The challenge of streaming model distillation lies in the fact that if the teacher model adopts a non-causal architecture, its latent representations contain future information that the student model cannot access during inference, requiring the design of alignment strategies to prevent the student from being forced to fit information it cannot reproduce.