LearnGUI: A Unified Demonstration Benchmark for Mobile GUI Agents

📄 Paper | 💻 Code | 🌐 Project Page

Overview

LearnGUI is the first comprehensive dataset specifically designed for studying demonstration-based learning in mobile GUI agents. It comprises 2,353 instructions across 73 applications with an average of 13.2 steps per task, featuring high-quality human demonstrations for both offline and online evaluation scenarios.

🌟 Key Features

• Unified Benchmark Framework: Provides standardized metrics and evaluation protocols for demonstration-based learning in mobile GUI agents
• Dual Evaluation Modes: Supports both offline (2,252 tasks) and online (101 tasks) evaluation scenarios to assess agent performance
• Rich Few-shot Learning Support: Includes k-shot combinations (k=1,2,3) for each task with varying similarity profiles
• Multi-dimensional Similarity Metrics: Quantifies demonstration relevance across instruction, UI, and action dimensions
• Diverse Real-world Coverage: Spans 73 mobile applications with 2,353 naturally varied tasks reflecting real-world usage patterns
• Expert-annotated Trajectories: Contains high-quality human demonstrations with detailed step-by-step action sequences and element annotations

📊 Dataset Structure and Statistics

The dataset is organized into three main splits:

Dataset Statistics

Split	K-shot	Tasks	Apps	Step actions	Avg InsSim	Avg UISim	Avg ActSim	UISHActSH	UISHActSL	UISLActSH	UISLActSL
Offline-Train	1-shot	2,001	44	26,184	0.845	0.901	0.858	364	400	403	834
Offline-Train	2-shot	2,001	44	26,184	0.818	0.898	0.845	216	360	358	1,067
Offline-Train	3-shot	2,001	44	26,184	0.798	0.895	0.836	152	346	310	1,193
Offline-Test	1-shot	251	9	3,469	0.798	0.868	0.867	37	49	56	109
Offline-Test	2-shot	251	9	3,469	0.767	0.855	0.853	15	42	55	139
Offline-Test	3-shot	251	9	3,469	0.745	0.847	0.847	10	36	49	156
Online-Test	1-shot	101	20	1,423	-	-	-	-	-	-	-

Each task in LearnGUI contains:

• High-level instruction
• Low-level action sequences
• Screenshot of each step
• UI element details
• Ground truth action labels
• Demonstration pairings with varying similarity profiles

📁 Directory Structure

LearnGUI/
├── offline/                            # Offline evaluation dataset
│   ├── screenshot.zip                  # Screenshot archives (multi-part)
│   ├── screenshot.z01-z05              # Screenshot archive parts
│   ├── element_anno.zip                # Element annotations
│   ├── instruction_anno.zip            # Instruction annotations
│   ├── task_spilit.json                # Task splitting information
│   └── low_level_instructions.json     # Detailed step-by-step instructions
│
└── online/                             # Online evaluation dataset
    ├── low_level_instructions/         # JSON files with step instructions for each task
    │   ├── AudioRecorderRecordAudio.json
    │   ├── BrowserDraw.json
    │   ├── SimpleCalendarAddOneEvent.json
    │   └── ... (98 more task instruction files)
    └── raw_data/                       # Raw data for each online task
        ├── AudioRecorderRecordAudio/
        ├── BrowserDraw/
        ├── SimpleCalendarAddOneEvent/
        └── ... (98 more task data directories)

🔍 Comparison with Existing Datasets

LearnGUI offers several advantages over existing GUI datasets:

Dataset	# Inst.	# Apps	# Step	Env.	HL	LL	GT	FS
PixelHelp	187	4	4.2	❌	✅	❌	✅	❌
MoTIF	276	125	4.5	❌	✅	✅	✅	❌
UIBert	16,660	-	1	❌	❌	✅	✅	❌
UGIF	523	12	6.3	❌	✅	✅	✅	❌
AITW	30,378	357	6.5	❌	✅	❌	✅	❌
AITZ	2,504	70	7.5	❌	✅	✅	✅	❌
AndroidControl	15,283	833	4.8	❌	✅	✅	✅	❌
AMEX	2,946	110	12.8	❌	✅	❌	✅	❌
MobileAgentBench	100	10	-	❌	✅	❌	❌	❌
AppAgent	50	10	-	❌	✅	❌	❌	❌
LlamaTouch	496	57	7.01	✅	✅	❌	✅	❌
AndroidWorld	116	20	-	✅	✅	❌	❌	❌
AndroidLab	138	9	8.5	✅	✅	❌	❌	❌
LearnGUI (Ours)	2,353	73	13.2	✅	✅	✅	✅	✅

Note: # Inst. (number of instructions), # Apps (number of applications), # Step (average steps per task), Env. (supports environment interactions), HL (has high-level instructions), LL (has low-level instructions), GT (provides ground truth trajectories), FS (supports few-shot learning).

📄 License

This dataset is licensed under Apache License 2.0.