Teaching Diffusion to Speculate Left-to-Right

In three linesDiffusion models can accelerate LLM speculative decoding by generating token blocks in parallel. This paper identifies a gap between bidirectional training and left-to-right verification, proposing three interventions (positional weighting, first-error focal loss, chain loss) that increase accepted draft length by 21-76% across six benchmarks without additional compute.

Summary generated by Claude — human-verified