MobiTwin: Toward building end-to-end framework for user's mobile twins

Creating authentic digital twins of mobile users through realistic user behavior simulation is critical to truly understand and anticipate customer needs at scale. Toward this, we introduce Digital TwIns of MObile User (TIMO), an end-to-end framework for high-fidelity mobile user simulation that addresses four key limitations in existing approaches: subjective decision-making diversity, scalable experience retention, simultaneous multi-option processing, and granular mobile interaction fidelity. Our approach integrates adaptive persona generation, self-evolving contextual memory, multi-faceted behavioral simulation, and high-fidelity mobile device control to create digital twins that replicate real user behavior patterns. Unlike existing methods that model decisions in isolation, our framework enables simulated agents to simultaneously evaluate multiple options through joint decision-making architectures, closely mirroring how real users process competing choices within limited mobile screen real estate. Evaluated on both public benchmarks and proprietary datasets from real-world mobile scenarios, our framework demonstrates significant improvements over existing baselines in user decision prediction accuracy, robustness to positional bias, and computational efficiency. Beyond task-related evaluation, we also validate our digital twins through reasoning analysis, and dedicated emotion perception experiments, showing how they align closely with authentic human behavioral patterns, providing a more holistic perspective on scalable simulacrum.