Hugging Face researchers have released GST-Bench, a new benchmark designed to test whether vision-language models (VLMs) can develop global spatial awareness from continuous video streams. Unlike existing benchmarks that focus on local spatial perception from single or few viewpoints, GST-Bench challenges models to reason about space over long-horizon visual input.
The benchmark comprises 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 and to map egocentric observations onto global top-down images.
In a comprehensive evaluation of 22 state-of-the-art VLMs, the strongest zero-shot model achieved only 42.68, far below the human score of 79.08. This striking gap suggests that current models struggle to consolidate long-horizon observations into a globally consistent scene representation.
To probe the cause, the researchers constructed GST-Bench-Local, which shows that models have strong local spatial understanding under the same task formulation but still fail at global integration. They also provide GST-Train, a dataset for global spatial reasoning, as a complementary resource to facilitate future research.