Xiaomi-GUI-0 Technical Report

2026-06-30 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Internet of Things (IoT) & Connected Devices · Depth: Expert, quick

Summary

Xiaomi-GUI-0 is a native multimodal graphical user interface (GUI) agent designed for real mobile environments, aiming to close the gap between benchmark scores and real-world usability. It completes user tasks end-to-end through interface actions like tapping and text entry. The agent is trained and evaluated within a real-device closed loop, utilizing a real-device-dominant hybrid infrastructure where physical devices are the primary execution environment. Its training data is multi-source, covering high-frequency tasks, long-tail intents, and capability-enhancement data for reflection and memory, augmented by an error-driven data flywheel. The model undergoes a three-stage progressive pipeline: supervised fine-tuning, step-level reinforcement learning, and agentic reinforcement learning. Xiaomi-GUI-0 achieves 72.0% success on RealMobile and 78.9% on AndroidWorld, significantly improving execution stability and abnormal-state recognition in real tasks.

Key takeaway

For AI Engineers developing GUI agents for mobile applications, recognize that traditional benchmarks often misrepresent real-world performance due to dynamic interface layouts and abnormal states. You should prioritize a real-device closed-loop training and evaluation strategy, coupled with an error-driven data flywheel, to achieve robust execution stability and superior abnormal-state recognition. This approach moves beyond simulated environments, ensuring your agents are truly deployment-ready for complex, real-world mobile interactions.

Key insights

Real-device closed-loop training and error-driven data are crucial for robust GUI agent performance in dynamic mobile environments.

Principles

• Offline trajectories and simulated environments misrepresent real-world GUI agent performance.
• Real-device execution environments are essential for characterizing execution stability.
• An error-driven data flywheel improves agent performance by converting failures into corrected actions.

Method

Xiaomi-GUI-0 employs a progressive three-stage training pipeline: supervised fine-tuning, step-level reinforcement learning, and agentic reinforcement learning, integrated with an error-driven data flywheel.

In practice

• Develop GUI agents using a real-device-dominant hybrid infrastructure for data collection and evaluation.
• Implement an error-driven data flywheel to generate corrected actions and recovery demonstrations from agent failures.

Topics

• GUI Agents
• Mobile AI
• Reinforcement Learning
• Real-Device Training
• Multimodal AI
• Error-Driven Learning

Best for: AI Scientist, Research Scientist, Machine Learning Engineer, AI Engineer, MLOps Engineer

Related on AIssential

• GUI-Libra: Training Native GUI Agents to Reason and Act with Action-aware Supervision and Partially Verifiable RL — Tailored training and data curation significantly boost GUI agent performance on complex navigation tasks.
• not much happened today — Compact, multimodal AI models with extended context are enabling advanced on-device and agentic deployments, shifting focus to reliability and efficient operations.
• PhoneWorld: Scaling Phone-Use Agent Environments — PhoneWorld scales phone-use agent environment creation by converting real GUI trajectories into runnable, verifiable tasks and training rollouts.
• LAI #131: A Tool Call Can Succeed and Still Be the Wrong Tool — Data transparency and rigorous data curation are crucial for trustworthy AI model development.
• System Design for Maintaining Internal State Consistency in Long-Horizon Robotic Tabletop Games — Explicit state partitioning and recovery mechanisms are critical for robotic system reliability in long-horizon, turn-based interactions.
• The Sequence Radar #828: NVIDIA’s GTM Releases, Bezos’s $100B Bet, Xiaomi’s Ambush, and the Fracturing of OpenAI — The AI industry is rapidly shifting towards agentic infrastructure and physical world integration, moving beyond conversational paradigms.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.