Read the original at HF Daily Papers
Researchers introduce Pivot-SD, a self-distillation framework that trains masked diffusion language models on high-impact tokens using only 200 questions and four rollouts each.
Carried by: HF Daily Papers. First seen: .