Latent variables are underlying factors that influence observed data but aren't directly measured or recorded themselves, inferred instead from patterns in what actually can be observed. In a dataset of customer purchases, for instance, something like "underlying interest in a product category" might be a latent variable: it isn't recorded anywhere directly, but it shapes the purchase patterns a model can actually see.
Latent variable models are useful specifically because they let a system represent hidden structure in data more compactly than the raw observations would allow, which is a big part of why they show up throughout generative AI: a diffusion model's latent space or a language model's internal representations both function as compressed, latent encodings of complex information that make it computationally tractable to generate new, coherent outputs rather than working directly with raw, high-dimensional data at every step.
