TestFlow: Advancing Mobile UI Testing through Multi-Step Reinforcement Learning

GUI Agents have demonstrated promising applications in mobile UI testing. However, for complex testing tasks, UI agents tend to fail due to their greedy approach in executing step-by-step operations, leading to error accumulation and neglecting long-horizon dependencies. To address these limitations, we propose TestFlow, a novel multi-modal UI testing model that combines Supervised Fine-Tuning with a Task-aware Reinforcement Learning framework. Our approach implements a two-phase training pipeline designed to optimize long-horizon instruction compliance and complex task completion. Additionally, we develop a tailor-made reward function that integrates both process and outcome rewards to improve the completion rate of multi-step tasks. The experimental results demonstrate that TestFlow significantly outperforms the baseline methods, achieving 33. 69% WTSR and 55. 37% SSR in cross-page test scenarios. These improvements highlight the practical value of TestFlow in addressing the challenges of modern mobile app testing, particularly in industrial settings requiring high adaptability and reliability.