Official code for "MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization"
See the codeREADME
MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization
-f5f5f5.svg?logo=wechat&logoColor=07C160&labelColor=ffffff){kind=icon}
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.
- 2026-06-23: Released the MobileForge codebase, 🤗 datasets, and 🤗 benchmark results.
- 2026-06-19: MobileForge preprint is available on arXiv.
- 2026-06-10: Released all MobileForge 🤗 model checkpoints.
📊 Main Results
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.
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.
📁 Repository Guide
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/ | Target-app exploration, APK cache, parallel exploration, and MobileGym-Curriculum task generation. |
| Run hint-guided rollouts and build GRPO data | rollout/ | Multi-attempt rollout, MobileGym-Critic feedback, hint reuse, and rollout-to-training-data conversion. |
| Train with HiFPO / step-level GRPO | training/ | Training environment, GRPO launch script, reward function, and utility tools. |
| Reproduce benchmark runs | evaluation/ and docs/evaluation_results.md | AndroidWorld and MobileWorld artifact mapping and evaluation notes. |
| Inspect release manifests | docs/ and metadata/ | Model list, dataset release notes, pipeline overview, and model-to-result mapping. |
📦 Release Index
| Artifact | Link | Details |
|---|---|---|
| Models | 🤗 MobileForge Models collection | Main ForgeQwen3 / ForgeOwl checkpoints and scaling-ablation checkpoints. See docs/models.md. |
| Datasets | 🤗 MobileForge Datasets collection | Training data, exploration trajectories, and generated tasks. See docs/data_release.md. |
| Benchmark results | 🤗 lgy0404/mobileforge-benchmark-results | AndroidWorld and MobileWorld archives. See docs/evaluation_results.md. |
| Paper | arXiv:2606.19930 | Technical report and citation metadata. |
Citation
@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.
🙏 Acknowledgements
MobileForge builds on open-source resources including AndroidWorld, MobileWorld, MobileAgent, Qwen3-VL, GUI-explorer, VERL, and GUI-R1.
Contributors
4 commits
Similar repos
Languages
Python
57.9%
HTML
20.7%
JavaScript
18.1%
Kotlin
1.7%
Official code for "MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization"
See the codeREADME
MobileForge: Annotation-Free Adaptation for Mobile GUI Agents with Hierarchical Feedback-Guided Policy Optimization
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.
- 2026-06-23: Released the MobileForge codebase, 🤗 datasets, and 🤗 benchmark results.
- 2026-06-19: MobileForge preprint is available on arXiv.
- 2026-06-10: Released all MobileForge 🤗 model checkpoints.
📊 Main Results
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.
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.
📁 Repository Guide
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/ | Target-app exploration, APK cache, parallel exploration, and MobileGym-Curriculum task generation. |
| Run hint-guided rollouts and build GRPO data | rollout/ | Multi-attempt rollout, MobileGym-Critic feedback, hint reuse, and rollout-to-training-data conversion. |
| Train with HiFPO / step-level GRPO | training/ | Training environment, GRPO launch script, reward function, and utility tools. |
| Reproduce benchmark runs | evaluation/ and docs/evaluation_results.md | AndroidWorld and MobileWorld artifact mapping and evaluation notes. |
| Inspect release manifests | docs/ and metadata/ | Model list, dataset release notes, pipeline overview, and model-to-result mapping. |
📦 Release Index
| Artifact | Link | Details |
|---|---|---|
| Models | 🤗 MobileForge Models collection | Main ForgeQwen3 / ForgeOwl checkpoints and scaling-ablation checkpoints. See docs/models.md. |
| Datasets | 🤗 MobileForge Datasets collection | Training data, exploration trajectories, and generated tasks. See docs/data_release.md. |
| Benchmark results | 🤗 lgy0404/mobileforge-benchmark-results | AndroidWorld and MobileWorld archives. See docs/evaluation_results.md. |
| Paper | arXiv:2606.19930 | Technical report and citation metadata. |
Citation
@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.
🙏 Acknowledgements
MobileForge builds on open-source resources including AndroidWorld, MobileWorld, MobileAgent, Qwen3-VL, GUI-explorer, VERL, and GUI-R1.
Similar repos
Contributors
4 commits
Languages
Python
57.9%
HTML
20.7%
JavaScript
18.1%
Kotlin
1.7%