Search papers, labs, and topics across Lattice.
This paper introduces Internalized Visual Thinking (IVT), a novel framework that allows multimodal large language models to learn visual reasoning during training and apply it directly at inference without the need for generating intermediate visual representations. By optimizing both textual and next-embedding predictions over unlabeled videos, IVT effectively captures essential dynamics such as motion and interactions, leading to improved performance in proactive video reasoning tasks. The results show that IVT not only outperforms traditional fine-tuning methods across various evaluation settings but also reduces inference latency by over five times compared to explicit Visual CoT methods.
Internalized Visual Thinking enables models to reason about future video frames directly, slashing inference time while boosting accuracy.
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 next-embedding prediction over unlabeled videos. Given a partially observed video, IVT predicts latent representations of future frames together with the target textual answer, encouraging the model to capture motion, object transitions, interactions, and latent intent. At inference, IVT generates the answer directly without synthesizing or re-encoding future frames. We conduct controlled studies across target representations, decoder designs, prediction horizons, data mixtures, training curricula, and predictive objectives. IVT improves over direct-answer fine-tuning on all six evaluation settings while retaining the same inference pathway. Compared with explicit Visual CoT, IVT achieves comparable or better performance and reduces average end-to-end latency by more than 5x. Together, our findings suggest that explicit pixel-space generation at inference time, as used in visual chain-of-thought, may not be necessary for effective proactive video reasoning. Predictive world modeling can be internalized during training to produce multimodal reasoners that are both more accurate and substantially more efficient.
