WebWorld: A Large-Scale World Model for Web Agent Training

Zikai Xiao; Jianhong Tu; Chuhang Zou; Yuxin Zuo; Zhi Li; Peng Wang; Bowen Yu; Fei Huang

cs.AIFeb 16, 2026

WebWorld: A Large-Scale World Model for Web Agent Training

Authors: Zikai Xiao, Jianhong Tu, Chuhang Zou, Yuxin Zuo, Zhi Li, Peng Wang, Bowen Yu, Fei Huang, +2 more

Organizations: Qwen Team, Alibaba Group · Zhejiang University

Abstract

Web agents require massive trajectories to generalize, yet real-world training is constrained by network latency, rate limits, and safety risks. We introduce WebWorld series, the first open-web simulator trained at scale. While existing simulators are restricted to closed environments with thousands of trajectories, WebWorld leverages a scalable data pipeline to train on 1M+ open-web interactions, supporting reasoning, multi-format data, and long-horizon simulations of 30+ steps. For intrinsic evaluation, we introduce WebWorld-Bench with dual metrics spanning nine dimensions, where WebWorld achieves simulation performance comparable to Gemini-3-Pro. For extrinsic evaluation, Qwen3-14B trained on WebWorld-synthesized trajectories improves by +9.2% on WebArena, reaching performance comparable to GPT-4o. WebWorld enables effective inference-time search, outperforming GPT-5 as a world model. Beyond web simulation, WebWorld exhibits cross-domain generalization to code, GUI, and game environments, providing a replicable recipe for world model construction.

Figures & tables

Figure 1: Overview of WebWorld. WebWorld is a large-scale world model for the open web, trained on over 1M real-world trajectories. It supports long-horizon, multi-format simulation, enabling agents trained with its data to achieve significant performance gains.

Model	Size	Open Access		Data Source		Formats	Model Capabilities
Model	Size	Model	Data	Type	Scale	Formats	Open-Web	Reason	Long-Horizon
API-Based Prompting Methods
UI-Simulator ( Wang et al., 2025 )	GPT-4o-mini	-	-	Prompting	-	A11y	$\checkmark$	$\checkmark$	$\checkmark$
Simia ( Li et al., 2025b )	o4-mini	-	-	Prompting	-	Text	$\checkmark$	$\checkmark$	$\checkmark$
Trained World Models
DreamGym ( Chen et al., 2025 )	8B	$\times$	$\times$	Benchmarks	$\sim$ 14K	Text	$\times$	$\checkmark$	Full Traj.

Table 1: Comparison of World Models for Web Agents. We categorize existing approaches into API-based prompting methods and trained world models . Unlike proprietary API-based simulators or prior open-weights models restricted to closed web environments, WebWorld is a generalist world model trained on large-scale (1M+) real-world trajectories, supporting internal reasoning , long-horizon consistency, and open-web generalization.

Figure 2: WebWorld Example. Left: Agent. Right: WebWorld.

Domain	Task	Input	Output	Description
Web	1. Multi-Format Simulator	$\mathcal{S}_{t}+\mathcal{A}_{t}$	$\mathcal{S}_{t+1}$	Predict next state in A11y, HTML, XML, or Markdown.
	2. Web Generation	$\mathcal{T}_{intent}$	$\mathcal{S}_{page}$	Generate a full webpage structure from user requirements.
	3. Descriptive Simulator	$\mathcal{S}_{t}+\mathcal{A}_{t}$	$\mathcal{T}_{desc}$	Interpret visual changes and output a text summary.
General	4. General World Model	$\mathcal{T}_{t}+\mathcal{A}_{t}$	$\mathcal{T}_{t+1}$	Simulate state transitions purely in natural language.
General	5. General Chat	$\mathcal{T}_{query}$	$\mathcal{T}_{response}$	Standard dialogue to preserve conversational capabilities.

Table 2: Data Enrichment Tasks. Overview of the five auxiliary tasks across Web and General domains. Notation:

\mathcal{S}

represents structured web states (A11y, HTML, XML, Markdown);

\mathcal{T}

denotes natural language text;

\mathcal{A}

indicates agent actions.

Panel of Figure 3 — Figure 3: Statistics of the WebWorld Dataset. (a) Diverse coverage across domains like Lifestyle, Tech, and Education. (b) Token distribution showing the model’s exposure to varying context lengths. (c) Interaction turns distribution confirming the inclusion of long-horizon tasks (up to 30+ steps).

