Apple IVT Cuts Visual Chain-of-Thought Latency by More Than 5×

Apple researchers test a faster alternative to explicit visual chain-of-thought

Apple researchers published Beyond Visual CoT: Internalized Visual Thinking for Proactive Video Reasoning on Apple Machine Learning Research on August 24, 2026. The work introduces Internalized Visual Thinking (IVT), a post-training framework designed to give multimodal models useful visual foresight without forcing them to generate intermediate images during inference.

The underlying paper was submitted to arXiv on August 16. This is therefore a research release, not a newly launched Apple consumer product or generally available model API.

Visual chain-of-thought can add major inference overhead

Some multimodal reasoning approaches generate an intermediate future image or visual state, then feed that generated result back into the model before producing an answer. That can make the reasoning process more explicit, but it also adds the cost of generating and re-encoding visual content.

IVT moves that predictive work into training. Given a partially observed video, the method trains the model to predict latent representations of future frames while also learning the target textual response. At inference time, the predictive branch is removed and the model answers directly.

The central hypothesis is that a model may benefit from learning to predict future visual states without needing to materialize those states as pixels every time it reasons.

The method jointly learns text and future visual embeddings

During post-training, IVT combines the standard text-generation objective with an auxiliary objective for predicting future visual embeddings. The researchers study several choices for those target representations, decoder designs, prediction horizons, data mixtures and training curricula.

Apple's summary says the method improves over text-only post-training across all six evaluation settings while preserving the same direct inference path. The paper also reports that reconstruction-oriented visual representations can be particularly useful for learning fine-grained spatial and appearance information needed for future-state reasoning.

These results come from the authors' controlled research experiments and should not be treated as universal performance guarantees across all multimodal models or video tasks.

Apple reports more than a 5× latency reduction versus explicit Visual CoT

The researchers report that IVT achieves comparable or better performance than explicit Visual CoT while reducing end-to-end inference latency by more than 5× in their evaluated setup.

The efficiency gain comes from eliminating inference-time image synthesis and re-encoding. Once training is complete, the model uses the representations learned through the predictive objective but produces the final answer directly.

That design is especially relevant to applications where latency matters, including video assistants, embodied systems and other multimodal agents that must reason about events as they unfold.

Why proactive video reasoning is different from normal video understanding

The research focuses on proactive video reasoning: interpreting a partial sequence and reasoning about what is happening or what is likely to happen next. That requires more than describing visible frames. A useful system must infer motion, object transitions, interactions and latent intent from incomplete evidence.

The researchers evaluate the approach across video reasoning settings including early-event and next-event prediction. Their findings suggest that richer predictive supervision during training can improve those capabilities without requiring a slower visual-generation loop at runtime.

What this research does and does not establish

Released: Apple has published the IVT research and the associated paper describing the method and evaluation.

Not announced: Apple has not announced a new commercial model, API or on-device product using IVT, and the research does not establish that every multimodal system will see the same speed or accuracy gains.

The broader significance is architectural. As multimodal models move toward real-time video, robotics and embodied interaction, developers need ways to preserve predictive reasoning while reducing the inference cost of explicit intermediate generations. IVT presents one evidence-backed route: train models to internalize visual prediction, then keep the deployment-time reasoning path lightweight.