input hidden states are sampled from the sequence of states produced by earlier MTP steps 1,...,k − 1, rather than relying exclusively on step k − 1
1 more from these models
we continue applying the hidden-state-mixing strategy described in Section 2.1.