自動駕駛／多模態模型
XCoT-VLA Compresses Autonomous-Driving Reasoning into 2–6 Action Tokens, with Interface Latency as Low as 38.6 ms on H100
XPeng introduces XCoT-VLA, replacing dozens of natural-language reasoning tokens with a limited set of semantic action tokens, then using flow matching to generate continuous trajectories. Open-loop tests show lower lateral error, but no closed-loop safety evaluation has been reported, and the latency figures do not cover the full in-vehicle hardware and software stack.
XPeng’s research team has unveiled XCoT-VLA, an attempt to resolve an engineering tradeoff in vision-language-action models: natural-language reasoning can express scenes and causal relationships, but autoregressive token-by-token decoding is difficult to fit into real-time planning cycles. The new method compresses reasoning into 2–6 tokens from a finite vocabulary, such as `LEFT_TURN_PREPARE`, `DECELERATE`, and `RED_LIGHT_HOLD`. These tokens remain in the attention context and directly condition fixed trajectory queries.
The architecture shares multimodal self-attention, but uses deterministic routing to send reasoning tokens to the Reason FFN and trajectory queries to the Control FFN, preventing reasoning fine-tuning from disrupting control representations. Rather than outputting discrete driving commands, the control branch uses conditional flow matching to predict longitudinal acceleration and yaw changes for the next 24 steps, which are then integrated into a six-second trajectory. The training dataset contains 3.62 million samples: 3.1 million automatically labeled driving logs, 200,000 human-annotated samples, and 320,000 targeted-mining lane-change cases.
On in-distribution tests, longitudinal ADE fell from 1.645 meters to 1.323 meters. In the lane-change ablation study, lateral FDE dropped from 1.616 meters to 1.112 meters, with XCoT alone delivering a 25.1% reduction. The authors measured an inter-token latency of 3.25 ms on an H100. With inputs ranging from 2K to 6K tokens, the XCoT reasoning interface is estimated to take 38.6–66.3 ms, below the 83.3 ms budget for planning at 12 Hz, while an 80-token textual CoT requires 279.1–306.8 ms.
These figures should not be interpreted as production-vehicle safety results or end-to-end latency. The evaluation covers only open-loop trajectory error, while the latency measurements exclude perception preprocessing, trajectory postprocessing, and scheduling. The paper’s proposed XCPO reinforcement-learning stage is also not quantified. The model, data, and code have not been released, and long-horizon longitudinal error remains slightly worse than that of the trajectory-only baseline. The next points to watch are full-stack latency on actual in-vehicle chips and closed-loop results for collisions, disengagements, traffic violations, comfort, and route-completion rate.