MindWorldBench is a carefully constructed dataset designed to evaluate the physical and social cognitive reasoning of vision-language models and video generation models. It structurally maps cognitive tasks using the BDI (Belief, Desire, Perception) theory of mind framework.
images/: The core visual context for each scenario.
prompt.json: The hierarchical, full metadata containing all tasks strictly categorized by… See the full description on the dataset page:
https://huggingface.co/datasets/Richard24601/MindWorldBench.
