WebSynthesis

World Model-Guided MCTS for Efficient WebUI-Trajectory Synthesis

WebSynthesis is a framework integrating world model learning and Monte Carlo Tree Search (MCTS), designed to significantly reduce the cost of online synthesis of high-quality Web UI trajectories. Through a two-stage curriculum, including UI fundamental understanding and UI behavior cloning, the policy agent acquires web navigation capabilities.

News

• [08/07/2025] 🎉 Our Paper has been accepted as main conference in ACL 2026.
• [08/07/2025] 📢 Our Paper is now available in Arxiv.
• [07/07/2025] 📢 We upload some of the training data and LoRA weights.
• [06/07/2025] 📢 We release the code of WebSynthesis.

Two-stage Curriculum Learning

Project Structure

The WebSynthesis project is organized as follows:

WebSynthesis/
├── config_files/              # Configuration files for different tasks (0.json, 1.json, ...)
├── data/                      # Generated data from MCTS runs (created during execution)
├── figure/                    # Images and figures used in the README and documentation
├── models/                    # Model-related code
│   ├── get_response.py        # Functions for getting responses from models
│   └── models.py              # Model definitions and implementations
├── utils/                     # Utility functions and helper modules
│   ├── new_obs_opt.py         # New observation optimization utilities
│   ├── obs_opt.py             # Observation optimization utilities
│   ├── prune_mcts.py          # MCTS pruning utilities
│   ├── query_llm.py           # LLM query utilities
│   ├── search_utils.py        # Search utilities
│   ├── text_utils.py          # Text processing utilities
│   ├── traj_utils.py          # Trajectory utilities
│   └── treeNode.py            # TreeNode implementation for MCTS
├── webMCTS/                   # WebMCTS core implementation
│   ├── base.py                # Base classes for MCTS
│   ├── mcts.py                # MCTS algorithm implementation
│   ├── prompt.py              # Prompt templates for MCTS
│   └── task.py                # Task definitions for MCTS
├── webmcts-ttraj/             # Traceable trajectories (created during execution)
├── webmcts-vtraj/             # Valuable trajectories (created during execution)
├── data/                      # Data generated by MCTS runs (created during execution)
├── fuzzy_match.json           # Fuzzy matching cache
├── merge.py                   # Script to merge and process MCTS data
├── README.md                  # Project documentation
├── run.py                     # Main script to run MCTS tasks
└── run.sh                     # Shell script to run MCTS on all config files

Key directories and files:

• config_files/: Contains JSON configuration files for different MCTS tasks (numbered 0.json, 1.json, etc.)
• webMCTS/: Core implementation of the WebMCTS algorithm
• utils/: Helper modules for various functions like observation optimization, LLM querying, and trajectory processing
• models/: Model-related code for the world model, reward model, and policy agents
• run.py: Main entry point for running MCTS tasks
• run.sh: Shell script to automatically run MCTS on all configuration files
• merge.py: Script to process and merge the generated MCTS data into valuable and traceable trajectories

Data Collection (MCTS)

1. Clone the GitHub Repository:

   git clone https://github.com/LucusFigoGao/WebSynthesis.git
   

2. Collection:

   cd WebMCTS
   bash run.sh          # run MCTS
   pyhton merge.py      # merge the data
   

Data Resources

UI Fundamental Understanding

Model Name	Base Model	Training Data	LoRA
TextUI-Cap-7B	Qwen2.5-Instruct-7B	TextUI-dense-caption-training-data	🤗 link
TextUI-Func-7B	Qwen2.5-Instruct-7B	TextUI-functionality-training-data	🤗 link
TextUI-Trans-7B	Qwen2.5-Instruct-7B	TextUI-state-transition-training-data	🤗 link

UI Behavior Cloning

Model Name	Base Model	Training Data	LoRA
WebSynthesis-7B	Qwen2.5-Instruct-7B	WebSynthesis-training-data	🤗 coming soon
OS-Genesis-TextUI-7B	Qwen2.5-Instruct-7B	OS-Genesis-training-data	🤗 coming soon

Collection of World Model State Transition

Model Name	Base Model	Training Data	LoRA
World Model 7B	Qwen2.5-Instruct-7B	world-model-training-data	🤗 coming soon

🙏 Many thanks to the following open-source projects for their raw data contributions:

• [ICLR'25] OS-Genesis-Web-Data
• [ACL'25] AgentTrek-Web-Data
• [ICLR'24] WebArena-Web-Data

Main Experiment

Ablation Studies

Scaling Analysis

Case Study on WebMCTS

Case Study on World Model

Citation 📖

🫶 If you are interested in our work or find this repository / our data helpful, please consider using the following citation format when referencing our paper:

@misc{gao2025websynthesisworldmodelguidedmctsefficient,
   title={WebSynthesis: World-Model-Guided MCTS for Efficient WebUI-Trajectory Synthesis}, 
   author={Yifei Gao and Junhong Ye and Jiaqi Wang and Jitao Sang},
   year={2025},
   eprint={2507.04370},
   archivePrefix={arXiv},
   primaryClass={cs.AI},
   url={https://arxiv.org/abs/2507.04370}, 
}