Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data 1