I don't think that's accurate. The logits actually have high dimensionality, and they are intermediate outputs used to sample tokens. The latent representations contain contextual information and are also high-dimensional, but they serve a different role--they feed into the logits.