AppDeltaWorld predicts the next mobile screen as an executable HTML diff rather than a generated image

arXiv·low signalXBlueskyLinkedInCopy link

Xu et al. (arXiv 2608.05891, submitted August 6) tackle the problem that real mobile GUI trajectories are unobtainable for sensitive apps while simulated environments are expensive to scale. Their world model predicts the next GUI as a reachable code update: it retrieves app-specific Level-1 HTML references under an action-transition constraint, generates Level-2 executable HTML conditioned on the current screen, action, predicted next-screen text and retrieved structure, then inserts visual assets into image slots before browser rendering. It posts the highest fidelity on CMGUIBench-500 under Code2World evaluation over image-only and code-only baselines, and as a training environment enables AppDeltaAgent to reach state of the art on AndroidLens with gains on MobileGym and MobileWorld, plus test-time RL adaptation without touching real apps.