Tag
This paper introduces See2Think, a unified evaluation framework to test whether multimodal models genuinely rely on intermediate visual states during reasoning. It finds that visual reasoning is highly model- and environment-dependent, with faithful rendering being the main bottleneck.