Model	Long-Horizon		Base Sem.		Fine-grain		Multi-tab		Multi-format Robustness								Web2NAL		Average
	(Consistency)		(Semantics)		(Sensitivity)		(Multi-page)		XML		HTML		Markdown		Playwright		(Nat. Lang.)		(All)
	Fact.	Tur.	Fact.	Tur.	Fact.	Tur.	Fact.	Tur.	Fact.	Tur.	Fact.	Tur.	Fact.	Tur.	Fact.	Tur.	Fact.	Tur.	Fact.	Tur.
Proprietary LLMs
GPT-4o	69.2	25.0	55.3	43.0	81.0	30.0	69.9	32.0	63.5	25.0	47.3	22.0	64.1	36.0	51.3	44.0	34.2	62.0	59.5	35.4
Claude-Sonnet-4.5	78.1	37.0	60.4	42.0	81.1	34.0	68.8	31.0	60.4	31.0	42.4	20.0	63.3	34.0	51.7	36.0	31.4	61.0	59.7	36.2
Claude-Opus-4.1	82.9	34.0	72.1	53.0	79.4	35.0	79.3	43.0	76.4	42.0	68.8	47.0	77.5	54.0	75.5	56.0	29.4	63.0	71.3	47.4

Table 3: Performance comparison across nine evaluation dimensions. Each dimension reports both Factuality Score (Fact.) and Web Turing Score (Tur.) in paired columns. Models are categorized into Proprietary and Open-source. The best results in each metric are bolded . All scores are normalized to [0, 1], with higher values indicating better performance.

Model	GPT-4o		Claude-Opus-4.1		Average
Model	Fact.	Turing	Fact.	Turing	Fact.	Turing
Proprietary
GPT-4o	59.5	35.4	51.6	21.0	55.6	28.2
Claude-Sonnet-4.5	59.7	36.2	59.9	31.9	59.8	34.1
Gemini-3-Pro	70.3	43.2	72.8	36.5	71.6	39.9
Open-weights

Table 4: Judge Consistency. We evaluate the consistency of model rankings using two state-of-the-art judges: GPT-4o and Claude-Opus-4.1 . Despite variations in absolute strictness, the relative ranking remains robust across different evaluators.

Model	MiniWob++			WebArena
	SR $\uparrow$	Std	Steps $\downarrow$	SR $\uparrow$	Std	Steps $\downarrow$	Domain Breakdown (SR %)
	(%)	( $\pm$ )	(Avg)	(%)	( $\pm$ )	(Avg)	E-Comm	GitLab	Reddit	Others
GPT-4o	64.3	0.019	4.12	26.6	0.016	11.96	26.8	27.5	24.5	25.3
Qwen3-8B (Base)	49.4	0.020	4.88	9.8	0.018	15.24	17.1	9.4	5.0	18.2
Qwen3-8B + WebWorld (Ours)	59.3	0.020	4.39	20.7	0.014	19.25	20.7	21.4	23.3	17.5
Improvement (8B)	+9.9%	-	-0.49	+10.9%	-	+4.01	+3.6%	+12.0%	+18.3%	-0.7%

Table 5: Downstream Performance. Comparison of the base Qwen3-8B and Qwen3-14B models versus the models fine-tuned on WebWorld-synthesized trajectories. We report Success Rate (SR %) , Standard Error (Std), and Average Steps. The symbols

\uparrow

and

\downarrow

indicate that higher and lower values are better, respectively. For WebArena, we detail performance across sub-domains.

Model & Search Configuration					Result
World Model	Value Model	Fmt	Score	Alg ( $k$ )	Reward	$\Delta$
Baselines
-	-	-	-	Greedy	64.3	-
GPT-4o	GPT-4o	A11y	Point	BoN (3)	63.8	-0.5%
Impact of Scoring & Value Model
Ours (WebWorld)	GPT-4o	A11y	Point	BoN (3)	64.8	+0.5%

Table 6: Inference-Time Lookahead Search on MiniWob. Fmt : NL = Natural Language, A11y = A11y Tree. Scoring : Point = Pointwise, Pair = Pairwise. Alg : BoN = Best-of-

N

Figure 4: Scaling Law of WebWorld. Larger models achieve lower eval loss. Stars indicate predictions for the 72B model, suggesting continued performance gains with model scaling.

Model Source	Data Scale	Factuality Score	Web Turing Score	Total Score
From Qwen3-8B (Direct Reasoning Tuning)
Qwen3-8B	500	0.502	0.277	0.390
	1k	0.511	0.296	0.403
	2k	0.541	0.319	0.430
	10k	0.625	0.394	0.510
From Stage 1 Model (1.06M Transition Modeling)

Table 7: Reasoning Activation Ablation. Comparison under varying reasoning data scales. WebWorld-8B achieves superior performance with only 1k samples .

