---

🚨 News

• 2026-08-24: 🌍 PhoneWorld is now publicly released for research, including the benchmark runner, 120 evaluation tasks, 300 verified training tasks, and the Research Source Edition for 34 mock apps. Research APKs are available through the gated Hugging Face dataset.
• 2026-06-26: 📄 arXiv preprint is live: arXiv:2606.23049.
• 2026-06-15: 🤗 PhoneBuddy models are public on Hugging Face: PhoneBuddy-4B, PhoneBuddy-4B-RealApp, and PhoneBuddy-0.8B.
• 2026-06-12: 🌐 Project page launched: phonebuddyai.github.io.
• 2026-06-11: 📄 Paper snapshot and result figures added to the project page.
• 2026-06-10: 🧭 The project page now connects the five-work phone-agent research line listed above.

---

✨ What Is PhoneBuddy?

Most mobile agents are evaluated as GUI controllers: observe a screen, tap, type, swipe, repeat. PhoneBuddy studies a training recipe for open phone-use models that can improve under real execution feedback while also benefiting from scalable mock-app supervision.

PhoneBuddy compares a shared SFT checkpoint, real-app RL, and mixed real+mock RL. The mixed recipe uses PhoneWorld-style mock apps as resettable environments with automatic verifiers, then evaluates whether this scalable signal transfers back to real-phone tasks and AndroidWorld.

---

🌍 Get PhoneWorld

PhoneWorld is the environment stack behind PhoneBuddy's mock-app training. Its public research release includes the evaluation runner, task definitions and verifiers, 120 benchmark tasks, 300 verified training tasks, and the Research Source Edition for all 34 mock Android apps.

• Code, tasks, source apps, and setup: github.com/EthanLeo-LYX/PhoneWorld
• Research APKs: EthanLeoLYX/PhoneWorld-APKs — gated access for non-commercial academic research
• Paper: PhoneWorld: Scaling Phone-Use Agent Environments

Please follow the PhoneWorld repository's research license, APK access terms, and benchmark reporting guidance.

---

🤗 Model Zoo

The public model release follows the Qwen-style XML tool-call format defined in the model chat_template.jinja. Dataset artifacts are not planned for public release at this stage.

---

📊 Results Snapshot

Takeaway. Real-app RL substantially improves over SFT. Adding mock-app RL further improves the average result, with the strongest gains on single-app tasks and AndroidWorld.

---

🧭 Phone-Agent Research Gallery

PhoneBuddy is one piece of a larger phone-agent stack: environments, training, runtime, privacy, and safety.

---

🗂️ Repository Layout

phonebuddy/
├── assets/
│   ├── figures/      # Paper and project figures
│   └── paper.pdf     # Current paper snapshot
├── docs/             # Public documentation drafts
└── README.md

---

📌 Release Plan

• ✅ Project page and paper snapshot
• 🚧 PhoneBuddy-4B model release
• 🚧 Lightweight and ablation checkpoints
• 🚧 Code release and evaluation documentation
• ❌ No public dataset release planned at this stage

---

📚 Citation

@misc{tang2026phonebuddytrainingopenmodels,
      title={PhoneBuddy: Training Open Models for Agentic Phone Use},
      author={Zhengyang Tang and Xin Lai and Pengyuan Lyu and Xinyuan Wang and Tianyi Bai and Chenxin Li and Yiduo Guo and Huawen Shen and Yuxuan Liu and Junyi Li and Zhengyao Fang and Yang Ding and Yi Zhang and Weinong Wang and Xingran Zhou and Liang Wu and Fei Tang and Sunqi Fan and Shangpin Peng and Zheng Ruan and Anran Zhang and Benyou Wang and Ji-Rong Wen and Rui Yan and Chengquan Zhang and Han Hu},
      year={2026},
      eprint={2606.23049},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2606.23049},
}

---