Mobile GUI agents are increasingly deployed on interfaces shaped by users' histories and preferences, yet their reliability across different users remains largely unstudied. We introduce PAIR (Personalized Application‑state Instantiation and Rendering), a pipeline that constructs user‑conditioned application states, enabling controlled evaluation of the same task across multiple users. Building on this, we propose RePAIR (Reinforcement learning with Personalization‑Aware Interaction Rewards), a training approach that leverages cross‑user differences in sub‑goal outcomes to improve reliability in user‑conditioned mobile environments. Experiments with six agents reveal substantial variation in task success across users and a consistent drop in sub‑goal achievement within user‑conditioned UI contexts (decreases of 6.98–15.4 percentage points). When targets are drawn from each user's own content, the gap widens further (8.77–22.0 percentage points). Failures often involve selecting an unintended item before the target is exposed. RePAIR improves user‑conditioned success rate (+5.87 pp), all‑success rate (+7.50 pp), and overall task success rate (+9.42 pp) over its supervised fine‑tuning parent on unseen users, providing early evidence that explicitly learning from cross‑user variation can boost GUI‑agent reliability.
Review