XCoT-VLA: Executable Chain-of-Thought for Vision-Language-Action Driving

A recent study published on arXiv (2608.10976) presents XCoT-VLA, a model aimed at enhancing Vision-Language-Action (VLA) systems for self-driving vehicles. The researchers contend that conventional verbose natural-language Chain-of-Thought (CoT) reasoning is not suitable for real-time control due to its open-ended nature, high decoding expenses, and optimization challenges. XCoT-VLA substitutes lengthy rationales with concise executable CoT tokens derived from automatically generated Reason-Action supervision. Action evidence is drawn from logged trajectories, while scene context provides causal semantics. The XCoT sequence is contextually relevant and conditions fixed trajectory queries via shared multimodal self-attention. Deterministic token-function routing employs the Reason FFN for XCoT tokens and the Control FFN for trajectory queries to facilitate flow-matching trajectory generation. Additionally, the paper presents XCoT Policy Optimization (XCPO) as an optional enhancement. This research tackles a significant issue in autonomous driving: achieving a balance between semantic reasoning and real-time control efficiency.