Understanding and manipulating spatial relations is fundamental for AI systems that reason about physical environments. Yet most evaluations of multimodal large language models (MLLMs) ask static questions about small or synthetic scenes, and rarely test whether a model can change a spatial configuration while respecting explicit geometric constraints.
GeoSpatialBench tests an escalating progression of spatial cognition: from understanding spatial relations, through geometric reasoning, to constrained spatial editing.
- 8,400 instances derived from real OpenStreetMap data
- Six cities with diverse geographic contexts: Stockholm, Paris, Hong Kong, Wuhan, Detroit and Nairobi
- 28 spatial relations recognized in spatial databases and GIS
- Tasks that combine spatial question answering with constrained geometric editing
Findings#
- Models given GeoJSON generally reason about space better than models given map images, pointing to an underexplored challenge in using maps as a visual medium for precise reasoning.
- Metric reasoning remains the most persistent bottleneck.
- Editing performance on ordering relations declines relative to question answering, and most failed edits violate at least one specified spatial constraint.
The paper points to explicit spatial representations — such as spatial tokenization — as one possible direction for bridging spatial recognition, geometric reasoning and spatial intervention.
Cite#
@misc{ren2026mindspatialgap,
title = {Mind the Spatial Gap: A Benchmark for Geospatial Reasoning and Constrained Spatial Editing},
author = {Ren, Zheng and Wang, Yanning and Huang, Ziheng and Huang, Fei and Wang, Shiqi and Ye, Ziqian and Yue, Yang},
year = {2026},
eprint = {2610.12881},
archivePrefix = {arXiv},
primaryClass = {cs.AI}
}