Diffusion language models (DLMs) win speed advantages through parallel decoding, but training cost has always been a hard engineering hurt. arXiv:2606.06712 (v1, 2026-06-08) cuts in from the "ARLM to DLM engineering transition" angle, proposing a self-distillation + on-policy transition paradigm, pulling DLM pretraining from "retrain" back to the "ARLM post-training" scope.
Core idea
The traditional practice is to directly take an ARLM and change to bidirectional attention, then put a DLM objective on top for pretraining. The authors point out that this triggers two kinds of distribution drift: target drift (switching from next-token prediction to random-mask prediction loses world knowledge) and trajectory drift (training goes through random masks, inference goes through confidence-based decoding, and the two are inconsistent).
OPDLM uses self-On-Policy Distillation to solve this: the student (ARLM with bidirectional attention) generates its own trajectories, and the original frozen ARLM serves as the teacher to provide target logits on the same trajectory. The student is pulled directly to the distribution the inference-time rollout will actually walk.
Results
The paper reports that on multiple tasks, the number of training tokens is reduced 15× to 7000× while still maintaining strong performance. This folds two long-conflicting needs — "wanting DLM inference speed" and "not wanting to pay for DLM pretraining again" — into a standard ARLM post-training flow.
Industry significance
For small labs that already hold ARLM weights, this is equivalent to a low-cost DLM entry ticket, meaning the red-sea competition of DLM inference speed will further compress the ARLM engineering space. But the method is still bounded by the ceiling of the teacher model itself — to have DLM quality truly reverse-surpass ARLM, it still needs innovation in alignment or training paradigms head-on.