Latent Space

A mathematical space of internal representations learned by a model. Points and directions in this space may capture useful regularities in the data, making it possible to compare content, interpolate attributes, or steer generation without operating directly on raw input. In transformers, the term is also used more broadly for intermediate hidden states.

Its geometry depends on the model, layer, and training objective. Vector proximity suggests a statistical relationship inside that representation, not guaranteed semantic equivalence for humans. A direction observed in one model may also disappear in another layer or version; claims about latent space require probes, interventions, and evals.