This event reports the introduction of Internalized Visual Thinking (IVT), a post-training framework that lets multimodal models answer directly about partially observed videos without generating intermediate reasoning images at inference. The arXiv abstract and the Apple research page both state that IVT jointly optimizes textual prediction and next-embedding prediction over unlabeled videos, and that the model is trained to predict latent representations of future frames alongside the target answer.12

Context

The sources frame Visual CoT as useful for spatial, temporal, and embodied reasoning but costly because it synthesizes or re-encodes future frames at inference. IVT transfers this cost to training by learning a latent world model, so the deployment-time path remains the same as a direct-answer model. Both documents describe controlled studies across target representations, decoder designs, prediction horizons, data mixtures, training curricula, and predictive objectives, while giving none of the underlying results.12

Assessment

The genuinely new claim is that predictive world modeling can be internalized during training, making explicit pixel-space generation at inference unnecessary for proactive video reasoning. The constraint that moves is where the computational cost of visual thinking is paid: Visual CoT places it at inference in the form of intermediate image generation, while IVT places it in post-training. This changes the deployment-time trade-off for latency-sensitive video reasoning, provided the reported parity with Visual CoT and the 5× latency reduction are reproducible.12

Caveats

AI hypothesisThe evidence is thin: two summaries, no numerical tables, no dataset names, and no release of code or trained weights. The baseline labels also shift ('direct-answer fine-tuning' in the arXiv abstract, 'text-only post-training' on the Apple page), leaving a minor ambiguity in the comparison. What would change my mind is evidence that next-embedding prediction adds no benefit over text-only training on the same architecture, or that a Visual CoT decoder can reach the same latency with quantization or speculative decoding; absent such data, the main claims are plausible but unverified.12