A latent representation is an internal pattern of numbers that captures features, concepts or relationships learned by a model. It is not ordinary human-readable language, even when it helps the model understand or generate language.
When reasoning happens through latent representations, more of the processing may remain inside the network rather than appearing as a written chain of thought. This can support compact computation but can also make the process harder to inspect.
ELI5
A latent representation is like the private set of shapes and signals an AI uses in its head. People usually cannot read those signals as sentences.
For example, the model might internally group cats with whiskers, fur and animals using numbers, even though it never writes those groupings out in words.
Can people directly read a latent representation?
Usually not. Researchers need specialized analysis tools because the representation is distributed across numerical activations rather than stored as a plain sentence.
Why reason in latent representations?
They may allow richer or more compact internal processing than expressing every intermediate step in natural language.
