What's this about?
Latent world models plan by predicting the future in a compressed representation space, and their quality hinges on how well-behaved that space is. QQWorld swaps the usual Gaussian regularizer for quantile-quantile matching, improving the learned latent distribution.
What exactly is a "world model"?
A model that learns to simulate its environment internally — "if I do X, the world becomes Y" — so an agent can plan by imagining outcomes instead of trying everything for real. The "latent" is the compressed sketch of the world it imagines in; if that sketch is distorted, every plan built on it inherits the distortion (11 upvotes).
Why it matters
World models are the backbone of the robotics and video-model wave (see Gemini Robotics ER 2 today); regularization details like this are the unglamorous plumbing that decides whether they work.