLM vs UDLM on zero-shot average perplexity | XDLM · Zero-shot, 7 datasets | perplexity: 59.574 -> 54.110 | Improved |
| XDLM vs MDLM on zero-shot average perplexity | XDLM · Zero-shot, 7 datasets | perplexity: 53.650 -> 54.110 | No clear effect |
| XDLM vs MDLM, few-step image generation without CFG | XDLM · ImageNet-1K, 4 steps, no CFG | FID: 80.752 -> 54.085 | Improved |
| XDLM vs MDLM, 16 steps with CFG | XDLM · ImageNet-1K, 16 steps, CFG 2.0 | FID: 6.725 -> 8.625 | Hurt |
| XDLM vs UDLM, few-step image generation with CFG | XDLM · ImageNet-1K, 4 steps, CFG 2.0 | FID: 14.055 -> 13.550 | Improved |
| XDLM vs UDLM, 16 steps with CFG | XDLM · ImageNet-1K, 16 steps, CFG 2.0 | FID: 8.980 -> 8.625 | Improved |
| continual pretraining of LLaDA with XDLM formulation (600 steps) | LLaDA-8B · MBPP | score: 6.8 -> 15.0 | Improved |
| continual pretraining with standard MDLM objective instead of XDLM (ablation) | LLaDA-8B · OpenCompass benchmarks | score | Hurt |
| use XDLM-trained model instead of only XDLM sampling at inference (ablation) | LLaDA-8B · GSM8k, MATH, BBH, MBPP | score | Improved |
| use mixing ratio k=0.1 between uniform and masked noise (ablation) | XDLM · OWT generation vs zero-shot PPL | Pareto trade-off | Improved |
Closest papers by meaning
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.91)
- SemanticDLM+: Improving Diffusion LLMs through Bias-variance Trade-off in Transition Kernel Design (NeurIPS 2026 · similarity 0.90)
- Simplified and Generalized Masked Diffusion for Discrete Data (NeurIPS 2024 · similarity 0.90)
- LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling (NeurIPS 2026 · similarity 0.90)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · 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.
