Research map / Image generation / Fast diffusion sampling
Breaking the Factorization Barrier in Diffusion Language Models
Ian Li, Zilei Shao, Benjie Wang, Rose Yu, Guy Van den Broeck, Anji 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 |
|---|---|---|---|
| add CoDD to Low Confidence decoding | LLaDA-Instruct-8B · MATH500 (256 steps) | accuracy: 36.00 -> 41.00 | Improved |
| add CoDD to Entropy decoding at 64 steps | Dream-Instruct-7B · GSM8K (full diffusion) | accuracy: 33.97 -> 56.41 | Improved |
| add CoDD to Entropy decoding at 128 steps | Dream-Instruct-7B · GSM8K (full diffusion) | accuracy: 56.18 -> 67.02 | Improved |
| add CoDD to Entropy decoding at 128 steps | Dream-Instruct-7B · GPQA (full diffusion) | accuracy: 14.06 -> 18.08 | Improved |
| add CoDD to Low Confidence decoding | LLaDA-Instruct-8B · MBPP (256 steps) | accuracy: 34.80 -> 34.00 | Hurt |
| add CoDD to Margin decoding | LLaDA-Instruct-8B · GSM8K (64 steps) | accuracy: 40.41 -> 40.41 | No clear effect |
| add CoDD to Entropy decoding (block) | Dream-Instruct-7B · GPQA (block, 64 steps) | accuracy: 5.13 -> 4.91 | Hurt |
| add any-order autoregressive sampling to CoDD (ablation) | CoDD (LLaDA) · MATH500 (64 steps) | accuracy: 15.80 -> 17.0 | Improved |
| add CoDD to Dream | Dream (low-confidence) · WikiText-103 (16 steps) | perplexity: 3.32 -> 2.92 | Improved |
| add CoDD to Dream | Dream (low-confidence) · WikiText-103 (16 steps) | Rep-2: 0.783 -> 0.812 | Hurt |
| add CoDD to Dream backbone | Dream-Instruct-7B · inference latency | seconds per sample overhead: 4-5% | No clear effect |
| raise PC fraction gamma from 0.3 to 0.6 (tau 0.2) (ablation) | CoDD (LLaDA, Random) · GPQA (256 steps) | accuracy: 21.65 -> 26.34 | Improved |
Closest papers by meaning
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
- Reinforcement Learning for Diffusion LLMs with Entropy-Guided Step Selection and Stepwise Advantages (NeurIPS 2026 · similarity 0.90)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.90)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.90)
- Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models (ICLR 2026 · similarity 0.90)
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.
