In brief: Xiaomi-GUI-0 is trained on real devices rather than in simulated environments, closing the gap between lab benchmarks and actual application stability in production.
Xiaomi has developed a GUI agent trained on real mobile devices that executes mobile tasks in real application environments with 72% success rate. The model is not guided by simulations, but by real-world scenarios such as authentication dialogs and risk controls from production operation.
Existing GUI agents are based on vision-language models and perform mobile tasks through direct interface interactions – tapping, swiping, text input, navigation. The problem: their training is predominantly conducted on offline recordings in simulated environments and standardized benchmarks. These differ significantly from real application scenarios in layout, interaction logic, and error distribution.
Xiaomi closes this gap through Xiaomi-GUI-0 with a hybrid-physical infrastructure where real devices serve as the primary execution environment and sandboxes function only in a supporting role. This ensures that data collection, training, rollout, and evaluation cover the same distribution as actual deployment. The model learns from three data sources: frequently performed head tasks, generalizable data for edge cases, and capability-enhancement data for reflection and memory. An “error-driven data flywheel” converts failed trajectories into corrected actions, reflective explanations, and recovery demonstrations.
Training occurs in three phases: supervised fine-tuning, step-level reinforcement learning, and agentic RL. On the internal RealMobile benchmark, Xiaomi-GUI-0 achieves a 72.0% success rate; on the public AndroidWorld benchmark, 78.9%. Crucially, the model demonstrates improved stability in anomalous states in real scenarios – authentication dialogs, permission requests, and payment verification, where traditional agents frequently fail.
Source: arxiv.org · Published 29 June 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification through Lumi News Pipeline v1.7.2.