FedMABench: Benchmarking Mobile GUI Agents on Decentralized Heterogeneous User Data

Accepted by EMNLP 2025 Main Conference ⭐️

Overview

FedMABench is an open-source benchmark for federated training and evaluation of mobile agents, specifically designed for heterogeneous scenarios.

FedMABench includes the following key features:

• 6 datasets with 30+ subsets （over 800 apps across 5 categories）
• 3 types of heterogeneity (e.g., App Category Distribution, Specific App Preference, etc.)
• 8 federated learning algorithms (e.g., FedAvg, FedProx, SCAFFOLD, FedAvgM, etc.).
• 10+ base models covering Qwen2-VL-2B/7B-Instruct, InternVL2-1B/2B/4B/8B, DeepseekVL2 and more.

Setup

Clone the repo, submodules and install the required packages.

git clone --recursive --shallow-submodules https://github.com/wwh0411/FedMABench.git
cd FedMABench
conda create -n fedma python=3.10
conda activate fedma
pip install -r requirements.txt

Data Directory Tree

The six datasets of FedMABench are organized in seperate directories. We also provide test sets and temporary split files for reference.

.
├── App-Level 
│   ├── App Half-Skew.jsonl
│   ├── App IID.jsonl
│   ├── App Non-Uniform.jsonl
│   ├── App Skew.jsonl
│   ├── App-Level Val.jsonl
│   ├── train_split.json
│   └── val_split.json
├── Basic-AC
│   ├── Basic-AC Entertainment.jsonl
│   ├── Basic-AC Lives.jsonl
│   ├── Basic-AC Office.jsonl
│   ├── Basic-AC Shopping.jsonl
│   ├── Basic-AC Traveling.jsonl
│   ├── Basic-AC c10n1000.jsonl
│   ├── Basic-AC c10n200.jsonl
│   ├── Basic-AC c10n3000.jsonl
│   ├── Basic-AC c10n500.jsonl
│   ├── Basic-AC c10n5000.jsonl
│   ├── Basic-AC c10n7000.jsonl
│   ├── Basic-AC c30n3000.jsonl
│   ├── Basic-AC c50n5000.jsonl
│   ├── Basic-AC c70n7000.jsonl
│   ├── Val_100.jsonl
│   └── val_100.json
├── Basic-AitW
│   ├── Basic-AitW G-Apps.jsonl
│   ├── Basic-AitW General.jsonl
│   ├── Basic-AitW Install.jsonl
│   ├── Basic-AitW Single.jsonl
│   └── Basic-AitW WebShopping.jsonl
├── Category-Level 
│   ├── App Random.jsonl
│   ├── App Skew.jsonl
│   ├── Category Half-Skew.jsonl
│   ├── Category IID.jsonl
│   ├── Category Non-Unifrom.jsonl
│   ├── Category Skew.jsonl
│   ├── Category Val.jsonl
│   ├── train_split.json
│   └── val_split.json
├── ScaleApp
│   ├── ScaleApp IID.jsonl
│   ├── ScaleApp Random.jsonl
│   ├── ScaleApp Skew.jsonl
│   ├── ScaleApp Val_250.jsonl
│   ├── train_split.json
│   └── val_split.json
├── Step-Episode 
│   ├── Both Skew.jsonl
│   ├── Episode Skew.jsonl
│   ├── Step Skew.jsonl
│   ├── Step-Episode IID.jsonl
│   ├── Step-Episode Val.jsonl
│   └── val_split.json
└── README.md

Citation

@misc{wang2025fedmabenchbenchmarkingmobileagents,
      title={FedMABench: Benchmarking Mobile Agents on Decentralized Heterogeneous User Data}, 
      author={Wenhao Wang and Zijie Yu and Rui Ye and Jianqing Zhang and Siheng Chen and Yanfeng Wang},
      year={2025},
      eprint={2503.05143},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2503.05143}, 
}