# D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding

**Authors:** Baoyou Chen, Ruiqiao Mei, Guancong Lin, Tao Song, Weihao Yuan, Qun Liu, Siyu Zhu  
**Lab:** Fudan University · **Published:** 3 Nov 2026 · **Upvotes:** 20 · **Level:** Advanced  
**Links:** [arXiv](https://arxiv.org/abs/2611.06425) · [PDF](https://arxiv.org/pdf/2611.06425) · [Code](https://github.com/fudan-generative-vision/D3D) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.06425.html)

## Curator's notes

**TL;DR:** D3D removes target-model hidden state dependence in diffusion drafting, enabling asynchronous speculative decoding with up to 7.48× speedup.

- **What's new:** It introduces a standalone diffusion drafter 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

- **36**: Removing five of 36 layers retains over 99\% of mean acceptance length across three benchmarks with…

## Abstract

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 length to reduce the cost of packed drafting. Removing five of 36 layers retains over 99\% of mean acceptance length 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.

## What you could build

1. Implement a real‑time chat assistant using D3D for low‑latency responses.
2. Create a batch inference service that drafts multiple sequences in parallel with asynchronous verification.
3. Develop a tool that dynamically prunes diffusion layers based on desired speed‑quality trade‑off.

## Quiz yourself

How does D3D achieve independence from target hidden states during drafting?

**Answer:** By combining blockwise denoising with auxiliary autoregressive supervision, allowing it to draft successor blocks without waiting for updated target-model states.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.