A new training method lets AI video models skip the step of literally imagining future frames, and still reason just as well, five times faster.
Researchers built a post-training technique called Internalized Visual Thinking (IVT) for multimodal AI models that reason over video using visual chain-of-thought, a method where the model generates actual intermediate images predicting what happens next before it answers a question. That image generation is slow, which is a problem for models meant to reason about video as it happens. IVT instead trains the model to predict latent representations, essentially compressed internal snapshots, of future frames alongside the text answer, then drops the image-generation step entirely once deployed. Across six evaluation settings, it beat standard fine-tuning every time and matched or beat the image-generating approach, while cutting average end-to-end latency by more than 5x.
Proactive video reasoning, think security cameras, robots, or live-stream assistants that need to anticipate what happens next, has always traded foresight for speed. This result suggests that tradeoff is not fundamental: models can learn to think in pictures during training and skip the visible sketching at answer time. That matters for anyone building video agents where a delay of even a second or two defeats the point.
It's one arXiv preprint, not a shipped product, and the six benchmarks were chosen by the researchers themselves, so treat the 5x figure as a promising lab result, not an industry standard, until someone else reproduces it.