The model encodes input into a latent space, predicts/processes latent representations, and only then decodes them into text tokens.
Operating solely on discrete tokens hinders longer-horizon planning and abstraction.
