What's this about?
HappyWorld-Bench evaluates video worlds, spatial scenes and embodied environments under exploration, interaction and editing instead of judging only a polished generated clip.
What exactly is state consistency?
If an agent moves an object, leaves and returns, the object should still be there. A visually impressive world fails as a simulator when it forgets or silently rewrites prior events.
Why does this matter?
author-reported preprint; 39 HF upvotes.
- The suite contains 1,138 video prompts, 300 spatial scenes and 254 embodied cases.
- Human A/B arenas are paired with behavioural correctness metrics.
- Every tested track shows reliability gaps during longer or altered interactions.