Understanding and manipulating spatial relations is fundamental for AI systems that build grounded world models of physical environments. Existing evaluations of multimodal large language models (MLLMs) focus on static question‑answering over small or synthetic scenes and rarely test whether a model can alter a spatial configuration while respecting explicit geometric constraints. To address this, we introduce GeoSpatialBench, a benchmark that probes an escalating hierarchy of spatial cognition—from relation comprehension and geometric reasoning to constrained spatial editing. The benchmark is built on real‑world urban data from six geographically diverse cities and provides both structured GeoJSON and map‑image modalities. It covers a taxonomy of 28 topological, directional, metric and ordering relations, with ground‑truth labels and constraints verified through GIS operations. Empirical results on state‑of‑the‑art MLLMs reveal a pronounced spatial gap: structured GeoJSON consistently outperforms map images, and multimodal fusion does not yield reliable gains over GeoJSON alone. Models show reasonable performance on qualitative topological and ordering relations but suffer steep drops on quantitative metric relations. Moreover, editing performance on ordering relations is lower than QA, and most failed edits violate at least one specified spatial constraint. These findings highlight the gap between passively recognizing spatial relations and actively manipulating geometric configurations, offering a rigorous testbed for physically grounded world models and embodied spatial AI.
Review