Recently, the Princeton University AI Lab team officially launched a new spatial intelligence benchmark — SSoN. The test aims to strip away the interference of dynamic interaction and action control, specifically assessing AI's 3D spatial understanding under static cross-view conditions. The results show that on indoor spatial questions requiring participants to "pick the odd one out among four," humans achieved an accuracy as high as 83.3%, while the top-performing mainstream vision-language models reached only 39.7%, and even most models performed below the random baseline for a four-choice task, exposing a core weakness in current AI's spatial geometric reasoning.
Core Test Mechanism: Four Static Images Test Spatial Consistency
The core task logic of the SSoN benchmark is straightforward: by providing a tested model with four static indoor scene images, it requires the model to accurately identify the "odd image" that does not belong to the same scene.
In this way, the test can directly examine whether AI possesses true cross-view 3D spatial integration ability. However, the test results reveal a huge gap: humans can easily identify spatial logical inconsistencies through intuition and life experience, while current vision-language models struggle when faced with such tests.
Three Core Pain Points: Clear Weaknesses in Geometric Reasoning, Questionable Explanation Validity
After in-depth analysis of model performance, the Princeton research team found that current large models expose three notable pain points in spatial intelligence:
Movement and Rotation Reasoning Is the Biggest Weakness: Across various geometric reasoning question types, especially scenarios involving spatial transformations such as relative object movement and viewpoint rotation, models' error rates remain high, making this weakness particularly prominent.
Knowing the Answer but Not Why: Among the few cases where models chose the correct answer, nearly 30% of their reasoning explanations were logically invalid, a classic case of "guessing the right answer but being unable to explain why."
Overall Performance Worse Than Random Guessing: Most tested models' overall accuracy was even below the random baseline of a four-choice task, indicating that they have not yet truly established a stable 3D spatial intuition.
Follow-up Intervention Experiments: Targeted Fine-Tuning and Camera Parameters Are Key to Breaking Through
To explore how to help AI bridge this technological gap, the research team then conducted a series of intervention experiments:
Prompt Guidance and 3D Reconstruction: Experiments showed that text prompt guidance or introducing a traditional 3D reconstruction pipeline only brought small improvements to model performance.
Providing Real Camera Parameters: When models were directly given real camera parameters, their performance improved substantially; even so, they still could not reliably answer more than half of the questions correctly.
Targeted Supervised Fine-Tuning: Encouragingly, targeted supervised fine-tuning for this task enabled models to nearly double their accuracy on test questions featuring entirely new, unseen rooms.
This result fully demonstrates that large models' ability to judge spatial consistency is not insurmountable and can be transferred and improved through targeted supervised training. The launch of the SSoN benchmark successfully strips away extraneous confounding factors, points directly at the core gap in current AI's cross-view 3D spatial integration, and provides a clear direction for researchers worldwide to iterate on spatial intelligence technology going forward.
