Xiaomi-GUI-0: A Real-World Mobile GUI Agent for Enhanced Stability.

Traditional GUI agents often struggle in real-world mobile applications because their training and evaluation environments, like offline trajectories or simulations, don't accurately reflect the complexities of live usage. Factors such as account states, permission dialogues, and payment authentication constantly introduce variability that benchmarks fail to capture. This discrepancy leads to a significant gap between reported performance and actual usability. To address this, researchers have developed Xiaomi-GUI-0, a native multimodal GUI agent specifically designed for real mobile environments. Its core innovation is a hybrid infrastructure where physical devices are the primary execution environment, supported by sandboxes. This setup ensures that data collection, training, deployment, and evaluation occur in a distribution closely mirroring real-world conditions. The agent utilizes a multi-source training data approach, incorporating high-frequency tasks, generalization data for diverse intents, and capability-enhancement data for reflection and memory. An error-driven data flywheel converts failure trajectories into corrective actions and recovery demonstrations. Through a three-stage progressive training pipeline, Xiaomi-GUI-0 achieves strong success rates on both public benchmarks and its in-house RealMobile evaluation, significantly boosting execution stability and abnormal-state recognition in practical tasks.