Latent reasoning refers to a model carrying out intermediate reasoning steps within its internal hidden states — its latent space — rather than by generating explicit intermediate text tokens that a person could read, as in a written chain of thought.
Why this distinction matters
Explicit chain-of-thought reasoning is legible: every intermediate step is a token a person can inspect. It's also comparatively expensive, since each reasoning step costs a full forward pass to generate and consumes context-window space. Latent reasoning instead lets a model iterate on its internal representation — for instance, by running a recurrent block multiple times before producing output — without materializing each intermediate step as text. This can be considerably cheaper per reasoning step, at the cost of interpretability: there's no intermediate text to inspect if the reasoning goes wrong.
Relationship to recurrent-depth architectures
A recurrent-depth model that applies a shared block repeatedly before producing output is a direct implementation of latent reasoning: each additional pass through the block is an additional reasoning step taken entirely in hidden-state space, with depth exposed as a controllable runtime parameter rather than fixed at training time.