Research map / Image generation / Fast diffusion sampling
Plan for Speed: Dilated Scheduling for Masked Diffusion Language Models
Omer Luxembourg, Haim H. Permuter, Eliya Nachmani
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 |
|---|---|---|---|
| replace self-confidence planner with DUS at 6.4x speedup | LLaDA-Instruct-8B · GSM8K, B=32 | accuracy: 38.74 -> 65.73 | Improved |
| replace self-confidence planner with DUS at 10.7x speedup | LLaDA-Base-8B · GSM8K, B=64 | accuracy: 8.04 -> 35.18 | Improved |
| replace self-confidence planner with DUS | DiffuCoder-Base-7B · HumanEval, B=16 | pass@1: 6.71 -> 38.41 | Improved |
| replace self-confidence planner with DUS at B=8 | Dream-Instruct-7B · BBH, B=8 | accuracy: 55.93 -> 54.81 | Hurt |
| replace self-confidence planner with DUS | LLaDA/Dream models · MMLU-Pro | accuracy: +5-9% | Improved |
| use dilated-incremental schedule instead of fixed-k for self-confidence planner (ablation) | self-confidence planner · GSM8K 300 samples, LLaDA-Base B=16 | accuracy: 55.00 -> 49.33 | Hurt |
| use dilated-incremental schedule for random planner instead of fixed-k (ablation) | random planner · GSM8K 300 samples, LLaDA-Base B=16 | accuracy: 24.67 -> 43.67 | Improved |
| add dilated spacing (DUS) on top of incremental schedule vs. best non-DUS planner (ablation) | DUS · GSM8K 300 samples, LLaDA-Instruct B=32 | accuracy: 53.67 -> 66.67 | Improved |
| add dilated spacing post-filter to EB-Sampler (ablation) | EB-Sampler gamma=2 · HumanEval, LLaDA-Instruct, B=32 | accuracy: 24.4 -> 37.8 | Improved |
| add dilated spacing post-filter to CB-Sampler (ablation) | CB-Sampler tau=0.5 · HumanEval, LLaDA-Instruct, B=32 | accuracy: 22.6 -> 34.8 | Improved |
| add dilated spacing post-filter to adaptive samplers (ablation) | EB/CB-Sampler · Table 4 settings | NFE: EB HumanEval NFE 35 -> 59 | Hurt |
| use very small block size B=8 with DUS (ablation) | DUS · general | accuracy gain | No clear effect |
Closest papers by meaning
- Accelerated Sampling from Masked Diffusion Models via Entropy Bounded Unmasking (NeurIPS 2025 · similarity 0.92)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.92)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.92)
- UnMaskFork: Test-Time Scaling for Masked Diffusion via Deterministic Action Branching (ICML 2026 · similarity 0.92)
- Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling (ICLR 2025 · similarity 0.91)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · 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.
