Researchers have uncovered a surprising limitation in how advanced vision-language models (VLMs) understand images. While these AI systems are becoming better at linking words and pictures, new findings show they often fail to accurately represent spatial relationships—how objects are arranged in space—even when they seem to perform well on tests. This discovery is important because understanding spatial relations is crucial for tasks like describing scenes, answering questions about images, or navigating environments.
Key Takeaways
- Vision-language models represent spatial relationships less reliably than they do simple word information.
- A new metric called EICS (Evaluation-Invariant Concept Stability) was developed to measure how consistently models encode concepts under controlled changes.
- The CASP-VL benchmark tests models using carefully designed spatial perturbations based on real-world scene graphs.
- Spatial relation information is present in the models’ token-level outputs but gets lost when these outputs are combined into a single overall representation.
The research, published in a new paper by Shubh Garg and Chen Shani, addresses a key question in AI: do vision-language models truly understand the relationships between objects in a scene, or are they just guessing based on word patterns? To investigate, the authors introduced two new tools. The first, EICS, is a way to check if a model’s internal representations of concepts remain stable when one part of the input is changed deliberately. This helps reveal whether the model really “knows” a concept or is relying on shortcuts.
The second tool, CASP-VL, is a benchmark that tests models with spatial perturbations—small, controlled changes in how objects are arranged in images. These perturbations are drawn from Visual Genome scene graphs, which map out objects and their relationships in real images. By tweaking spatial relations while keeping other factors constant, the researchers could see how well models maintained an accurate understanding of space.
Using these tools, the team evaluated several popular vision-language models. They found that while models could recognize words and object labels reliably, they struggled to encode spatial relations like “above,” “next to,” or “behind” in a stable way. Interestingly, the spatial information was detectable at the level of individual tokens (the small pieces of language the model processes), but when these tokens were combined into a single summary representation—often used for downstream tasks—the spatial details disappeared. This suggests a bottleneck in how models aggregate information.
In simpler terms, the models “know” spatial relations when looking at specific words or parts, but lose this knowledge when forming an overall picture. This problem is not obvious if you only look at the model’s final answers or outputs, highlighting why traditional evaluation methods might miss important weaknesses.
These insights could have significant implications for the development of AI systems that need to reason about the physical world. For example, robots or virtual assistants that rely on vision-language models might misunderstand instructions involving object placement or navigation. The research points to the need for new architectures or training methods that better preserve relational information throughout the model’s processing pipeline.
Looking ahead, the authors suggest that further work could explore how to maintain spatial relations in global representations and whether incorporating causal interventions, like those used in their benchmark, can improve model robustness. This study serves as a reminder that even state-of-the-art AI can have hidden blind spots, and careful evaluation is key to building truly reliable systems.
Based on research published on arXiv by Shubh Garg, Chen Shani.