Mobile GUI Agent Privacy Personalization with Trajectory Induced Preference Optimization

Researchers propose Trajectory Induced Preference Optimization (TIPO), a novel method for training mobile GUI agents to respect user privacy preferences while maintaining task execution capability. The approach addresses the challenge that privacy-conscious users generate structurally different execution patterns than utility-focused users, requiring specialized optimization techniques to properly align agent behavior with individual privacy preferences.

This research addresses a critical but underexplored intersection of AI capability and user autonomy in mobile computing. As multimodal large language models increasingly control device interactions, the ability to personalize agent behavior around privacy preferences becomes essential infrastructure rather than optional feature. The paper identifies a fundamental technical problem: users with different privacy philosophies produce trajectories too structurally divergent for standard preference optimization methods, which assume relatively homogeneous behavioral patterns.

The development reflects broader maturation in AI systems design. Early MLLM-powered agents focused purely on task completion metrics, treating all users as identical. This work recognizes that optimization for a single objective function—whether task success or privacy compliance—often produces tension. TIPO's solution using preference-intensity weighting and padding gating represents practical progress toward multi-objective agent training.

For the AI industry, this signals growing recognition that personalization cannot remain an afterthought. Users increasingly demand agency over how automated systems interact with their devices, particularly regarding permissions and data exposure. The reported metrics—65.60% success rate, 46.22% compliance score, and 66.67% persona distinction—demonstrate the approach successfully balances competing objectives, though room for improvement exists.

The public release of code and a Privacy Preference Dataset creates valuable infrastructure for the community. This mirrors patterns where reproducible benchmarks accelerate progress. Organizations deploying mobile agents will likely need similar personalization capabilities to meet user expectations and regulatory requirements, making TIPO's technical contributions directly applicable to production systems.

• →TIPO enables mobile GUI agents to respect individual privacy preferences without sacrificing task execution performance.
• →Privacy-conscious users exhibit structurally different execution patterns requiring specialized optimization techniques distinct from standard preference learning.
• →The method achieves 65.60% success rate while maintaining 46.22% privacy compliance across diverse mobile tasks.
• →Public release of code and Privacy Preference Dataset will accelerate community research in personalized AI agent behavior.
• →Privacy personalization is evolving from optional feature to essential requirement as autonomous agents gain deeper device control.