Representational Priors Influence World Action Model Generalization

Research shows that the generative backbone of World Action Models (WAMs) dictates generalization via its representational geometry. Models with higher nuisance orthogonality—separating task-relevant features from visual variations—demonstrate superior robustness to out-of-distribution visual shifts.