See2Think: Do Multimodal Models Really Use Intermediate Visual States? Paper • 2607.26769 • Published Jul 29 • 25