Research map / Image generation / Fast diffusion sampling
Trajectory-Level Speculative Decoding for Diffusion Language Models
Tianxiang Pan, Baitao Gong, Mo Guang, Hongwei Yong, Tianpeng Jiang, Yaqian Li, Zheng Cao, Kaiwen Long
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 trajectory-level speculative decoding | Fast-dLLM (DC) · GSM8K (LLaDA) | denoising steps: 110.3 -> 71.3 | Improved |
| add trajectory-level speculative decoding | Fast-dLLM (DC) · HumanEval (LLaDA) | denoising steps: 179.2 -> 106.7 | Improved |
| add trajectory-level speculative decoding | Fast-dLLM (DC) · HumanEval (LLaDA) | TPS: 70.6 -> 94.4 | Improved |
| add trajectory-level speculative decoding | Fast-dLLM (DC) · GSM8K (Dream) | accuracy: 74.6 -> 73.7 | Hurt |
| add trajectory-level speculative decoding | Fast-dLLM (DC) · HumanEval (Dream) | accuracy: 51.2 -> 54.3 | Improved |
| add trajectory-level speculative decoding | Fast-dLLM (DC) · MATH level 4 (LLaDA) | accuracy: 30.4 -> 25.1 | Hurt |
| add trajectory-level speculative decoding | Fast-dLLM (DC) · IFEval (LLaDA, H800) | TPS: 36.8 -> 69.1 | Improved |
| add intra-block tree W1D1(1) (ablation) | Fast-dLLM (DC) · HumanEval (LLaDA) | tokens per step: 2.61 -> 3.51 | Improved |
| grow intra-block tree from W2D2(3) to W3D3(6) (ablation) | trajectory speculation · HumanEval (LLaDA) | latency (s): 5.3 -> 5.5 | Hurt |
| add inter-block speculation to W2D2(3) (ablation) | trajectory speculation · HumanEval (LLaDA) | denoising steps: 124.0 -> 106.7 | Improved |
| add inter-block speculation to W2D2(3) (ablation) | trajectory speculation · HumanEval (LLaDA) | latency (s): 5.3 -> 5.0 | Improved |
| use full tree W3D3(9) instead of hybrid W3D3(6) (ablation) | trajectory speculation · HumanEval (LLaDA) | total latency (s): 5.5 -> 6.6 | Hurt |
Closest papers by meaning
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.94)
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.91)
- DFlash: Block Diffusion for Flash Speculative Decoding (ICML 2026 · similarity 0.91)
- Accelerating Diffusion Language Models via Structured Suffix Modeling (NeurIPS 2026 · similarity 0.91)
- A Theoretical Perspective for Speculative Decoding Algorithm (NeurIPS 2024 · similarity 0.91)
- Mean-Field Parallel Decoding for Discrete 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.
