WebWorld - A series of large-scale web world models open-sourced by Alibaba's Qwen team.

What is WebWorld?

WebWorld is a series of large-scale web world models open-sourced by Alibaba's Qwen team. Trained on the Qwen3 platform, it includes 8B, 14B, and 32B versions. By predicting web page state transitions in a simulated browser environment, the model provides high-quality training data and an inference environment for Web Agents, avoiding network latency, rate limitations, and security risks associated with real-world web page training. The model supports multiple state representation formats, including A11y Tree, HTML, XML, Markdown, and natural language, and can achieve long-term consistent simulations of 30+ steps, possessing explicit chained reasoning capabilities.

WebWorld's main functions

• Webpage status predictionGiven the current page state (A11y Tree/HTML/XML, etc.) and Agent actions, predict the complete page state at the next moment and simulate real browser behavior.
• Long-term multi-round simulationSupports continuous interactive simulation of more than 30 steps, maintaining state consistency, and is suitable for complex multi-step web page tasks.
• Multi-format status representationIt natively supports A11y Tree and can also handle HTML, XML, Markdown, and natural language descriptions, enhancing the model's generalization ability.
• Reasoning ability activatedThrough a two-stage training program, large-scale dynamic knowledge from web pages is first injected, and then explicit causal reasoning is activated using a small amount of CoT data.
• Cross-domain generalizationIt demonstrates excellent portability in code environments, GUI desktops, and game scenarios.
• Trajectory data synthesisIt can be used as a data synthesizer to generate large-scale, high-quality training trajectories for downstream agents, thereby improving performance in real-world tasks.

WebWorld's Technical Principles

• Regressive browser simulator modelingWebWorld formalizes the browser environment as an autoregressive sequence generation task, learning conditional probability distributions based on causal language models. Pθ (st+1 |I,ht ) In the given task instructions I and interaction history ht =(s0 ,a0 ,…,st ,at ) In the case of predicting the action to be performed at Next page status st+1 End-to-end training is performed on complete trajectory data using maximum likelihood estimation.
• Three-tiered data collection pipelineTo overcome the data bottlenecks of closed environments, WebWorld constructed a scalable three-layer collection strategy: the first layer, "random crawling," performs random actions on websites corresponding to the pre-training corpus, obtaining 43.3% of the breadth data; the second layer, "autonomous exploration," deploys an LLM Agent to generate targets and explore websites, producing 20.4% of real long-term trajectories; the third layer, "task-oriented execution," synthesizes diverse variants based on seed tasks and executes them by the Agent, obtaining 16.1% of high-quality task trajectories. The three layers combined yield over 1.06 million real open webpage interaction data.
• A11y Tree Master State Representation and Multiformat EnhancementThe model uses the A11y Tree extracted by Playwright as the main state representation because it has the universality across web pages and GUIs, high information density and LLM-friendly structure. At the same time, the trajectory is extended to multiple formats such as HTML, XML, and Markdown through post-transformation, natural language page description is introduced, and a five-dimensional instruction fine-tuning dataset is constructed to avoid overfitting the model to a single representation and prevent catastrophic forgetting.
• Two-layer data filtering and quality controlData cleaning employs a two-tiered mechanism combining rule-based heuristics and LLM scoring: First, website accessibility is verified and sensitive keywords are filtered using scripts, retaining only 15.7% of the original URLs. Then, LLM scores the data across four dimensions: accessibility, content suitability, interactivity, and engineering quality, eliminating low-scoring sites. At the trajectory level, invalid transitions without state changes are further pruned, and excessively long samples exceeding 30 rounds or 30K tokens are discarded, without introducing any inductive bias specific to the model throughout the process.
• Two-stage course training strategy
• The training follows a curriculum design of "injecting knowledge first, then activating reasoning": the first stage involves large-scale dynamic modeling on 1.06 million trajectories, enabling the model to master a wide range of webpage state transition patterns; the second stage uses only 1,000 synthetic CoT samples for fine-tuning, requiring the model to output explicit analysis of page structure, user intent, and state changes before prediction, thereby externalizing implicit reasoning ability into an interpretable chain-like thinking pattern.
• WebWorld-Bench Multidimensional Evaluation System
• To comprehensively measure the quality of the simulation, the team constructed an internal benchmark that includes nine evaluation dimensions: the factual score uses LLM to judge point by point whether the predicted state correctly reflects the functional causal effect of the action; the Web Turing score uses adversarial pairwise comparison to test the indistinguishability between the simulated state and the real web page. The combination of the two measures quantifies the world model's capabilities from both objective correctness and subjective realism.

How to use WebWorld

• Environmental preparationExecute after cloning the repository pip install -r requirements.txt And decompress the data packet.
• Model loadingLoading models via HuggingFace (e.g.) Qwen/WebWorld-8B),use AutoModelForCausalLM Cooperate trust_remote_code=True initialization.
• Single-step prediction: Construct a dialog that includes system prompts (declared as a web world model) and user messages (initial page state + action), and call model.generate Predict the state of the next page.
• Multi-round simulationThe first round provides the initial state and the first action; subsequent rounds use fixed continuation prompts. CONTINUE_PROMPTIt takes the previous predicted state as history, inputs new actions to continue generating, and can cycle for 30+ rounds.
• Agent trainingWe use WebWorld to synthesize trajectory data and generate diverse task trajectories through an abstract-and-instantiate strategy. This allows us to fine-tune the base model to improve the performance of downstream agent benchmarks.
• BenchmarkingUse WebWorld-Bench for internal evaluation, or verify the agent training effect through external benchmarks such as MiniWob++ and WebArena.

WebWorld's core advantages

• Leading in scaleTraining is based on the trajectories of over 1 million real open web pages, with a data volume 100 times that of previous work, covering diverse fields such as e-commerce, social media, and news.
• Open sourceThe model weights and training data (WebWorldData) are both open source under the Apache 2.0 license, providing a complete and reproducible technical roadmap.
• A comprehensive evaluation systemWe have developed our own WebWorld-Bench internal benchmark, which comprehensively evaluates simulation quality from two dimensions: factality and Web Turing, and nine sub-dimensions.
• Outstanding training efficiencyOnly 1000 CoT samples are needed to activate strong reasoning ability, proving the efficiency of large-scale dynamic pre-training for reasoning activation.
• Significant Agent Training GainsQwen3-8B, after being fine-tuned with WebWorld synthetic data, showed a 10.9% improvement on WebArena, and version 14B was close to the level of GPT-4o.

WebWorld's project address

• GitHub repositoryhttps://github.com/QwenLM/WebWorld
• HuggingFace model libraryhttps://huggingface.co/datasets/Qwen/WebWorldData
• arXiv technical paper: https://arxiv.org/pdf/2602.14721

WebWorld's Competitive Comparison

WebWorld Application Scenarios

• Web Agent Training and EvaluationIt provides a low-cost, high-throughput simulation training environment for automated web page operation agents, replacing expensive real web page interactions.
• Data augmentation and synthesisGenerate large-scale synthetic trajectories for web tasks lacking labeled data, for supervised fine-tuning or reinforcement learning.
• Planning and Search during ReasoningDuring Agent execution, it acts as a "world model" to perform action look-ahead simulations and assist in selecting the optimal action sequence.
• Cross-domain world model researchIts technological paradigm can be transferred to a wider range of digital world modeling tasks, such as GUI automation, code environment simulation, and game state prediction.
• Browser automated testingSimulates user interaction paths for webpage function testing, compatibility verification, and user experience preview.