MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization

[![Project Page](https://img.shields.io/badge/🌐%20Project-Page-orange.svg)](https://mobile-forge.github.io/) [![Media 机器之心 (Synced China)](https://img.shields.io/badge/Media-%E6%9C%BA%E5%99%A8%E4%B9%8B%E5%BF%83%20(Synced%20China)-f5f5f5.svg?logo=wechat&logoColor=07C160&labelColor=ffffff)](https://mp.weixin.qq.com/s/la86XLaDIwOB2mHHbl0oDQ) [![arXiv](https://img.shields.io/badge/arXiv-2606.19930-b31b1b.svg)](https://arxiv.org/abs/2606.19930) [![HuggingFace Models](https://img.shields.io/badge/🤗%20HuggingFace-Models-purple.svg)](https://huggingface.co/collections/lgy0404/mobileforge-models) [![Datasets](https://img.shields.io/badge/🤗%20HuggingFace-Datasets-yellow.svg)](https://huggingface.co/collections/lgy0404/mobileforge-datasets) [![Benchmark Results](https://img.shields.io/badge/🤗%20Benchmark-Results-blue.svg)](https://huggingface.co/datasets/lgy0404/mobileforge-benchmark-results) [![License](https://img.shields.io/badge/License-Apache_2.0-green.svg)](https://opensource.org/licenses/Apache-2.0) MobileForge turns real target-app interaction into executable curricula, hierarchical rollout feedback, and hint-contextualized policy updates without human-written tasks, demonstrations, or reward labels.

## 🔥 News - **2026-07-09**: MobileForge was featured by 机器之心: [不用人工标注，GUI Agent跑起「数据飞轮」：快手、浙大开源MobileForge](https://mp.weixin.qq.com/s/la86XLaDIwOB2mHHbl0oDQ). - **2026-06-23**: Released the MobileForge [codebase](https://github.com/kwai/MobileForge), [🤗 datasets](https://huggingface.co/collections/lgy0404/mobileforge-datasets), and [🤗 benchmark results](https://huggingface.co/datasets/lgy0404/mobileforge-benchmark-results). - **2026-06-19**: MobileForge preprint is available on [arXiv](https://arxiv.org/abs/2606.19930). - **2026-06-10**: Released all MobileForge [🤗 model checkpoints](https://huggingface.co/collections/lgy0404/mobileforge-models). ## 📊 Main Results

MobileForge main benchmark performance

MobileForge improves mobile GUI agents through annotation-free target-app adaptation. With GUI-Owl-1.5-8B, MobileForge reaches **67.24% Pass@1** and **77.59% Pass@3** on AndroidWorld, and **41.03% SR** on MobileWorld. With Qwen3-VL-8B, MobileForge raises AndroidWorld Pass@3 to **67.24%**. ## 🧩 Overview MobileForge adapts mobile GUI agents without collecting task-specific human annotations. It combines **MobileGym** for target-app exploration and automatic curriculum generation with **HiFPO** for hint-guided rollout, hierarchical trajectory feedback, and step-level GRPO training. ### MobileGym: target-app interaction and hierarchical feedback MobileGym grounds the adaptation loop in real target-app interaction. It explores Android apps, mines executable curriculum tasks from interaction traces, executes rollouts, and evaluates completed attempts with outcome labels, step-level process feedback, and corrective hints.

MobileGym framework

### HiFPO: feedback-guided policy optimization HiFPO turns MobileGym feedback into training signals. It runs hint-guided multi-attempt rollouts, filters mastered tasks and low-quality steps, retains informative experience, and trains the agent with hint-contextualized step-level GRPO.

HiFPO optimization pipeline

## 📁 Repository Guide ```text MobileForge/ |-- explore/ # Target-app exploration and MobileGym-Curriculum task generation |-- rollout/ # Hint-guided rollout, critic feedback, and rollout-to-GRPO processing |-- training/ # VERL-derived MobileForge step-level GRPO training stack |-- evaluation/ | |-- androidworld/ # AndroidWorld evaluation fork and reproduction utilities | `-- mobileworld/ # MobileWorld reproduction notes and helpers |-- docs/ # Models, data release, pipeline, and evaluation-result mapping |-- metadata/ # Public release manifests and model/result maps |-- CITATION.cff |-- citations.bib `-- README.md ``` ## 🚀 Where to Start The root README is intentionally concise. Detailed setup and commands live in the component README files. | Goal | Start here | What it covers | | --- | --- | --- | | Explore target apps and generate tasks | [`explore/`](explore/README.md) | Target-app exploration, APK cache, parallel exploration, and MobileGym-Curriculum task generation. | | Run hint-guided rollouts and build GRPO data | [`rollout/`](rollout/README.md) | Multi-attempt rollout, MobileGym-Critic feedback, hint reuse, and rollout-to-training-data conversion. | | Train with HiFPO / step-level GRPO | [`training/`](training/README.md) | Training environment, GRPO launch script, reward function, and utility tools. | | Reproduce benchmark runs | [`evaluation/`](evaluation/) and [`docs/evaluation_results.md`](docs/evaluation_results.md) | AndroidWorld and MobileWorld artifact mapping and evaluation notes. | | Inspect release manifests | [`docs/`](docs/) and [`metadata/`](metadata/) | Model list, dataset release notes, pipeline overview, and model-to-result mapping. | ## 📦 Release Index | Artifact | Link | Details | | --- | --- | --- | | Models | [🤗 MobileForge Models collection](https://huggingface.co/collections/lgy0404/mobileforge-models) | Main ForgeQwen3 / ForgeOwl checkpoints and scaling-ablation checkpoints. See [`docs/models.md`](docs/models.md). | | Datasets | [🤗 MobileForge Datasets collection](https://huggingface.co/collections/lgy0404/mobileforge-datasets) | Training data, exploration trajectories, and generated tasks. See [`docs/data_release.md`](docs/data_release.md). | | Benchmark results | [🤗 `lgy0404/mobileforge-benchmark-results`](https://huggingface.co/datasets/lgy0404/mobileforge-benchmark-results) | AndroidWorld and MobileWorld archives. See [`docs/evaluation_results.md`](docs/evaluation_results.md). | | Paper | [arXiv:2606.19930](https://arxiv.org/abs/2606.19930) | Technical report and citation metadata. | ## Citation ```bibtex @article{liu2026mobileforge, title={MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization}, author={Liu, Guangyi and Zhao, Pengxiang and Wu, Gao and Yin, Yiwen and Li, Mading and Liu, Liang and Liu, Congxiao and Qi, Zhang and Wang, Mengyan and Guo, Liang and others}, journal={arXiv preprint arXiv:2606.19930}, year={2026} } ``` ## Contact For questions about the paper, code, or released artifacts, contact [guangyiliu@zju.edu.cn](mailto:guangyiliu@zju.edu.cn). ## 🙏 Acknowledgements MobileForge builds on open-source resources including [AndroidWorld](https://github.com/google-research/android_world), [MobileWorld](https://github.com/Tongyi-MAI/MobileWorld), [MobileAgent](https://github.com/X-PLUG/MobileAgent), [Qwen3-VL](https://github.com/QwenLM/Qwen3-VL), [GUI-explorer](https://github.com/iLearn-Lab/ACL25-GUI-explorer), [VERL](https://github.com/verl-project/verl), and [GUI-R1](https://github.com/ritzz-ai/GUI-R1).