Do maps still matter for machines: revisiting the role of choropleth maps in foundation model spatial understanding

Spatial understanding is crucial for foundation models (FMs), and visual maps have long been a key tool for humans to understand space. Recent advances in FMs have dramatically improved their ability to process and reason over text, images, and other multimodal data, leading to a growing debate over whether traditional human-designed abstractions, such as maps, are still necessary for machine reasoning. This study revisits the role of choropleth maps in the spatial understanding of FMs, asking whether maps, as a long-standing cognitive medium for spatial understanding, still offer benefits in the era of multimodal models. To this end, we construct ChoroplethMap-Bench, a controlled benchmark consisting of 2,400 synthetic choropleth maps, corresponding GeoJSON data, and 12,000 questions, spanning five hierarchical cognitive dimensions: Identify, Spatial Recognition, Compare, Rank, and Delineate. The benchmark compares three input conditions—Data Only, Map Only, and Data + Map—across 22 frontier models, including both open-source and proprietary systems. The results demonstrate that maps, particularly when combined with symbolic data, substantially enhance model performance, especially for tasks requiring higher-level spatial understanding. We further analyze how cartographic design factors, including map type, color hue, and spatial structure, influence reasoning performance. In addition, we examine language effects and response stability, showing the cross-lingual robustness and repeatability of modern multimodal models. Overall, our findings show that incorporating maps into FMs leads to improved geospatial reasoning accuracy, with the strongest gains observed under the Data + Map condition. This work highlights the continued relevance of maps as external representations for machine intelligence and suggests future directions for evaluating broader map forms and more complex reasoning tasks.