Think3D: Thinking with Space for Spatial Reasoning
Vision-Language Models are limited by 2D vision for spatial reasoning. Think3D introduces a framework that enables these models to reason interactively in 3D space by using a suite of 3D manipulation tools and an iterative 'observe-manipulate-reflect' loop. This active exploration approach improves performance on benchmarks and allows smaller models to learn effective 3D navigation through reinforcement learning.