WebDreamer is a planning framework that enables efficient and effective planning for real-world web agent tasks. Check our paper for more details. This work is a collaboration between OSUNLP and Orby AI.

Repository: https://github.com/OSU-NLP-Group/WebDreamer
Paper: https://arxiv.org/abs/2411.06559
Point of Contact: Kai Zhang

Models

Dreamer-7B:

Data:

Dreamer Training Data

root
 |-- prompt: string
 |-- image: binary
 |-- response: string
 |-- action: string

Results

Strong performance on VisualWebArena and Mind2Web-live

Benchmark	Method	Success Rate
VisualWebArena	GPT-4o + Reactive	17.6%
	GPT-4o + Tree Search	26.2%
	GPT-4o + WebDreamer	23.6% (↑34.1%)
Online-Mind2Web	GPT-4o + Reactive	26.0%
	GPT-4o + WebDreamer	37.0% (↑42.3%)
Mind2Web-live	GPT-4o + Reactive	20.2%
	GPT-4o + WebDreamer	25.0% (↑23.8%)

Compared to the reactive baselines, WebDreamer significantly improves performance by 34.1%, 42.3%, and 23.8% on VisualWebArena, Online-Mind2Web, and Mind2Web-live, respectively.

Better efficiency than tree search with true interactions

WebDreamer effectively explores the search space through simulations, which largely reduces the reliance on real-world interactions while maintaining robust performance.

Inference

vLLM server

vllm serve osunlp/Dreamer-7B --api-key token-abc123 --dtype float16

python -m vllm.entrypoints.openai.api_server --served-model-name osunlp/Dreamer-7B --model osunlp/Dreamer-7B --dtype float16

You can find more instruction about training and inference in Qwen2-VL's Official Repo.

Prompt

Actually our model is quite robust to textual prompt so feel free to try various prompts which we didn't heavily explore.

python

def format_openai_template(description: str, base64_image):
    return [
        {
            "role": "user",
            "content": [
                {
                    "type": "image_url",
                    "image_url": {"url": f"data:image/jpeg;base64,{base64_image}"},
                },
                {
                    "type": "text",
                    "text": f"""
  Below is current screenshot. Please describe what you would see after a {action_description}"""
                },
            ],
        },
    ]


messages = format_openai_template(description, base64_image)

completion = await client.chat.completions.create(
    model=args.model_path,
    messages=messages,
    temperature=1.0
)

Citation Information

If you find this work useful, please consider citing our papers:

@article{Gu2024WebDreamer,
  author    = {Yu Gu and Kai Zhang and Yuting Ning and Boyuan Zheng and Boyu Gou and Tianci Xue and Cheng Chang and Sanjari Srivastava and Yanan Xie and Peng Qi and Huan Sun and Yu Su},
  title     = {Is Your LLM Secretly a World Model of the Internet? Model-Based Planning for Web Agents},
  journal   = {CoRR},
  volume    = {abs/2411.06559},
  year      = {2024},
  url       = {https://arxiv.org/abs/2411.06559},
  eprinttype= {arXiv},
  eprint    = {2411.06559},
}