GST-Bench: Can VLMs Develop Global Spatial Awareness from Video?

2026-08-06

AI Summary

This work addresses the limited global spatial awareness of current vision-language models in long-duration videos, which typically support only local spatial understanding. To bridge this gap, the authors introduce GST-Bench, the first evaluation benchmark for video-based global spatial intelligence, comprising 6,790 minutes of synthetically generated, human-verified question-answer pairs that require models to perform spatial reasoning from novel, unseen viewpoints and map egocentric observations onto a consistent global top-down layout. Accompanying datasets—GST-Bench-Local and GST-Train—are also released. The benchmark evaluates 22 state-of-the-art models within a VQA framework, revealing a significant performance gap: the best zero-shot model achieves only 42.68% accuracy, far below the human baseline of 79.08%, underscoring a fundamental deficiency in existing approaches to integrating long-horizon observations into coherent global scene representations.

Abstract

Spatial intelligence is fundamental to embodied agents, yet existing benchmarks focus on local spatial perception from single or few viewpoints, overlooking global spatial awareness over continuous, long-horizon visual streams. To address this limitation, we introduce the Global-Spatial-Temporal Benchmark (GST-Bench), a VQA benchmark for global spatial intelligence in video understanding, comprising human-verified questions derived from 6,790 minutes of synthetically generated video. It requires models to perform accurate spatial inference from novel viewpoints unseen in the input video and to map egocentric observations onto global top-down images. A comprehensive evaluation of 22 state-of-the-art VLMs exposes a striking gap between models and humans: the strongest zero-shot model attains only 42.68, far below the human score of 79.08. To probe the cause of this gap, we construct GST-Bench-Local and find that models, despite strong local spatial understanding under the same task formulation, still fail to consolidate long-horizon observations into a globally consistent scene representation. We further provide GST-Train, a dataset for global spatial reasoning, as a complementary resource to facilitate future research on this challenge.

Problem

• global spatial awareness
• video understanding
• visual question answering
• spatial intelligence
• embodied agents

Innovation

• Global Spatial Awareness
• Video Understanding
• Visual Language Models
• Spatial Reasoning
• VQA Benchmark