CausalWAM: Learning the Causes of Actions in World Action Models
arXiv cs.AI · 31 Oct 2026, 03:40 AEST
CausalWAM introduces a method for learning the causes of actions in World Action Models. This approach separates action-relevant latent variables from nuisance variables to improve action generation and prediction accuracy.
Key facts
- World action models (WAMs) learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions.
- We therefore propose CWAM (Causal WAM), which incorporates the constraint into a pretrained WAM by adding a sparse routing path to its action head, with the world modeling objective remaining unchanged.
- Experiments on both synthetic data and real-world robots validate the theory as well as showcase the improved efficiency in closed-loop success.
Read aloud
CausalWAM introduces a method for learning the causes of actions in World Action Models. This approach separates action-relevant latent variables from nuisance variables to improve action generation and prediction accuracy. World action models WAMs learn latent variables from observations and use these to both predict the evolution of the scene and to generate actions. We therefore propose CWAM Causal WAM, which incorporates the constraint into a pretrained WAM by adding a sparse routing path to its action head, with the world modeling objective remaining unchanged. Experiments on both synthetic data and real-world robots validate the theory as well as showcase the improved efficiency in closed-loop success.
Sources in this cluster
- CausalWAM: Learning the Causes of Actions in World Action Models
arXiv cs.AI · 31 Oct 2026, 03:40 AEST