left-to-right-verification

This paper proposes three training-time interventions (positional weighting, first-error focal loss, and chain loss) to align diffusion-based draft models with autoregressive verification in speculative decoding, improving accepted prefix length by 21–76% without extra inference cost.