Being-H0.7: A Latent World-Action Model from Egocentric Videos

A team of researchers has introduced Being-H0.7, a new latent world-action model that addresses a key limitation of Vision-Language-Action models (VLAs). Standard VLAs map observations directly to actions but often rely on shortcuts rather than understanding task dynamics, contacts, and progress. Meanwhile, world models improve future reasoning but typically require expensive pixel-space video generation that adds overhead and distracts from action-relevant features.

Being-H0.7 solves this by inserting learnable latent queries between perception and action, trained with a future-informed dual-branch design. The prior branch infers latent states from current context for deployment, while a posterior branch uses future observations during training. Aligning both branches forces the prior to learn future-aware, action-useful representations without ever generating frames at inference. The model achieves state-of-the-art or comparable performance across six simulation benchmarks and diverse real-world robotics tasks, offering efficient deployment like direct VLA policies but with the predictive benefits of world models.