StepReflect Enhances Mobile GUI Agent Accuracy and Efficiency

Autonomous agents designed to interact with mobile graphical user interfaces (GUIs) often struggle with long-horizon tasks due to the need for accurate action reflection. Current approaches typically rely on expensive, open-ended multimodal reasoning after each action, which is not well-suited to the structured nature of GUI state transitions. StepReflect addresses this by reframing per-step GUI reflection as a supervised structured prediction problem. It conditions this prediction on explicit transition specifications and visual evidence, making the reflection process more targeted and efficient. The model is trained through a multi-stage pipeline involving fine-tuning, distillation, and preference-based refinement. The resulting 8B model demonstrates significant improvements. Offline, it achieves 82.16% transition-level accuracy on AndroidWorld, outperforming zero-shot GPT-5.2 by a notable margin. Online, StepReflect leads to higher task success in most agent configurations and substantially reduces paid API charges compared to GPT-based reflection. This establishes StepReflect as a practical, locally deployable alternative for building reliable mobile GUI agents.