BayesianVLA: Bayesian Decomposition of Vision Language Action Models via Latent Action Queries - Explained Simply | ArXiv Explained