The paper proposes CWAM, a world action model that routes actions through only the latent variables that cause them, and argues it provably recovers the true causes.
The idea: World action models predict the whole scene, but only part of it drives actions. CWAM adds a sparse routing path to the action head of a pretrained model and leaves the world modeling objective unchanged.
The claim: Under a sparsity constraint, the action-relevant latent variables are identified up to a permutation. The paper says this needs no sufficient temporal dependence and no assumptions about how unobserved regimes evolve.
Testing: Experiments on synthetic data and real-world robots validate the theory and show improved efficiency in closed-loop success.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper proposes CWAM, a world action model that routes actions through only the latent variables that cause them, and argues it provably recovers the true causes.