By combining visual reasoning andcode execution, the model formulates plans to zoom in, inspect, and manipulate images step-by-step. Until now, multimodal models typically processed the world in a ...