When learning patterns during training, it finds non-linear ones. That is the purpose of the activation layers. But when interpolating in the latent space during inference, I believe it would always give a linear combination. I did think about this when I wrote that line and I am moderately confident this is correct.