WebGen-Agent

WebGen-Agent is an advanced website generation agent designed to autonomously create websites from natural language instructions. This repository contains both inference code for running the agent and training code for the Step-GRPO with Screenshot and GUI-agent Feedback approach. It is an official implementation of the paper WebGen-Agent: Enhancing Interactive Website Generation with Multi-Level Feedback and Step-Level Reinforcement Learning

Project Overview

WebGen-Agent combines state-of-the-art language models with specialized training techniques to create a powerful website generation tool. The agent can understand natural language instructions specifying appearance and functional requirements, iteratively generate website codebases, and refine them using visual and functional feedback.

Resources

Links to the data and model parameters are as follows:

Repository Structure

WebGen-Agent/
├── src/                    # Inference code for the WebGen-Agent workflow
│   ├── agent/              # Core agent implementation
│   ├── prompts/            # Prompt templates
│   ├── scripts/            # Utility scripts
│   ├── utils/              # Helper functions
│   ├── webvoyager/         # GUI agent integration
│   ├── infer_batch.py      # Batch inference script
│   └── infer_single.py     # Single inference script
├── requirements.txt        # Python dependencies for inference
├── data/
│   └── webgen-bench/       # WebGen-Bench dataset
└── WebGen-RL/              # Training code (Step-GRPO with Screenshot and GUI-agent Feedback)
    ├── src/
    │   ├── sft/            # Supervised Fine-Tuning code
    │   └── runs/           # GRPO training scripts
    └── ...

Installation

1. Clone the repository:

   git clone https://github.com/your-username/WebGen-Agent.git
   cd WebGen-Agent
   

2. Create a conda environment:

   conda create -n env/webgen-agent python=3.10
   conda activate env/webgen-agent
   

3. Install the required dependencies:

   pip install -r requirements.txt
   

4. Node.js and google-chrome are also necessary for the project. You can install them by running:

   bash src/scripts/install_node.sh
   bash src/scripts/install_chrome.sh
   

Running Inference

Before running inference, you should rename .env.template to .env and set the base urls and api keys for the agent-engine LLM and feedback VLM. They can be obtained from any openai-compatible providers such as openrouter, modelscope, bailian, and llmprovider.

You can also deploy open-source VLMs and LLMs by running src/scripts/deploy_qwenvl_32b.sh and src/scripts/deploy.sh. Scripts for single inference and batch inference can be found at src/scripts/infer_single.sh and src/scripts/infer_batch.sh.

Single Inference

python src/infer_single.py \
    --model deepseek-chat \
    --vlm_model Qwen/Qwen2.5-VL-32B-Instruct \
    --instruction "Please implement a wheel of fortune website." \
    --workspace-dir workspaces_root/test \
    --log-dir service_logs/test \
    --max-iter 20 \
    --overwrite \
    --error-limit 5

Batch Inference

python src/infer_batch.py \
    --model ${model_name} \
    --vlm_model Qwen/Qwen2.5-VL-32B-Instruct \
    --fb_model ${fb_model_name} \
    --data-path data/webgen-bench/test.jsonl \
    --workspace-root workspaces_root \
    --log-root service_logs \
    --max-iter 20 \
    --num-workers 4 \
    --eval-tag select_best \
    --error-limit 5 \
    --max-tokens -1 \
    --max-completion-tokens -1 \
    --temperature 0.5

WebGen-Bench Evaluation

The WebGen-Bench evaluation script is in run_ui_eval_with_answer.sh. For details of the WebGen-Bench dataset, see WebGen-Bench.

Step-GRPO with Screenshot and GUI-agent Feedback (WebGen-RL)

For Step-GRPO with Screenshot and GUI-agent Feedback training, please refer to the WebGen-RL README.

The WebGen-RL component includes:

• Supervised Fine-Tuning (SFT) code in WebGen-RL/src/sft/
• GRPO training scripts in WebGen-RL/src/runs/

How WebGen-Agent Works

WebGen-Agent follows an iterative, multi-step paradigm for website generation:

1. Code Generation: The agent generates code to create or edit website files based on natural language instructions
2. Code Execution: Dependencies are installed and the website service is started
3. Feedback Gathering:
   • A screenshot of the website is captured
   • A Visual Language Model (VLM) provides appearance feedback and scores
   • A GUI-agent tests the website functionality and provides functional feedback
4. Refinement: Based on the feedback, the agent continues to improve the website until it meets requirements

Key Features

• Iterative Refinement: Continuously improves website appearance and functionality
• Feedback Integration: Uses both visual and functional feedback for enhanced performance
• Backtracking Mechanism: Reverts to previous states when encountering persistent errors
• Best Step Selection: Selects the optimal version based on visual and functional scores

Related Works

• WebGen-Bench: A benchmark for evaluating LLM-based agents on generating interactive and functional websites from scratch. It contains 101 website-generation instructions, 647 test cases, and a training set of 6,667 instructions named WebGen-Instruct.
  GitHub: mnluzimu/WebGen-Bench
  Paper: WebGen-Bench: Evaluating LLMs on Generating Interactive and Functional Websites from Scratch

• FullStack-Agent: A unified agent system for full-stack web development, consisting of FullStack-Dev, FullStack-Learn, and FullStack-Bench. It targets production-level full-stack website generation by evaluating and improving frontend, backend, and database functionalities.
  GitHub: mnluzimu/FullStack-Agent
  Paper: FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation

Citation

If you find our project useful, please cite:

@misc{lu2026fullstackagentenhancingagenticfullstack,
      title={FullStack-Agent: Enhancing Agentic Full-Stack Web Coding via Development-Oriented Testing and Repository Back-Translation}, 
      author={Zimu Lu and Houxing Ren and Yunqiao Yang and Ke Wang and Zhuofan Zong and Mingjie Zhan and Hongsheng Li},
      year={2026},
      eprint={2602.03798},
      archivePrefix={arXiv},
      primaryClass={cs.SE},
      url={https://arxiv.org/abs/2602.03798}, 
}
@misc{lu2025webgenagentenhancinginteractivewebsite,
      title={WebGen-Agent: Enhancing Interactive Website Generation with Multi-Level Feedback and Step-Level Reinforcement Learning}, 
      author={Zimu Lu and Houxing Ren and Yunqiao Yang and Ke Wang and Zhuofan Zong and Junting Pan and Mingjie Zhan and Hongsheng Li},
      year={2025},
      eprint={2509.22644},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.22644}, 
}

@misc{lu2025webgenbenchevaluatingllmsgenerating,
      title={WebGen-Bench: Evaluating LLMs on Generating Interactive and Functional Websites from Scratch}, 
      author={Zimu Lu and Yunqiao Yang and Houxing Ren and Haotian Hou and Han Xiao and Ke Wang and Weikang Shi and Aojun Zhou and Mingjie Zhan and Hongsheng Li},
      year={2025},
      eprint={2505.03733},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.03733}, 
}