BrowserAgent

An agent that can interact with browser to complete tasks

Resources

• Project Page: https://tiger-ai-lab.github.io/BrowserAgent/
• Paper (arXiv): https://arxiv.org/abs/2510.10666
• Models (HF):
  • https://huggingface.co/TIGER-Lab/BrowserAgent-SFT
  • https://huggingface.co/TIGER-Lab/BrowserAgent-RFT
• Dataset (HF): https://huggingface.co/datasets/TIGER-Lab/BrowserAgent-Data
• Seed Dataset (HF): https://huggingface.co/datasets/TIGER-Lab/BrowserAgent-SeedData

Installation

1. Clone this repository and navigate to the folder:

git clone https://github.com/TIGER-AI-Lab/BrowserAgent.git
cd BrowserAgent

Install the inference package:

conda create -n browseragent python=3.10.12
conda activate browseragent
pip install -e .
cd verl-tool
pip install -e .
pip install -e verl
pip install vllm==0.8.4
pip install --upgrade opentelemetry-api opentelemetry-sdk
pip install flash-attn --no-build-isolation
pip install -e ".[acecoder,torl]"
pip uninstall uvloop
playwright install

2. Data preparation

To conduct Web-Information-Seeking Task, download 📊 BrowserAgent-SeedData and place it in the data folder, the final structure should look like this:

benchmark
-- | nq
-- | hotpot
-- | 2wiki
-- | popqa
-- | musique
-- | bamboogle

3. Deploy the wiki webpage:

We adopt the WikiPedia from WebArena.

To deploy the Wiki webpage locally, run the following Docker command and open http://localhost:22015.

docker run -d --name=wikipedia --volume=<your-path-to-downloaded-folder>/:/data -p 22015:80 ghcr.io/kiwix/kiwix-serve:3.3.0 wikipedia_en_all_maxi_2022-05.zim

For public deployment, see our live example at tigerai.ca/wiki/.../Landing and apply the Nginx template in this repo, updating proxy_pass to match your Docker port.

location = /search {
    return 301 /wiki/search?$args;
}

location = /random {
    return 301 /wiki/random?$args;
}

location ~ ^/wikipedia_en_all_maxi_2022-05 {
    return 301 /wiki$request_uri;
}


location /wiki/ {
    proxy_pass http://localhost:22015/;
    proxy_buffering off;
    proxy_http_version 1.1;
    proxy_set_header Upgrade $http_upgrade;
    proxy_set_header Connection "upgrade";
    proxy_set_header Host $host;

    sub_filter_once off;

    sub_filter '<a href="/' '<a href="/wiki/';
    sub_filter '<link href="/' '<link href="/wiki/';
    sub_filter '<script src="/' '<script src="/wiki/';
    sub_filter '<img src="/' '<img src="/wiki/';

    sub_filter 'href="/search/' 'href="/wiki/search/';
    sub_filter 'href="/random/' 'href="/wiki/random/';
    sub_filter 'href="/wikipedia_en_all_maxi_2022-05/' 'href="/wiki/wikipedia_en_all_maxi_2022-05/';
    sub_filter 'action="/search' 'action="/wiki/search';
    sub_filter 'src="/wikipedia_en_all_maxi_2022-05/' 'src="/wiki/wikipedia_en_all_maxi_2022-05/';

4. SFT and RFT model:

Download 📊 BroswerAgent-SFT or 📊 BroswerAgent-RFT, and deploy using vllm.

Terminal 1:
conda activate browseragent
cd BrowserAgent
bash deploy_vllm.sh /path/to/your/model

5. Data generation and model evaluation:

(1) For SFT data generation

Terminal 2:
conda activate browseragent
cd BrowserAgent
bash verl-tool/examples/train/wikiRL/wikiRL_server.sh

Terminal 3:
conda activate browseragent
cd BrowserAgent
python data_generate.py /path/to/your/output_file /path/to/your/sft_data_path

Then you can run the following code to convert the generated data into the ms-swift training format.

conda activate browseragent
cd BrowserAgent
python judge_sft.py /path/to/your/sft_data_path /path/to/your/previous_step_output_file /path/to/your/output_file
python swift_switch.py /path/to/your/previous_step_output_file /path/to/your/output_file

(2) For RFT data generation

Terminal 2:
conda activate browseragent
cd BrowserAgent
bash verl-tool/examples/train/wikiRL/wikiRL_server.sh

Terminal 3:
conda activate browseragent
cd BrowserAgent
python data_generate_rft.py /path/to/your/output_file /path/to/your/rft_data_path

Then you can run the following code to convert the generated data into the ms-swift training format.

conda activate browseragent
cd BrowserAgent
python judge_rft.py /path/to/your/sft_data_path /path/to/your/previous_step_output_file /path/to/your/output_file
python swift_switch.py /path/to/your/previous_step_output_file /path/to/your/output_file

Alternatively, you can also get our SFT and RFT Dataset via https://huggingface.co/datasets/TIGER-Lab/BrowserAgent-Data

(3) For model evaluation

Terminal 2:
conda activate browseragent
cd BrowserAgent
bash verl-tool/examples/train/wikiRL/wikiRL_server.sh

Terminal 3:
conda activate browseragent
cd BrowserAgent
python run_model.py /path/to/your/benchmark_path

Then you can run the following code to calculate the rule-based accuracy.

conda activate browseragent
cd BrowserAgent
python val_answer.py /path/to/your/benchmark_path /path/to/your/previous_step_output_file

Then you can run the following code to calculate the model-based accuracy.

conda activate browseragent
cd BrowserAgent
python val_answer_model_based.py /path/to/your/benchmark_path /path/to/your/previous_step_output_file /path/to/your/output_file

Citation

If you find it useful for your research and applications, please cite related papers/blogs using this BibTeX:

@misc{yu2025browseragentbuildingwebagents,
      title={BrowserAgent: Building Web Agents with Human-Inspired Web Browsing Actions}, 
      author={Tao Yu and Zhengbo Zhang and Zhiheng Lyu and Junhao Gong and Hongzhu Yi and Xinming Wang and Yuxuan Zhou and Jiabing Yang and Ping Nie and Yan Huang and Wenhu Chen},
      year={2025},
      eprint={2510.10666},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.10666}, 
}