Tag
The paper introduces OPDLM, a method that transforms autoregressive language models into diffusion language models via on-policy distillation, requiring 15x to 7000x fewer training tokens while retaining knowledge from the original model.