SWAM: Scaling World Action Models with Cross-Modal Alignment
World action models jointly predict video and actions conditioned on language. Existing studies focus primarily on architectures and training mechanisms, typically treating language-video alignment acquired through video…