Tag
The paper introduces BVB, a benchmark for agentic video understanding via programmatic reconstruction in Blender, evaluating models on perceptual similarity and spatiotemporal fact retention.