You can deploy this model and start earning money today!

senapati484/
WebWorld-32B

Name: WebWorld-32B
Rating: 2.5 (6 reviews)

synthetic-trajectories

instruction-tuning

conversational

Views

No views yet

WebWorld 🌐

License

GitHub

Dataset

14B

32B

📚 Introduction

WebWorld is a large-scale open-web world model series for training and evaluating web agents. It is trained on 1M+ real-world web interaction trajectories via a scalable hierarchical data pipeline, supporting:

Long-horizon simulation (30+ steps)
Multi-format state representations: A11y Tree, HTML, XML, Markdown, and natural language
CoT-activated reasoning for transition prediction
Cross-domain generalization to code, GUI, and game environments

Agents trained on WebWorld-synthesized trajectories achieve +9.9% on MiniWob++ and +10.9% on WebArena. When used for inference-time lookahead search, WebWorld outperforms GPT-5 as a world model.

🎯 Model Series

Model	Base Model	HuggingFace Link	ModelScope Link
WebWorld-8B	Qwen3-8B	🤗 HuggingFace	🤖 ModelScope
WebWorld-14B	Qwen3-14B	🤗 HuggingFace	🤖 ModelScope
WebWorld-32B	Qwen3-32B	🤗 HuggingFace	🤖 ModelScope

WebWorldData: Huggingface: Qwen/WebWorldData, ModelScope: Qwen/WebWorldData

💡 Recommendation: Use 8B for fast simulation and data synthesis; use 14B/32B for higher-fidelity simulation and better long-horizon robustness. For best results in a specific environment, we recommend task-specific fine-tuning on in-domain trajectories.

🛠️ Requirements

transformers (recommended: latest version)
torch
Optional: accelerate, vllm for efficient serving

🚀 Quick Start

Key Notes:

WebWorld predicts the next page state given the current state and an action.
It strictly preserves the input/output format (A11y / HTML / XML / Markdown / NL).
Supports multi-turn trajectory simulation up to 30+ steps.

Single-Step Prediction

💻 Click to expand code

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Qwen/WebWorld-8B"  # or WebWorld-14B, WebWorld-32B
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).eval()

system_prompt = (
    "You are a web world model. I will provide you with an initial page state "
    "and a sequence of actions. For each action, predict the resulting page state.\n"
    "Strictly maintain the original format. Output only the full page state "
    "without explanations, code, or truncation."
)

current_state = """RootWebArea 'Global Start - Your Daily Portal', focused
\t[1] banner 'Top Header', visible
\t\t[2] link 'Set as Homepage', clickable, visible
\t\t[3] link 'Feedback', clickable, visible
\t\t[5] region 'Weather Widget', visible
\t\t\tStaticText 'New York, USA'
\t\t\t[6] image 'Sunny', visible
\t\t\tStaticText '24°C'
\t\t[8] link 'Sign In', clickable, visible
\t[10] region 'Search Area', visible
\t\t[11] image 'Global Start Logo', visible
\t\tStaticText 'Search the entire web'
\t\t[12] tablist 'Search Engine Selector', orientation='horizontal'
\t\t\t[13] tab 'Google', selected=True, clickable
\t\t\t[14] tab 'Bing', selected=False, clickable
\t\t\t[15] tab 'DuckDuckGo', selected=False, clickable
\t\t[18] combobox 'Web Search', clickable, visible, autocomplete='both', expanded=False
\t\t\t[19] textbox 'Type keywords or URL...', clickable, visible, editable, value=''
\t\t[20] button 'Search', clickable, visible
\t[30] navigation 'Category Bar', visible
\t\t[31] link 'Home', clickable, selected=True
\t\t[32] link 'News', clickable
\t\t[33] link 'Video', clickable
\t\t[34] link 'Shopping', clickable
\t\t[35] link 'Social', clickable
\t[50] main 'Site Directory', visible
\t\t[51] region 'Top Recommended', visible
\t\t\t[52] heading 'Most Popular', visible
\t\t\t[53] list 'Top Sites Grid', visible
\t\t\t\t[54] link 'Facebook', clickable
\t\t\t\t[56] link 'YouTube', clickable
\t\t\t\t[58] link 'Amazon', clickable
\t\t\t\t[60] link 'Twitter / X', clickable
\t\t\t\t[62] link 'Instagram', clickable
\t\t\t\t[64] link 'Wikipedia', clickable
\t\t\t\t[66] link 'Netflix', clickable
\t\t\t\t[68] link 'LinkedIn', clickable
\t\t[80] region 'News & Media', visible
\t\t\t[81] heading 'Latest News', visible
\t\t\t[82] link 'CNN', clickable
\t\t\t[83] link 'BBC', clickable
\t\t\t[84] link 'The Verge', clickable
\t\t[90] region 'Shopping', visible
\t\t\t[91] heading 'E-Commerce', visible
\t\t\t[92] link 'eBay', clickable
\t\t\t[93] link 'Walmart', clickable
\t\t\t[94] link 'Best Buy', clickable
\t[200] complementary 'Ads', visible
\t\t[201] image 'Ad: Travel to Japan'
\t\t[202] link 'Book Now', clickable
\t[300] contentinfo 'Footer', visible
\t\tStaticText '© 2026 Global Start Inc.'"""

