Block aim to speed up generation by denoising multiple tokens per block. A practical route is to adapt a pretrained autoregressive (AR) model into this form using distillation. Earlier corrected the mismatch in training states, but left a supervision mismatch: the student sees the whole partially denoised block, while the AR teacher target is computed only from the causal prefix.
d-OPD fixes this by reweighting the AR teacher's next-token distribution using the visible future tokens in the current block. This makes the teacher's supervision match the information the student actually has. On Qwen3 models from 0.6B to 8B, d-OPD raised the six-benchmark average by up to 4.0 points over the OPDLM baseline and reduced training time by 1.35 to 1.58 times.
The result is interesting because it separates two different alignment problems in distillation: state distribution and target conditioning. Even a perfect rollout match is not enough if the teacher and student condition on different context. For people converting AR models to parallel block diffusion, this points to a concrete correction that improves both quality and training cost.