Perceptual Constancy is a multimodal benchmark designed to evaluate high-level perceptual invariance in large vision-language models (VLMs). It probes a model’s understanding of physical and geometric stability under varying sensory appearances. This dataset is part of the Grow AI Like a Child benchmark initiative.
The Perceptual Constancy dataset focuses on appearance-invariant reasoning using both static images and short… See the full description on the dataset page:
https://huggingface.co/datasets/grow-ai-like-a-child/perceptual-constancy.
