Large Vision-Language Models (LVLMs) have become adept at describing what appears in an image, but they still struggle to reconstruct and reason about the 3D structure behind a 2D scene — a capability the authors call spatial intelligence. Existing methods often rely on real-scene spatial question-answering datasets that demand dense geometric annotations, which are costly, time-consuming, and prone to noise from external perception modules.
In a new paper, researchers propose a paradigm inspired by human cognitive development: learning foundational spatial skills through structured block-manipulation tasks. They introduce SpatialBlock-15k, a synthetic dataset of 15,000 block-stacking problems covering 3D-to-2D projection, viewpoint transformation, and structural combination. The dataset also incorporates controlled color modulation as visual cues to encourage anchor-based reasoning in visually complex conditions.
Experiments show that LVLMs trained on SpatialBlock-15k — through either direct answering or reasoning-based prediction — significantly outperform baselines and generalize to real-world spatial tasks, despite the dataset's synthetic and compact nature. Two training strategies are released: a direct model that predicts the answer immediately, and a reason model that thinks before answering.
The authors note that both models transfer to real-scene spatial benchmarks even though they were trained only on synthetic data at a small scale. Code and data are available at https://github.com/rsoohyun/SpatialBlock.