TL;DR of the paper, generated automatically.

HiFOR is a research paper proposing a way for vision-language-action robot models to use predicted future states when generating actions.

• Hierarchical prediction: The model jointly predicts short-term state changes, semantic milestones and the final goal state, with dependencies across these time horizons set by a Hierarchical Dependency Mask.
• World modeling and retrieval: Future latent representations are organized along horizon, view and spatial dimensions, then retrieved based on evolving action representations during denoising.
• Testing: The authors report experiments on two simulation benchmarks and six real-world manipulation tasks.