Spatial Reasoning Evaluation of 21 Vision-Language Models Shows Max 58.0% Accuracy

On November 13, 2026, Leiphone published a report stating that a team led by Zhuang Yueting at Zhejiang University has proposed a new benchmark for spatial intelligence, designed to evaluate how well vision-language models (VLM) perform predictive spatial reasoning. The original report notes that spatial intelligence is emerging as a key arena in the AI race, and that vision-language models (Vision-Language Model, VLM) are already capable of answering a fair number of spatial questions.

Using this benchmark, the team evaluated the predictive spatial reasoning performance of 21 mainstream vision-language models. The results show that current general-purpose models still face a clear gap in spatial understanding, with the highest accuracy among the 21 models reaching just 58.0%. The report's headline also raises the question of whether dedicated spatial models actually fall short of general-purpose models.

At this point, the publicly available information stops at these evaluation results. No further details about the benchmark itself, nor any follow-up developments, have been disclosed yet, leaving the current status of the work at the stage of the released test outcomes and the team's initial findings.