Latent action models (LAMs) are designed to learn action-like representations from unlabeled video by compressing what changes between frames. The idea is that the compressed signal corresponds to the agent's own actions, without needing manual labels.

According to the arXiv preprint, the approach runs into trouble on in-the-wild footage. Such videos contain not only the agent's state but also exogenous elements—background motion, lighting shifts, camera movement, or other actors—that are unrelated to the agent's actions. These extra changes can contaminate the compression target, causing the model to learn spurious representations.

The paper's title promises a fix, and the abstract indicates that prevention is possible, but the provided excerpt cuts off before describing the proposed solution. Readers will need to consult the full text for the method. As with any single preprint, the claims have not yet been peer-reviewed or independently replicated.