CausalWAM: Learning the causal sources of actions in world action models

The study proposes CWAM (Causal WAM), which adds a sparse routing path to the action head of a pretrained world action model, separating action-relevant latent variables from the other variables that world modeling depends on, while keeping the world modeling objective unchanged. The authors prove that under a sparsity constraint, the latent variables used to generate actions are merely an invertible function of the action-relevant latent variables, that the true structure is identifiable up to a permutation of the latent variables, and that this holds without relying on sufficient temporal dependence or assumptions about the evolution of unobserved states. Synthetic data and real-robot experiments validate the theory and show improved closed-loop success rate efficiency.