Physical Reasoning in Video

Physical reasoning in video asks whether systems can infer objects’ states, forces, interactions, and causal consequences from changing images, then predict or generate what happens next. Common approaches include video-language models, explicit scene graphs or spatial codes, learned dynamics and differentiable simulators, and retrieval or reasoning modules that test physical laws. Evaluation increasingly uses controlled clips, counterfactuals, temporal and spatial grounding, and measurable criteria such as trajectories, conservation, and event timing, rather than visual quality alone. Current systems often recognize plausible events better than they explain anomalies or sustain accurate causal dynamics, leaving open whether their apparent gains reflect grounded physics or semantic priors and benchmark-specific cues.