Large language models generate text one word at a time, making them slow. uses a lightweight model to propose multiple future words, which the large model then verifies in parallel. But many proposal models depend on the large model's internal states, forcing them to wait and limiting speed.
D3D is an independent proposal model, called a diffusion drafter, that predicts a block of words without needing the large model's hidden states. This lets it prepare the next block while the large model is still verifying the current block. Training combines blockwise denoising and autoregressive supervision; unneeded layers are pruned to keep it fast.
Across seven benchmarks, D3D achieves average decoding speedups of 5.00 times and 5.71 times with two different large models, exceeding prior methods. This matters because it makes large language models much faster without changing their output.