Research map / Image generation / Fast diffusion sampling
NI Sampling: Accelerating Discrete Diffusion Sampling by Token Order Optimization
Enshu Liu, Xuefei Ning, Yu Wang, Zinan Lin
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 |
|---|---|---|---|
| use neural indicator instead of confidence threshold | LLaDA-8B-Instruct · HumanEval-512 | tokens/s: 100.8 -> 172.4 | Improved |
| use neural indicator instead of full-step sampling | LLaDA-8B-Instruct · MBPP-512 | speedup: 19.1 -> 273.8 tokens/s (14.3x) | Improved |
| use neural indicator instead of confidence threshold | Dream-7B-Base · MBPP-256 | accuracy: 53.40 -> 56.00 | Improved |
| use neural indicator instead of confidence threshold | Dream-7B-Base · GSM8K-256 | accuracy: 72.78 -> 74.45 | Improved |
| use trajectory-preserving order instead of confidence threshold | LLaDA-8B-Instruct · MBPP-256 | sampling steps: 33.48 -> 10.52 | Improved |
| use final-results-preserving order instead of trajectory-preserving order | LLaDA-8B-Instruct · GSM8K-256 | sampling steps: 22.36 -> 6.95 | Improved |
| add dual caching to NI Sampling | LLaDA-8B-Instruct · GSM8K-512 | tokens/s: 197.7 -> 360.6 | Improved |
| add dual caching to NI Sampling | LLaDA-8B-Instruct · GSM8K-512 | accuracy: 75.44 -> 73.84 | Hurt |
| use NI Sampling instead of Token-Critic at about 15 steps | LLaDA-8B-Instruct · GSM8K | accuracy: 33.59 -> 60.35 | Improved |
| train indicator on GSM8K+MATH mix instead of ShareGPT (ablation) | NI Sampling indicator · GSM8K and MATH vs code | accuracy | Mixed |
| add hidden states and logits as indicator inputs (ablation) | NI Sampling indicator · LLaDA input ablation | accuracy | Improved |
| add additional top-K token embeddings as indicator input (ablation) | NI Sampling indicator · LLaDA input ablation | accuracy | Improved |
Closest papers by meaning
- Scalable Discrete Diffusion Samplers: Combinatorial Optimization and Statistical Physics (ICLR 2025 · similarity 0.91)
- Fast Sampling via Discrete Non-Markov Diffusion Models with Predetermined Transition Time (NeurIPS 2024 · similarity 0.90)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.90)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.90)
- Jump Your Steps: Optimizing Sampling Schedule of Discrete Diffusion Models (ICLR 2025 · similarity 0.90)
- Attention-Based Sampler for Diffusion Language Models (NeurIPS 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.
