Dataset card for DAC-Bench

• Repository: https://github.com/melonthrower/DAC-Bench/tree/master
• Paper: https://openreview.net/forum?id=0u75PAejgs

💡 Introduction

DAC-Becch is a benchmark designed to evaluate mobile GUI agents on long-horizon planning and complex decision-making tasks. While existing benchmarks for mobile GUI agents primarily focus on short, linear workflows, they offer limited insight into an agent's capabilities for conditional reasoning and long-term planning. DAC-Bench addresses this gap by introducing graph-structured, decision-aware tasks that systematically evaluate and advance the capabilities of modern agents in realistic scenarios.

📖 Data structure

After unzipping screenshots.zip and annotations.zip in the DAC-Becch/data directory, you will get the following directory structure

DAC-Becch
├── data
│   ├── annotations ## The files obtained after unzipping "annotations.zip"
│   │   └── *.json ## JSON files that store instruction annotation data
│   ├── screenshots 
│   │   └── data_* ## The files obtained after unzipping "screenshots.zip"
│   │        └── *.png

In each annotated json file, the structure and meaning of each field are as follows:

• episode_id (str): The id identifying the episode
• w, h (int): The width and height of the screenshot in the episode
• device_info (dict): A dictionary containing device information
• instruction (str): The natural language instruction of the episode
• category (str): The type of the episode, it belongs to one of the types: single, and(Conjunctive), chain(Sequential), selection(Conditional), nested(Hierarchical).
• APP (list): The list of applications involved in the episode
• steps (list): A list of each step consisting of screenshots path, step-instructions, and actions
• atomic_instructions (dict): A dictionary containing each atomic actions
• seq (dict): A dictionary containing each branch sequence, only the selection and nested episodes include this item
• seq_step: the step id of branch steps, only the selection and nested episodes include this item
• instruction_cn (str): the chinese version of instruction

🎮 action space

the action and its parameters will be one of the format listed below.

• Click[int, int, int, int]: Click within the [x, y, m, n] area on the screen. The origin of the coordinates is at the top-left corner, with the width direction as the x-axis and the height direction as the y-axis. (x, y) and (m, n) are the coordinates of the top-left corner and the bottom-right corner of the area, respectively. The area refers to the range where the button can respond to clicks.
• long_press[int, int, int, int]: Similar to click, press and hold a certain position on the screen
• scroll[int, int , int, int]：Scroll operation, where the numbers correspond to the coordinates of the specified starting and ending points.
• type(str): Input text, the parameter is the text to be input.
• navigate_home: Return to the home page.
• navigate_back: Return to the last page.
• complete: the task has been completed.
• impossible: the task cannot been completed.
• wait: wait for few seconds.

Citation

If you find DAC-Becch useful in your research, please consider citing our paper:

@inproceedings{
anonymous2026dacbench,
title={{DAC}-Bench: A Decision-Aware Benchmark for Compositional Mobile {GUI} Tasks},
author={Anonymous},
booktitle={The 64th Annual Meeting of the Association for Computational Linguistics},
year={2026},
url={https://openreview.net/forum?id=0u75PAejgs}
}