Causal Reasoning in Videos

The field asks whether models can explain why events happen, how entities affect one another, and what would change under a different action—not merely recognize visible content. Common approaches represent events as causal chains or graphs, ground evidence in video frames and timestamps, and use counterfactual interventions to separate causes from correlations. Benchmarks such as CLEVRER and video question-answering datasets test event prediction and explanation, while newer evaluations examine real-world safety, long-range memory, and the physical consistency of generated videos. A central challenge is that strong recognition and fluent explanations often coexist with weak visual grounding, causal attribution, and prediction of unseen outcomes.