Research map / Image generation / Fast diffusion sampling
WavefrontDiffusion: Dynamic Decoding Schedule for Improved Reasoning
Haojin Yang, Rui Hu, Zequn Sun, Rui Zhou, Yujun Cai, Yiwei Wang
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 fixed block schedule with dynamic wavefront | BlockDiffusion · GSM8K (LLaDA-8B-Instruct) | accuracy: 80.74 -> 82.03 | Improved |
| replace fixed block schedule with dynamic wavefront | BlockDiffusion · HumanEval (LLaDA-8B-Instruct) | accuracy: 45.73 -> 47.56 | Improved |
| replace fixed block schedule with dynamic wavefront | BlockDiffusion · MBPP (LLaDA-1.5) | accuracy: 44.04 -> 46.20 | Improved |
| replace fixed block schedule with dynamic wavefront | BlockDiffusion · MATH (LLaDA-8B-Instruct) | accuracy: 40.62 -> 41.04 | Improved |
| replace fixed block schedule with dynamic wavefront | BlockDiffusion · WikiText | BERTScore F1: 0.7946 -> 0.8094 | Improved |
| increase wavefront size F from 4 to 8 (ablation) | WavefrontDiffusion · HumanEval | accuracy: 45.12 -> 47.56 | Improved |
| increase wavefront size F from 8 to 16 (ablation) | WavefrontDiffusion · HumanEval | accuracy: 47.56 -> 45.12 | Hurt |
| increase expansion radius R from 2 to 8 (ablation) | WavefrontDiffusion · HumanEval | accuracy: 47.56 -> 42.07 | Hurt |
| increase expansion radius R from 2 to 8 (ablation) | WavefrontDiffusion · MATH | accuracy: 41.04 -> 41.00 | No clear effect |
| use wavefront instead of block at 256 steps | BlockDiffusion · GSM8K (256 steps) | accuracy: 43.14 -> 42.38 | Hurt |
| use wavefront instead of block at 512 steps | BlockDiffusion · GSM8K (512 steps) | accuracy: 67.02 -> 73.79 | Improved |
| wavefront instead of block with entropy priority | BlockDiffusion · GSM8K (Dream-7B, entropy) | accuracy: 81.96 -> 82.79 | Improved |
Closest papers by meaning
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.89)
- Dynamically Structured Diffusion Language Model Decoding via Bayesian Inference (NeurIPS 2026 · similarity 0.89)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.89)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.88)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.88)
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.
