“While Large Language Models (LLMs) primarily predict the next token, vision models are often trained to fill in blanks or reconstruct different parts of an image.”