Research map / Image generation / Fast diffusion sampling
Consistent Diffusion Language Models
Hasan Amin, Yuan Gao, Yaser Souri, Subhojit Som, Ming Yin, Rajiv Khanna, Xia 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 |
|---|---|---|---|
| train with multi-path discrete consistency instead of MDLM objective | 110M masked diffusion LM · OpenWebText unconditional, 8 steps | generative perplexity: 682.7 -> 246.9 | Improved |
| tune CDLM for perplexity (shrink max step range, anneal max-step weight) (ablation) | MCDLM · OpenWebText unconditional, 64 steps | generative perplexity: 67.7 -> 40.1 | Improved |
| distill MCDLM with SDTT objective | MCDLM · OpenWebText unconditional, 1024 steps | generative perplexity: 31.2 -> 27.1 vs SDTT on MDLM | Improved |
| use uniform prior CDLM instead of UDLM/DUO | UCDLM · OpenWebText unconditional, 1024 steps | generative perplexity: 84.0 -> 71.7 | Improved |
| remove max-step diffusion anchor (ablation) | CDLM · OpenWebText unconditional | entropy: 5.3 -> 3.2 at 8 steps | Hurt |
| raise max-step weight from 0.4 to 1.0 (ablation) | CDLM · OpenWebText unconditional, 64 steps | generative perplexity: 32.8 -> 75.1 | Hurt |
| use forward KL instead of JS divergence (ablation) | CDLM · OpenWebText unconditional, 8 steps | generative perplexity: 110.6 -> 8e4 | Hurt |
| use backward KL instead of JS divergence (ablation) | CDLM · OpenWebText unconditional, 1024 steps | generative perplexity: 19.7 -> 44.2 with entropy 5.2 -> 4.3 | Hurt |
| use staged increasing step-size scheduler instead of random (ablation) | CDLM · OpenWebText unconditional, 8 steps | generative perplexity: 110.6 -> 160.8 | Hurt |
| use linear increasing step-size scheduler instead of random (ablation) | CDLM · OpenWebText unconditional | generative perplexity: 19.7 -> 17.9 at 1024 steps; 110.6 -> 117.6 at 8 steps | Mixed |
| use CDLM instead of DUO-DCD greedy for conditional generation | MCDLM-PPLOptimized · Conditional generation (OOD sets) | BLEU2: 26.7 -> 62.8 on OpenWebText 8 step | Improved |
Closest papers by meaning
- The Diffusion Duality (ICML 2025 · similarity 0.93)
- SemanticDLM+: Improving Diffusion LLMs through Bias-variance Trade-off in Transition Kernel Design (NeurIPS 2026 · similarity 0.92)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.92)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.92)
- LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling (NeurIPS 2026 · similarity 0.92)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.92)
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.
