XPENG unveils X-Mind framework to advance predictive AI for autonomous driving

XPENG has introduced its new X-Mind technology framework, an AI architecture designed to enhance predictive reasoning and decision-making in autonomous vehicles. By utilizing an embedded predictive world model and a Visual Chain of Thought, the system allows vehicles to simulate future traffic scenarios and evaluate outcomes before taking action. This shift from reactive to proactive AI represents a significant step toward achieving safer, more human-like driving capabilities in production-ready autonomous systems.

XPENG’s General Intelligence Center head, Liu Xianming, detailed the X-Mind framework at the CVPR 2026 Workshop, emphasizing three core capabilities: proactive reasoning, controllable generation, and long-horizon temporal prediction. Unlike traditional perception-to-action pipelines that react only to immediate sensor data, X-Mind enables an in-vehicle AI agent to perform complex spatiotemporal reasoning. This approach addresses the limitations of current AI methods, such as text-based reasoning’s inability to grasp geometric relationships and the computational inefficiency of image-based future generation.

The framework introduces a Visual Chain-of-Thought (Visual CoT) that allows the model to carry out explicit reasoning before generating control actions. To make this practical for real-time deployment, XPENG utilizes thought sketches—compact visual representations that combine bird’s-eye-view layouts with abstract driving priors like lane boundaries, obstacles, and traffic signal status. By using a deep compression autoencoder (DC-AE), the system compresses 12 future frames of predictive data into just 96 tokens, significantly reducing the computational load compared to processing high-resolution images or 3D reconstructions.

X-Mind serves as a reasoning layer for XPENG’s vision-language-action (VLA) models, working alongside the previously announced X-Foresight framework. While X-Foresight focuses on predicting future multi-view scenes, X-Mind provides the cognitive inference necessary to make the model's decision-making process transparent and interpretable. Together, these technologies aim to advance VLA architecture toward a general-purpose physical AI capable of understanding real-world physics and anticipating future events. To further improve predictive efficiency, XPENG developed a Recursive Block Diffusion (RBD) mechanism, which performs the generative process internally across different layers of the driving model.