Do Vision--Language Models Understand 3D Scenes or Just Catalogue Objects?

TL;DR

This paper introduces a comprehensive benchmark to evaluate vision-language models' understanding of 3D spatial relationships, revealing significant limitations in their grasp of occlusion and reflections despite good object naming.

Contribution

It presents a novel, human-curated benchmark targeting three aspects of 3D spatial understanding in vision-language models, with detailed analysis of model failures.

Findings

Models excel at volumetric rearrangement planning with 53-97% accuracy.

Models perform poorly on occlusion and reflection tasks, below 45%.

Analysis shows spatial information loss occurs after visual token merging.

Abstract

Vision--language models reliably name objects in a scene, but do they represent the 3D layout those objects inhabit? We introduce a 3,034-sample human-curated benchmark targeting three components of spatial understanding: depth-ordered occlusion (probed via three independent counterfactual operationalisations), optical-geometry inference over visible reflections, and volumetric rearrangement planning. Six frontier and open-weight VLMs, scored by trained annotators on 18,204 responses with no LLM-as-judge, reveal a sharp dissociation: models that plan rearrangements over visible layouts at 53--97% accuracy and rarely violate collision constraints fall to 6--45% on occlusion and below 7% on reflections. An embodied-reasoning model reproduces the same profile. White-box analysis on Qwen3-VL-8B-Thinking localises the failure to the visual-token merger: spatial information recoverable throughout the vision encoder becomes inaccessible after token compression and only stabilises again when clean post-merger activations are patched into the language decoder.