There's a good deal of work on latent reasoning models, these days. Here is a recent survey paper.

https://arxiv.org/abs/2507.06203