Why Far Looks Up: Probing Spatial Representation in Vision-Language Models

Cheolhong Min, Jaeyun Jung, Daeun Lee, Hyeonseong Jeon, Yu Su, Jonathan Tremblay, Chan Hee Song, Jaesik Park

arXiv:2605.30161 · 2026-05-30 공개 · arXiv · PDF

vlm spatial-reasoning synthetic-benchmark model-robustness distance-entanglement spatial-axes embedding-analysis perspective-bias

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/.

한국어 요약

📋 한 줄 요약

**[VLM 공간 표현 / Probing]** "Why Far Looks Up"이 VLM 임베딩에서 vertical-distance entanglement 발견 — 모델이 vertical 이미지 위치와 distance 혼동, SpatialTunnel 합성 벤치마크로 model-intrinsic bias 입증.

🎯 핵심 기여도

💡 핵심 아이디어

VLM의 공간 추론 정확도는 진정한 3D 이해가 아닌 자연 이미지에서 학습된 perspective bias(vertical=먼 거리) shortcut에 크게 의존하며, 이는 model-intrinsic이고 data scaling으로 강화되며, well-separated spatial axis를 가진 모델이 더 robust한 추론을 한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: VLM 공간 추론의 본질적 bias 정량 노출, representation-level analysis 프레임워크 정립, SpatialTunnel로 evaluation 표준 제공, 잘 분리된 spatial axis가 robust reasoning과 상관됨 입증으로 모델 설계 가이드. **한계**: Vertical-distance가 한 axis에 국한·다른 spatial axis로 일반화 추가 검증, 합성 벤치마크의 실세계 분포와 격차, bias 완화 솔루션은 후속.

🚀 실용적 활용