“NVIDIA's research found that teacher models trained with substantially different pipelines cannot be effectively combined through on-policy distillation, leading to suboptimal performance.”