multimodalcritiqueneutralExisting world models incorrectly entangle physical dynamics with pixel rendering and require continuous visual observation to sustain motionComputer Vision27 Jul 2026http://arxiv.org/abs/2607.02517v1