Latent world models such as V-JEPA 2 predict future states in the feature space of a video encoder, but a single visual glance does not reveal how an object will rotate, bounce, or slide. The authors of a new arXiv paper argue that people learn these properties by trying — physically interacting with objects to see how they respond.
Their proposed method, IMPULSE, follows that intuition: push first, then predict. By adding an interaction step before forecasting in the V-JEPA 2 feature space, the model can probe physical characteristics that are not visible from appearance alone.
The abstract is brief and does not include experimental results, so the paper's effectiveness is not yet clear from this summary. Still, the core idea is a straightforward departure from passive observation, suggesting that active exploration may be a useful complement to latent world models.