“NVIDIA hypothesizes that a mismatch in SFT data between teacher and student models causes different reasoning behaviors and output distributions, reducing the quality of supervision signals during on-policy distillation.”