arXiv:2611.05448v1 Announce Type: cross Abstract: Large language models understand passengers' driving commands, but as driving agents they reason in generated text for seconds and must infer each action's consequences implicitly from the scene.