/ Paper arXiv 2611.06425
Curator's notes
D3D removes target-model hidden state dependence in diffusion drafting, enabling asynchronous speculative decodingspeculative decodingA technique where a fast draft model proposes tokens that are later verified by a larger target model. with up to 7.48× speedup.
D3D lets a language model write faster by guessing words ahead without needing extra checks each time.D3D removes target-model hidden state dependence in diffusion drafting, enabling asynchronous speculative decodingspeculative decodingA technique where a fast draft model proposes tokens that are later verified by a larger target model. with up to 7.48× speedup.Curator's notesWritten by AI from the paperEvery number here appears in the abstractReport a mistake
What's new
It introduces a standalone diffusion drafterdiffusion drafterA model that generates candidate tokens by iteratively denoising random noise. that drafts without target hidden states and prunes layers to keep cost low.
Key result
Across seven benchmarks D3D averages 5.00× and 5.71× speedup with Qwen3-8B and Qwen3-32B targets, reaching 7.48× on GSM8K with Qwen3-32B.
Why it matters
Builders can generate text much faster without sacrificing quality, saving compute and latency.
Key numbers
From the paperCopied from the abstract, not written by AI.
Abstract
From the paperThe authors' own words.
Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each block is generated in one pass. We introduce D3D, a standalone diffusion drafter that removes this inter-round dependency. D3D combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes without target hidden states. This independence lets D3D draft successor blocks during verification in one packed pass, enabling asynchronous speculative decoding. We progressively prune layers according to their effect on acceptance lengthacceptance lengthThe number of drafted tokens that the target model accepts without correction. to reduce the cost of packed drafting. Removing five of 36 layers retains over 99\% of mean acceptance lengthacceptance lengthThe number of drafted tokens that the target model accepts without correction. across three benchmarks with Qwen3-8B as the target. Across seven benchmarks, D3D averages 5.00times and 5.71times decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03times and 3.67times for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48times. Code is publicly available at https://github.com/fudan-generative-vision/D3D.
From the full paper
From the paperCopied from the paper itself by our code.
Reading the paper on arXiv…
What you could build
Written by AIAdvancedrequires understanding of diffusion models and speculative decoding pipelines
- Implement a real‑time chat assistant using D3D for low‑latency responses.
- Create a batch inference service that drafts multiple sequences in parallel with asynchronous verification.
- Develop a tool that dynamically prunes diffusion layers based on desired speed‑quality trade‑off.
Try it
Code, models and demos that use this paper.
- Code:
- ★ 20 on GitHub ↗
- Datasets:
- fudan-generative-ai/PerfectBlend-Qwen3-4B-Thinking-Greedy ↗8 datasets on Hugging Face
Ask this paper
/ Ask a questionAnswers come from the full paper and name the section
/ Quiz yourself
How does D3D achieve independence from target hidden states during drafting?
By combining blockwise denoising with auxiliary autoregressive supervision, allowing it to draft successor blocks without waiting for updated target-model states.
/ Your notes
Select text in the abstract to highlight it or add a note. They stay in this browser.
