Panoramic video captures everything happening around a camera, all directions at once. Traditional video only shows what the camera points at. This makes it much harder for AI models to find information scattered across different directions and times.
The researchers created PanoVideoBench, a test with 1000 questions to measure how well models understand panoramic video. They found current models struggle most when needed information appears in different views or times. They then built a training method called PanoVideoReasoner that teaches models to connect observations across directions and time.
The model improved by 12.5 percentage points on the , reaching 55.7% accuracy. This shows that training AI to reason across all directions and moments is possible and effective, which matters for robots and other systems that need complete environmental awareness.