Autoregressive language models are built around next-token cross-entropy, yet a growing family of systems also spend computation on hidden states that never surface as words. These include continuous thoughts, pause steps, and recurrent depth. A new arXiv note takes a step back to ask what these latent iterations actually contribute.
The paper organises the question around three ideas: possibility, gauge, and obstruction. Possibility asks whether predictive latents can improve expressiveness or efficiency; gauge concerns the freedom to transform hidden states without changing the model's behaviour; and obstruction points to limits or barriers that prevent latent reasoning from being fully effective.
Because the abstract is brief, the full details of the argument are not yet available. What is clear is that the note aims to bring theoretical clarity to a practice that has been spreading empirically, separating what is genuinely new from what is already implicit in the autoregressive framework.