A new paper on arXiv introduces "Same Space or Not," a benchmark designed to assess spatial intelligence by presenting models with cross-view contradictions. The authors argue that spatial intelligence—the capacity to represent and reason about physical space—is fundamental to embodied perception and real-world visual understanding, yet current evaluation methods capture only part of it.

The approach centers on whether a model can recognize when different views of a scene are inconsistent with a single physical space. By framing the test as a contradiction-detection problem, the benchmark moves beyond existing tasks that probe spatial relations and perspective-taking, offering a distinct way to measure how well a model understands space as a coherent whole.

Because this is a single preprint, the details of the methodology and results are limited to what the abstract reports. Future work or peer review may clarify how the benchmark performs in practice and whether it correlates with other measures of spatial reasoning. For now, the paper marks a useful step toward more comprehensive spatial intelligence evaluation.