We introduce the Latent Program Network (LPN), a new neural architecture that builds test-time search directly inside its inference, by learning and searching through a latent space of implicit programs before generating outputs.