GST-Bench: Can VLMs Develop Global Spatial Awareness from Video?
GST-Bench evaluates whether vision-language models can integrate long video streams into a globally consistent spatial representation. Its human-verified VQA questions are derived from 6,790 minutes of synthetic video and require reasoning from unseen viewpoints and mapping egocentric observations onto top-down images. Across 22 state-of-the-art VLMs, the strongest zero-shot result reported is 42.68, compared with a human score of 79.08. The authors also introduce GST-Bench-Local to distinguish local perception from global integration failures, plus GST-Train as a training resource for global spatial reasoning.
Why it's worth reading
As embodied systems increasingly rely on long-horizon visual memory, the reported 42.68-versus-79.08 gap pinpoints global scene consolidation as a concrete weakness in current VLMs.