Internalized Visual Thinking Reduces Video Reasoning Inference Overhead

Internalized Visual Thinking (IVT) is a post-training framework that optimizes textual and next-embedding prediction for video reasoning. It allows models to reason about future frames internally without the high inference cost of generating intermediate Visual CoT images.