--- license_link: https://www.apache.org/licenses/LICENSE-2.0 license: apache-2.0 language: - en - zh pipeline_tag: text-generation library_name: transformers tags: - WebWorld - web-agent - world-model - simulator - browser - a11y - html - xml - markdown - long-horizon - long-context - synthetic-trajectories - instruction-tuning base_model_relation: finetune base_model: - Qwen/Qwen3-8B datasets: - Qwen/WebWorldData --- # WebWorld 🌐 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/LICENSE-2.0) [![GitHub](https://img.shields.io/badge/GitHub-WebWorld-4b32c3?logo=github)](https://github.com/QwenLM/WebWorld) [![Dataset](https://img.shields.io/badge/HF%20Dataset-WebWorldData-yellow?logo=huggingface)](https://huggingface.co/datasets/Qwen/WebWorldData) [![MS Dataset](https://img.shields.io/badge/ModelScope-Dataset-7B42BC)](https://modelscope.cn/datasets/Qwen/WebWorldData) [![8B](https://img.shields.io/badge/Model-8B-green?logo=huggingface)](https://huggingface.co/Qwen/WebWorld-8B) [![MS 8B](https://img.shields.io/badge/ModelScope-8B-7B42BC)](https://modelscope.cn/models/Qwen/WebWorld-8B) [![14B](https://img.shields.io/badge/Model-14B-green?logo=huggingface)](https://huggingface.co/Qwen/WebWorld-14B) [![MS 14B](https://img.shields.io/badge/ModelScope-14B-7B42BC)](https://modelscope.cn/models/Qwen/WebWorld-14B) [![32B](https://img.shields.io/badge/Model-32B-green?logo=huggingface)](https://huggingface.co/Qwen/WebWorld-32B) [![MS 32B](https://img.shields.io/badge/ModelScope-32B-7B42BC)](https://modelscope.cn/models/Qwen/WebWorld-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](https://huggingface.co/Qwen/Qwen3-8B) | [🤗 HuggingFace](https://huggingface.co/Qwen/WebWorld-8B) | [🤖 ModelScope](https://modelscope.cn/models/Qwen/WebWorld-8B) | | **WebWorld-14B** | [Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B) | [🤗 HuggingFace](https://huggingface.co/Qwen/WebWorld-14B) | [🤖 ModelScope](https://modelscope.cn/models/Qwen/WebWorld-14B) | | **WebWorld-32B** | [Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) | [🤗 HuggingFace](https://huggingface.co/Qwen/WebWorld-32B) | [🤖 ModelScope](https://modelscope.cn/models/Qwen/WebWorld-32B) | **WebWorldData**: [Huggingface: Qwen/WebWorldData](https://huggingface.co/datasets/Qwen/WebWorldData), [ModelScope: Qwen/WebWorldData](https://modelscope.cn/datasets/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}, }