Multimodal Language Models Cannot Spot Spatial Inconsistencies

Conventional benchmarks ask models to describe a scene

Most benchmarks ask multimodal LLMs to describe a scene’s spatial attributes, such as whether a staircase is left or right of a window. Many SOTA models score very well, which can create the appearance of grounded spatial understanding.

We hypothesize that semantic priors and local visual cues can support plausible descriptions without a spatially coherent scene representation. We therefore propose a counterfactual test to probe this understanding.

We test spatial understanding with a simple counterfactual test

We introduce an inconsistency into a multiview scene and ask the model to identify which object violates its 3D structure. We hypothesize this requires it to reason about what is and isn’t spatially coherent.

We propose a simple cut-and-paste method to generate spatial inconsistencies. We first select a relatively large object in V₁. Then, we cut it out and inpaint the area in V₂. Finally, in that region, we paste in the same object from V₃. Since V₃’s pose is different from V₂, the change in object pose doesn’t match the camera motion and results in a spatial inconsistency.

All models do substantially worse than humans, and by a large margin too. This suggests that even if these models can describe spatial attributes of scenes, they lack grounded spatial understanding.

We empirically find that artifacts from cut-and-paste do not confound our results.

Notably, models select a cut-and-pasted object with a different pose (benchmark) far more than if we cut-and-paste the same pose. This suggests models look for 3D structure and not artifacts.

Increasing the reasoning budget does not improve spatial reasoning. This is counterintuitive and suggests that models’ reasoning abilities do not represent spatial information well.