Paper 2605.30161
Why Far Looks Up: Probing Spatial Representation in Vision-Language Models
01 In brief
Summary
The paper investigates whether vision-language models (VLMs) achieve spatial reasoning through structured 3D understanding or by exploiting statistical shortcuts in natural images.
The authors introduce a representation-level analysis framework using contrastive pairs to measure how spatial axes (horizontal, vertical, depth) are organized in VLM embeddings.
They identify a pervasive 'vertical-distance entanglement' where models conflate vertical image position with depth, mirroring perspective bias.
This bias causes a significant accuracy gap between perspective-consistent and counter-heuristic examples, and intensifies with data scaling.
To isolate this bias, they introduce SpatialTunnel, a synthetic benchmark that decouples vertical position from depth.
Experiments across multiple VLM families (Molmo, NVILA, Qwen) show that models with similar benchmark scores can have different internal representations, and that models with well-separated spatial axes (high distance coherence, low VD-Entanglement Index) exhibit greater robustness and accuracy across diverse spatial reasoning benchmarks.
The findings suggest that benchmark accuracy alone may overestimate VLM spatial reasoning capabilities, and that representational structure is a more reliable indicator of robust spatial understanding.
02 From the paper
Abstract
Vision-language models (VLMs) achieve strong performance on spatial reasoning benchmarks, yet it remains unclear whether this reflects structured 3D understanding or reliance on statistical shortcuts in natural images. We introduce a representation-level analysis framework that constructs minimal contrastive pairs to measure how spatial axes are organized and disentangled within VLM embeddings. Our analysis across multiple model families reveals a consistent vertical-distance entanglement: models conflate vertical image position with distance, mirroring the perspective bias of natural photographs. This bias produces a significant accuracy gap between perspective-consistent and counter-heuristic examples, and intensifies under data scaling even as overall benchmark accuracy improves. We further show that models with similar benchmark scores can exhibit different internal representations, and that these differences predict accuracy and robustness across diverse spatial reasoning benchmarks. To isolate this bias from evaluation-set skew, we introduce SpatialTunnel, a synthetic benchmark designed to expose spatial shortcut biases by removing common correlations present in natural images. Experiments confirm that the entanglement is model-intrinsic, and that models with well-separated spatial axes exhibit greater robustness, suggesting that well-structured spatial representations lead to more reliable spatial reasoning across diverse benchmarks. Code and benchmark are available on the project page: https://cheolhong0916.github.io/whyfarlooksup.github.io/.