Paper notes · ICML 2026
DFlash: Block Diffusion for Flash Speculative Decoding
Speculative decoding speeds up an AR model: a cheap drafter guesses several tokens, the target model checks them all in one pass, and the output is exactly what the target would have written alone. DFlash keeps the target AR and changes the drafter: instead of a small AR model that guesses one token per pass, a small block diffusion model guesses a whole block of 16 in one pass.
A dLLM is a weak writer on its own (M06, M12). As a drafter it does not need to be good, only often right, because every guess is checked. That is the paper’s reframing.
Three ingredients:
- Parallel drafting: a masked block of 16 positions, filled in one denoising step.
- The target’s features: hidden states from 5 layers of the target, fused and injected as extra keys and values into every drafter layer.
- Training for the job: blocks that start from a clean token, as at inference, and a loss that weights early positions more.
Loading PDF… open it directly
PDF reproduced unmodified under CC BY 4.0. Highlights and notes are mine.