TL;DR of the paper, generated automatically.

Paper introduces D3D, a diffusion drafter that speeds up LLM decoding.

• Change: drafts without target-model hidden states, so it runs asynchronously.
• Speedup: averages 5.00x and 5.71x on Qwen3-8B and 32B.
• Versus DFlash: 4.03x and 3.67x on the same targets.