Show, Don't Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text

Researchers introduced ProVisE, a benchmark-agnostic framework that evaluates image-generation models on spatial tasks using pixel-space answers.

They also released SpatialGen-Bench, a 470-sample benchmark covering 14 spatial subtasks.

Results show image-generation models can perform competitively when they respond visually, while text-based VLMs still lead on more compositional spatial reasoning.

The work offers a fairer way to measure spatial cognition and clarifies the strengths of visual vs. text-based systems.