Training diffusion language models typically follows one of two paths: on-policy distillation or verifier-based reinforcement learning (RL). Distillation offers dense supervision on the student's own samples, but its targets still come from the teacher. RL, on the other hand, uses an outcome signal for learning, yet provides no direct constraint on the generated samples.

The paper proposes a hybrid approach called outcome-routed distillation and RL. The idea is to use the outcome signal to decide, for each sample, whether to apply distillation or RL. This routing lets the model benefit from both dense teacher guidance and outcome-based feedback, potentially overcoming the limitations of each method alone.

The abstract does not include experimental details, but the conceptual contribution is a flexible training framework that adapts its supervision strategy based on sample quality. The work appears on arXiv and is aimed at improving diffusion language model training.