New AI Model Simulates Mobile App Interactions with Code.

Training mobile GUI agents to operate applications through pixel perception and touch actions is a promising avenue for developing long-horizon mobile interaction policies. However, obtaining real-world interaction data is often challenging due to privacy concerns and the sensitivity of certain applications. Existing simulated environments also face issues with scalability, while current GUI world models struggle with unstable generation, limited modality coverage, and inconsistent action-transition logic. To overcome these limitations, researchers propose AppDeltaWorld, a novel transition-grounded delta code world model. Instead of predicting the next GUI as an unconstrained image or text, AppDeltaWorld forecasts it as a reachable code update. This model retrieves app-specific Level-1 HTML references under an action-transition constraint, then generates Level-2 executable HTML conditioned on the current screen, action, predicted next-screen text, and retrieved structure. Finally, it inserts generated visual assets into image slots for browser rendering. As a world model, AppDeltaWorld demonstrates superior fidelity on the CMGUIBench-500 benchmark, showing clear improvements in structural layout and UI element reconstruction compared to image-only and code-only baselines. Furthermore, as a training environment, it supports the construction of filtered closed-loop SFT data, which, when combined with public supervision, allows AppDeltaAgent to achieve state-of-the-art performance on AndroidLens and consistent gains on MobileGym and MobileWorld. The model also enables policy adaptation through world-model-based test-time reinforcement learning, further enhancing performance without additional real-app interactions.