Video models are zero-shot learners and reasoners (arxiv.org)

Researchers behind the Veo 3 video model report that generative video models can exhibit broad zero-shot abilities previously mostly associated with large language models. Without task-specific fine-tuning, Veo 3 can perform object segmentation, edge detection, image editing, infer physical properties and affordances, simulate tool use, and even solve early visual reasoning tasks like mazes and symmetry puzzles. The paper argues these capabilities emerge from the same simple primitives that powered LLMs—large, generative models trained on web-scale data—suggesting video models may be evolving into generalist vision foundation models. This is significant for AI/ML because it shifts how we think about visual intelligence: instead of many task-specific vision systems, a single generative video foundation model could perceive, model, and manipulate the visual world in flexible, compositional ways. Technically, the work highlights emergent zero-shot generalization across perception, simulation, and reasoning domains, indicating that temporal and generative modeling of video provides rich priors for physical and causal inference. If these trends hold, future research could leverage large-scale video pretraining to unify vision tasks, accelerate few-shot/zero-shot deployment, and enable more capable embodied and multimodal agents.