Beyond Visual CoT: Internalized Visual Thinking for Proactive Video Reasoning
Apple Machine Learning Research
Internalized Visual Thinking (IVT) is introduced as a post-training framework that learns visual reasoning from unlabeled videos without generating future frames at inference time. IVT cuts end-to-end latency by more than 5× versus Visual CoT. As a result, it improves over text-only post-training and matches or exceeds Visual CoT performance while keeping the same efficient inference pathway.
Why it matters
Multimodal large language models increasingly use visual chain-of-thought (Visual CoT) to reason about spatial, temporal, and embodied environments. By generating intermediate reasoning images, Visual CoT provides an intuitive mechanism for visual foresight but introduces substantial inference overhead, which is particularly problematic for proactive video reasoning. We ask whether models can learn to think visually during training while reasoning directly at inference. We introduce Internalized Visual Thinking (IVT), a post-training framework that jointly optimizes textual prediction and…