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

<h1 id="technical" Technical Analysis</h1 Jump to Section: Motivation motivation | Results summary-of-results | Conclusions final-conclusions "Being-H0.7" https://research.beingbeyond.com/being-h07 is a latent world-action model that brings future-aware reasoning into VLA-style policies without generating future frames. Being-H0.7 inserts learnable latent queries between perception and action as a compact reasoning interface, and trains them with a future-informed dual-branch design: a deployable prior branch infers latent states from the current context, while a training-only posterior branch replaces the queries with embeddings from future observations. High-level Analogy : Imagine you're trying to catch a ball. The old way VLA is like instantly reacting to where the ball is right now – quick, but you might miss if it takes an unexpected bounce. Another way WAM is to try and predict every single pixel of the ball's future movement, like a super-detailed mental video – very accurate but incredibly slow and often overkill. Being-H0.7 is like a smart catcher who, when seeing the ball now , quickly processes only the most crucial future information needed to make the catch, such as its trajectory, speed, and likely landing spot. It doesn't waste time predicting every visual detail. To learn this skill, it uses a 'training-only cheat sheet' that shows it the actual future trajectory. But in a real game, it throws the cheat sheet away and uses its finely tuned intuition to make e