D3D: Decoupled Diffusion Drafting for Asynchronous Speculative Decoding

D3D. Decoupled Diffusion Drafting for Asynchronous Speculative Decoding. By Baoyou Chen and six others, from the Shanghai Innovation Institute, the Shanghai Academy of AI for Science, and three other institutions. Published on the fourth of November.

In speculative decoding, a drafter proposes candidate tokens and a target model verifies them. Diffusion drafters conditioned on the target predict their candidates in parallel, which sounds ideal. The catch is a dependency between rounds. Each new block must wait for updated hidden states from the target model, so drafting and verification are forced to run one after the other. Every block is generated in one pass, and the speedup is limited all the same.

D3D removes that inter-round dependency. It is a standalone diffusion drafter, one that does not need the target's hidden states at all. To still achieve long accepted prefixes without them, it combines blockwise denoising with auxiliary autoregressive supervision.

Independence is what unlocks the rest. Because D3D is not waiting on the target, it can draft successor blocks during verification, in one packed pass. Drafting and verification now overlap, which is what makes the speculative decoding asynchronous. Packed drafting carries a cost, so the authors progressively prune layers according to their effect on acceptance length. Removing five of thirty-six layers retains over ninety-nine percent of mean acceptance length across three benchmarks, with Qwen3-8B as the target.

Across seven benchmarks, D3D averages a decoding speedup of five times with a Qwen3-8B target. With a Qwen3-32B target, the average is five point seven one times. The comparison is DFlash, which reaches four point zero three times on the smaller target. On the larger one, DFlash reaches three point six seven times. The best single case is GSM8K with Qwen3-32B, where the speedup reaches seven point four eight times.

Two things in those numbers belong together. D3D's margin over DFlash is wider with the bigger target, where DFlash's own speedup is lower. And the gain is framed not as a faster drafter but as a removed wait: once drafting no longer needs the target's latest state, the two halves of the loop can run side by side.