user_message = (
    f"Initial Page State:\n{current_state}\n\n"
    f"First Action: 'click([32])'\n\n"
    f"Next Page State:"
)

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_message},
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=4096,
        do_sample=False,
    )

response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(response)

Multi-Turn Simulation

The first turn provides the initial state and first action. Each subsequent turn uses a fixed continuation prompt:

💻 Click to expand code

python

CONTINUE_PROMPT = (
    "Continue the trajectory. Given the previous state, "
    "predict the next page state after this action.\n\n"
    "Action: '{action}'\n\nNext Page State:"
)

# Turn 1
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": f"Initial Page State:\n{state_0}\n\nFirst Action: '{action_0}'\n\nNext Page State:"},
]
state_1 = generate(messages)  # your generate function

# Turn 2
messages.append({"role": "assistant", "content": state_1})
messages.append({"role": "user", "content": CONTINUE_PROMPT.format(action=action_1)})
state_2 = generate(messages)

# Turn 3, 4, ... up to 30+ turns: repeat the same pattern
messages.append({"role": "assistant", "content": state_2})
messages.append({"role": "user", "content": CONTINUE_PROMPT.format(action=action_2)})
state_3 = generate(messages)

🎮 Action Space

WebWorld supports a unified action space as Python-style function calls:

Category	Action	Description
Element	`click(bid, button, modifiers)`	Click a DOM element by its ID
	`fill(bid, text, press_enter)`	Type text into an input field
	`select_option(bid, options)`	Select from a dropdown / combobox
	`hover(bid)`	Hover over an element
Mouse	`mouse_move(x, y)`	Move cursor to coordinates
	`mouse_click(x, y, button)`	Click at coordinates
	`mouse_down(x, y)` / `mouse_up(x, y)`	Press / release (drag-and-drop)
Keyboard	`keyboard_press(key)`	Press a key (e.g., `Enter`, `Tab`)
	`keyboard_type(text)`	Type a string sequentially
Browser	`scroll(dx, dy)`	Scroll the viewport
	`goto(url)`	Navigate to a URL
	`go_back()` / `go_forward()`	Browser history navigation
	`tab_new()` / `tab_close()` / `tab_focus(index)`	Manage browser tabs
Meta	`send_msg_to_user(text)`	Send a message to the user
	`noop(wait_ms)`	Wait for a duration
	`infeasible(reason)`	Declare the task impossible

📊 Performance

Intrinsic Evaluation (WebWorld-Bench)

WebWorld-Bench evaluates models using Factuality Score (functional correctness) and Web Turing Score (perceptual realism) across nine dimensions:

Model	Avg Factuality	Avg Turing
GPT-4o	59.5	35.4
Claude-Opus-4.1	71.3	47.4
Gemini-3-Pro	70.3	43.2
Qwen3-8B (base)	26.9	17.4
WebWorld-8B	70.1	42.2
WebWorld-14B	70.7	44.7
WebWorld-32B	71.0	45.6

Extrinsic Evaluation (Agent Training)

Model	MiniWob++ SR	WebArena SR
GPT-4o	64.3%	26.6%
Qwen3-8B (base)	49.4%	9.8%
Qwen3-8B + WebWorld	59.3% (+9.9%)	20.7% (+10.9%)
Qwen3-14B (base)	54.9%	15.1%
Qwen3-14B + WebWorld	63.2% (+8.3%)	24.3% (+9.2%)

Cross-Domain Generalization

Environment	Qwen3-8B	WebWorld-8B	Gain
API Services	0.088	0.299	+0.211
Code	0.147	0.396	+0.249
Game	0.253	0.473	+0.220
GUI Desktop	0.322	0.705	+0.383

⚠️ Limitations

Sycophancy / optimism bias: the model may generate outcomes that are overly favorable to the agent's intended action.
Content generation fidelity: long-form, high-precision content (e.g., scientific articles) is not the primary target.
Text-only: WebWorld does not simulate visual / pixel-level rendering.

📝 Citation

bibtex

@misc{xiao2026webworldlargescaleworldmodel,
      title={WebWorld: A Large-Scale World Model for Web Agent Training}, 
      author={Zikai Xiao and Jianhong Tu and Chuhang Zou and Yuxin Zuo and Zhi Li and Peng Wang and Bowen Yu and Fei Huang and Junyang Lin and Zuozhu Liu},
      year={2026},
      eprint={2602.14721},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.14721}, 
}

senapati484/WebWorld-32B