PhoneBuddy: Training Open Models for Agentic Phone Use

A team of researchers has introduced PhoneBuddy, an open-model training recipe designed to improve the reliability of artificial intelligence agents operating smartphones, according to a paper submitted to the arXiv preprint server on 22 June 2026 [1]. The system addresses a core difficulty in training phone-based agents: the real-device environments needed for deployment are slow, stateful, and hard to reset, while scalable mock environments only approximate real behavior [1]. To bridge this gap, PhoneBuddy combines a real-app environment with a mock-app environment called PhoneWorld, which reconstructs runnable mock apps from real graphical user interface usage structure [1]. The training pipeline first builds a shared supervised fine-tuning stage from trajectories collected in both environments, then compares reinforcement learning (RL) performed solely on real apps against mixed RL conducted across both environments [1]. In a 150-task human evaluation on real phones spanning apps, mini-apps, and cross-app workflows, the task success rate improved from 36.67% after supervised fine-tuning to 40.67% after real-app RL and 45.33% after mixed RL [1]. On the AndroidWorld benchmark, the same progression rose from 60.3% to 77.2% to 83.2% [1]. The paper, authored by Zhengyang Tang and colleagues, concludes that mock-app training is not a replacement for real-app RL but a complementary source of scalable, resettable, and automatically checked interaction [1]. The gains were strongest on app and mini-app tasks, while long-horizontal cross-app workflows remain an open challenge [1]. The work arrives amid sustained investment in general-purpose AI agents. OpenAI’s ChatGPT, originally released in November 2022, accelerated an AI boom marked by rapid investment and public attention toward artificial intelligence [3]. By February 2026, the service had reached 900 million weekly active users [3]. PhoneBuddy’s focus on open models contrasts with the proprietary systems that dominate consumer deployments, offering a training recipe that other researchers can inspect and extend [1]. arXiv, where the paper was posted, is an open-access repository of electronic preprints that has grown to a submission rate of about 24,000 articles per month as of November 2024 [11]. The repository’s arXivLabs framework allows community collaborators to develop experimental tools that appear on article record pages, such as citation explorers and code finders, under guidelines that emphasize openness and user data privacy [9][10].