dOPSD: On-Policy Self-Distillation for Diffusion Language Models

The paper introduces dOPSD, an on-policy self-distillation method for diffusion language models. Instead of giving the teacher an external, instance-specific ground-truth reference, dOPSD derives privileged information from later, more-decoded steps in the student’s own denoising trajectory. This creates token-level, on-policy supervision without requiring tractable sequence likelihoods or sparse sequence-level reinforcement-learning rewards. Experiments on Dream and LLaDA reportedly show gains in in-domain mathematical reasoning and out-of-domain code generation, outperforming supervised fine-tuning and other on-policy baselines.