The paper shows that, when teaching a reasoning AI with step-by-step examples, repeating a small set many times can beat using a huge set only once.
Diffusion Language Models (DLMs) write by polishing whole sentences in several passes instead of one token at a time.