What's this about?
The paper finds that world-action models retain generalisation when the policy conditions on a single noisy future representation, avoiding repeated video denoising.
What exactly is future modelling?
The agent predicts what the scene may look like after an action, using that imagined consequence to select what to do now.
Why does this matter?
10 HF upvotes. It aims to preserve the benefit of explicit world models at the speed of latent variants, but results are author-reported.