Apple researchers have developed a way for vision-language models to learn visual foresight during training without generating intermediate images when answering a question. The method, called Internalized Visual Thinking, predicts hidden representations of future video frames alongside the correct text response.
Visual chain-of-thought systems can reason by creating possible future frames and then processing those images. That makes the reasoning easier to inspect, but adds substantial generation and re-encoding time. Apple's approach instead teaches motion, object transitions and interactions through unlabeled video, then uses the model's normal direct-answer path at inference time.
Across six evaluation settings, the researchers report improvements over text-only post-training and comparable or better performance than explicit visual chain of thought. End-to-end latency fell by more than five times. This is a research result rather than a product release, and the page does not establish how broadly it generalizes. It nevertheless suggests that video systems may gain predictive world knowledge during training without paying the full cost of synthesizing future pixels for every answer.