SPA-BENCH:
A COMPREHENSIVE BENCHMARK FOR
SMARTPHONE AGENT EVALUATION

Derek Yuen1 *, Jingxuan Chen1 *, Bin Xie2, Yuhao Yang1, Gongwei Chen2, Zhihao Wu1, Yixing Li2, Xurui Zhou2, Weiwen Liu1, Shuai Wang1, Kaiwen Zhou1, Rui Shao2 †, Liqiang Nie2, Yasheng Wang1, Jianye Hao1, Jun Wang3, Kun Shao1 †
1 Huawei Noah's Ark Lab
2 Harbin Institute of Technology, Shenzhen
3 University College London
* Equal Contribution
† Corresponding authors: shaorui@hit.edu.cn, shaokun2@huawei.com
Paper Code arXiv Task Collection

Comparison to Existing Benchmarks:

Dataset Third-party app? Cross-app? Chinese app? Difficulty level? Number of tasks Number of agents Number of metrics Free of hand-crafted validation? Information for success detection
AndroidArena ✗ ✓ ✗ ✗ 221 1 4 ✗ Action only
AndroidWorld ✓ ✓ ✗ ✓ 116 3 1 ✗ State only
LlamaTouch ✓ ✗ ✗ ✓ 495 4 1 ✗ State only
B-MoCA ✗ ✗ ✗ ✗ 60 3 1 ✗ State only
MobileAgentBench ✗ ✗ ✗ ✓ 100 5 6 ✗ Action and State
SPA-Bench ✓ ✓ ✓ ✓ 340 11 7 ✓ Action and State

Abstract

Smartphone agents are increasingly important for helping users control devices efficiently, with (Multimodal) Large Language Model (MLLM)-based approaches emerging as key contenders. Fairly comparing these agents is essential but challenging, requiring a varied task scope, the integration of agents with different implementations, and a generalisable evaluation pipeline to assess their strengths and weaknesses. In this paper, we present SPA-BENCH, a comprehensive SmartPhone Agent Benchmark designed to evaluate (M)LLM-based agents in an interactive environment that simulates real-world conditions. SPA-BENCH offers three key contributions: (1) A diverse set of tasks covering system and third-party apps in both English and Chinese, focusing on features commonly used in daily routines; (2) A plug-and-play framework enabling real-time agent interaction with Android devices, integrating over ten agents with the flexibility to add more; (3) A novel evaluation pipeline that automatically assesses agent performance across multiple dimensions, encompassing seven metrics related to task completion and resource consumption. Our extensive experiments across tasks and agents reveal challenges like interpreting mobile user interfaces, action grounding, memory retention, and execution costs. We propose future research directions to ease these difficulties, moving closer to real-world smartphone agent applications.

📋 Diverse and Realistic Task Design

Key Features:

  • 📦 340 Tasks - 300 Single-app Tasks and 40 Cross-app Tasks
  • 🌐 66 Apps – 52 Third-party Apps, 7 Google Apps and 7 System Apps
  • 🌍 2 Languages – Chinese and English Apps
  • 📊 Increased Difficulty Levels
  • 🎨 Human Annotated Trajectories & Key Components

🤖 Plug-and-Play Agent Framework

Key Features:

  • 🧠 11 Smartphone Agents Ready for Evaluation
  • 🧩 Easy Integration of Your Own Agents with Minimal Code Changes
  • 📱 Scalable Design – Multi-device support & Emulator Compatibility
  • 📸 Android Snapshot – Local Environment Setup and Data Reset for Consistent Testing

An overview of the agent framework using a multi-processing architecture. Each worker process connects an agent to an Android emulator, and they interact multiple times throughout the task (i.e., step 3 is repeated) until completion. The emulators are reset after the agent has executed all assigned tasks.

✅ Automatic and Scalable Evaluation Pipeline

Key Features:

  • 🔍 7 Evaluation Metrics for a Comprehensive Analysis
  • 📐 Coarse-and-Fine Success Detection Pipeline – Requires No Further Human Effort
  • 🔀 Trajectory Splitting & Subtask Evaluation – Tailored for Long-Sequence Tasks
  • 🏆 Single-app Tasks – Achieved F1-scores of 0.926 (English) and 0.884 (Chinese)
  • 🌟 Cross-app Tasks – Achieved F1-scores of 0.833 (English) and 0.857 (Chinese)

Results

Agent Success (%) Mean Step Ratio on Success Termination Reason Termination Inaccuracy Mean Exec Time per Step (sec) Mean Token Cost per Step (USD)
SRC (%) MSR (%) Error (%) Premature (%) Overdue (%)
Off-the-Shelf Model (GPT-4o)
AppAgent 0.340 1.33 0.327 0.507 0.166 0.347 0.197 26.5 0.014
AutoDroid 0.327 1.10 0.593 0.340 0.067 0.494 0.078 34.0 0.008
MobileAgent 0.387 1.24 0.367 0.633 0 0.109 0.095 27.1 0.053
MobileAgentV2 0.433 1.05 0.580 0.420 0 0.333 0.111 56.1 0.067
M3A 0.640 0.92 0.847 0.153 0 0.244 0 19.3 0.092
T3A 0.487 1.04 0.707 0.293 0 0.368 0.136 9.6 0.116
SeeAct 0.393 1.60 0.200 0.773 0.027 0.100 0.276 41.2 0.046
Fine-tuned Model
Auto-UI 0.013 1.50 0.060 0.940 0 1.000 0.015 - -
CogAgent 0.020 1.67 0.147 0.820 0.033 1.000 0.024 - -
DigiRL 0.020 1.52 0.227 0.607 0.166 0.971 0.022 - -
OdysseyAgent 0.053 2.00 0 1.000 0 - 0.013 - -

Task performance on single-app English tasks. SRC and MSR refer to Self-Reported Completion and Maximum Steps Reached, respectively. The execution time and token costs of the last four agents are omitted because they use locally hosted open-source models.

For full results and more details, please refer to our paper.

BibTeX


@inproceedings{chen2024spa,
  title={SPA-Bench: A Comprehensive Benchmark for SmartPhone Agent Evaluation},
  author={Chen*, Jingxuan and Yuen*, Derek and Xie, Bin and Yang, Yuhao and Chen, Gongwei and Wu, Zhihao and Yixing, Li and Zhou, Xurui and Liu, Weiwen and Wang, Shuai and Shao, Rui and Nie, Liqiang and Wang, Yasheng and Hao, Jianye and Wang, Jun and Shao, Kun},
  booktitle={NeurIPS 2024 Workshop on Open-World Agents},
  year={2024}
}