| license: apache-2.0 | |
| task_categories: | |
| - image-text-to-text | |
| tags: | |
| - vlm | |
| - spatial-reasoning | |
| # SpaceNum: Revisiting Spatial Numerical Understanding in VLMs | |
| [Project Page](https://sterzhang.github.io/SpaceNum-Home/) | [Paper](https://huggingface.co/papers/2605.23898) | |
| SpaceNum is a unified framework designed to evaluate Vision-Language Models (VLMs) on their ability to ground numerical outputs in spatial perception. It addresses two complementary settings: | |
| 1. **Dynamic transitions** during spatial exploration. | |
| 2. **Static layouts** in spatial reasoning. | |
| ## Tasks | |
| The benchmark formulates two bidirectional tasks to evaluate the mapping between vision-side spatial structure and language-side numerical representations: | |
| - **Num2Space**: Evaluating how models map numerical representations to spatial structures. | |
| - **Space2Num**: Evaluating how models extract numerical values from visual spatial observations. | |
| SpaceNum systematically studies whether current VLMs truly understand numerical values in spatial settings or rely on shallow spatial cues. |