AIAny
Icon for item

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

Evaluates VLMs' ability to form global spatial awareness from long-horizon egocentric video. Introduces GST-Bench: a VQA benchmark with human-verified questions from 6,790 minutes of synthetic video, reveals a large gap (best zero-shot 42.68 vs human 79.08) and provides GST-Train dataset.

Introduction

Why this matters

Global spatial awareness—being able to integrate dispersed egocentric observations into a consistent top-down scene representation—is essential for embodied agents and robotics. Most prior benchmarks measure local perception from single or few views; GST-Bench exposes that current VLMs struggle when required to consolidate long-horizon video into a global map and answer spatial queries from novel viewpoints.

Key Findings
  • Large human–model gap: a comprehensive evaluation over 22 state-of-the-art VLMs shows the best zero-shot score at 42.68 while human performance is 79.08, indicating substantial room for improvement in global reasoning.
  • Local competence, global failure: experiments with GST-Bench-Local demonstrate that models can succeed on short-range, same-view spatial tasks but fail to merge observations across time into a globally consistent layout — so good local perception does not imply global understanding.
  • Diagnostic benchmark design: GST-Bench uses human-verified VQA derived from 6,790 minutes of synthetically generated egocentric video and includes tasks that require inferring unseen viewpoints and mapping observations to top-down images — so it isolates long-horizon integration and cross-view generalization.
  • Training resource provided: GST-Train accompanies the benchmark to help researchers probe learning-based remedies and fine-tuning strategies for global spatial reasoning.
Who It's For and Trade-offs

Great fit if you are researching multimodal long-horizon reasoning, embodied perception, or evaluating VLMs for robotics and navigation. The benchmark is especially useful for method development that targets cross-view alignment, temporal memory, and scene reconstruction. Look elsewhere if your interest is purely single-image VQA or short-clip action recognition — GST-Bench emphasizes global consolidation over instantaneous perception. Note that the dataset is synthetically generated, which helps control variables and scale but may introduce a domain gap to real-world egocentric footage.

Where It Fits

GST-Bench complements local spatial and video reasoning benchmarks by focusing explicitly on global top-down mapping and cross-view inference from continuous egocentric streams. Use it when you need a targeted assessment of a model's ability to integrate dispersed visual evidence across long horizons.

Methodology (brief)

The paper constructs human-verified question-answer pairs from long synthetic egocentric trajectories, evaluates a broad suite of VLMs under zero-shot and controlled settings, introduces a local variant to separate local vs global capabilities, and releases GST-Train to enable supervised or self-supervised approaches aimed at closing the identified gap.

Information

  • Websitearxiv.org
  • AuthorsQifeng Zhang, Kaixiang Huang, Heng Dong, Huang Fang, Junting Chen, Junjie Zhu, Yonghang Chen, Zhiyu Zhang, Wei Li
  • Published date2026/08/06

More Items

Generates retrieval-centric Chain-of-Thought (RC-CoT) over initially retrieved candidates to improve unified multimodal retrieval via reranking or full-corpus re-retrieval with a dual-mode embedder. Trains an embedder–adviser framework (UniME-R1) using mined hard negatives, supervised learning, and retrieval-oriented reinforcement learning.

Evaluates vision-language model judges on computer-using agent (CUA) trajectories to measure verifier reliability. Provides OSReward-Hard and OSReward-Multi challenge sets, the OS-Shepherd-100K reasoning-annotated corpus, and trained OS-Shepherd reward models that match commercial judges at ~30–60× lower cost.

Generates large-scale, explorable 3D open-world scenes from open-ended text prompts, producing editable instance-level assets and a consistent global terrain. Uses agentic planning to convert text into region/terrain/asset specifications and a coarse-to-fine pipeline for terrain construction, mesh reconstruction, and render-based refinement.