Environment	1,500 Samples			3,000 Samples
	(Total Score)			(Total Score)
	Qwen3	Ours	Gain ( $\Delta$ )	Qwen3	Ours	Gain ( $\Delta$ )
API Services	0.088	0.299	+0.211	0.258	0.292	+0.034
Code	0.147	0.396	+0.249	0.196	0.471	+0.275
Game	0.253	0.473	+0.220	0.374	0.522	+0.148
GUI	0.322	0.705	+0.383	0.511	0.719	+0.208

Table 8: Cross-Environment Adaptation. WebWorld demonstrates strong adaptation capability across unseen environments.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Stage 1: Transition Modeling	Stage 2: Reasoning Activation
Training Configuration
Objective	Learn $(s_{t},a_{t})\rightarrow s_{t+1}$	Learn $(s_{t},a_{t})\rightarrow\text{thought}\rightarrow s_{t+1}$
Data Source	1.06M real-world trajectories	1K CoT samples
Initialization	Qwen3 Base Checkpoint	Stage 1 Checkpoint
Finetuning Type	Full Parameter	Full Parameter
Precision	BF16	BF16

Appendix

Table 9: Comparison of Hyperparameters Across Training Stages. Stage 1 focuses on large-scale dynamics learning with aggressive data throughput, while Stage 2 refines reasoning capabilities with conservative fine-tuning to prevent forgetting. Training times are reported for three model scales on NVIDIA A100 GPUs.

Model	Version
GPT-4o	gpt-4o-2024-11-20
Claude-Sonnet-4.5	claude-sonnet-4-5-20250929
Claude-Opus-4.1	claude-opus-4-1-20250805
Gemini-3-Pro	Gemini-3-Pro-preview
GPT-5	gpt-5-2025-08-07

Appendix

Table 10: API models and versions used for evaluations.

Category	Source	Lang.	Scale		Attributes
Category	Source	Lang.	# Traj.	Size	Real-World	Long-Seq.	Multi-Fmt.
Randomized Crawling	CCI 3.0 (Filtered)	CN	57,837	17.67G	$\checkmark$	$\times$	$\times$
	FineWeb Subset	EN	235,674	4.33G	$\checkmark$	$\times$	$\times$
	Subtotal	-	293,511	22.0G	-
Autonomous Exploration	FineWeb (LLM-Driven)	Mixed	36,474	9.9G	$\checkmark$	$\checkmark$	$\times$
	High-Freq Sites	Mixed	1,882	0.5G	$\checkmark$	$\checkmark$	$\times$

Appendix

Table 11: Statistics of the Constructed WebWorld Training Data. Attributes indicate if data is from Real-world websites, contains Long-horizon sequences (

>

10 steps), or supports Multi-format (HTML/A11y).

Figure 5: Domain and Source Distribution of WebWorld Training Data. The chart illustrates the composition of our trajectory dataset, which contains over one million samples, across 15 distinct data sources. The colors represent different semantic domains (e.g., Technology, E-Commerce), showing that our data collection pipelines significantly contribute to the diversity of open-domain topics compared to traditional web generation methods.

Category	Action Primitives	Description
Element Interactions	click(bid, button, mods)	Click a specific DOM element identified by bid .
	fill(bid, text, auto)	Input text into a focused field (supports autocomplete).
	select_option(bid, opts)	Select single or multiple options from a dropdown/combobox.
	hover(bid)	Hover the cursor over a specific element.
Coordinate & Mouse	mouse_move(x, y)	Move cursor to absolute screen coordinates.
	mouse_click(x, y, button)	Click at a specific coordinate (supports double click).

Appendix

Table 12: The unified action space available to the agent. Actions are categorized by interaction type. Parameters include element IDs ( bid ), coordinates (

x,y

), textual input ( text ), and navigation deltas.

Category	Action Primitive	Percentage	Category	Action Primitive	Percentage
Element Interactions	click	77.29%	Browser & Navigation	scroll	0.88%
	fill	5.12%		goto	10.06%
	select_option	0.96%		go_back	0.24%
	hover	0.06%		go_fwd	0.15%
	Subtotal	83.43%		tab_new	0.22%
Coordinate & Mouse	mouse_move	0.42%		tab_close	0.19%

Appendix

Table 13: Action Distribution by Category

Figure 6: LLM-based URL Filtering Pipeline. We evaluate candidate URLs across four dimensions—accessibility, content suitability, interactivity, and engineering quality—using an LLM judge. The chart shows the distribution of scores and the filtering threshold (red dashed line), which retains only the top 32% of URLs for data collection.

Category	Rank	Dataset Name	Project Source	Environment Type	Samples
A. Code & Development	1	wm_intercode_sql.jsonl	AgentBank/intercode_sql	code/IDE, API	4,522
	5	wm_full_sft.jsonl	Neulab/swe-smith	terminal/shell, code/IDE	17,380
	8	wm_train-00005-of-00012.jsonl	SWE-agent-trajectories	terminal/shell, code/IDE	6,665
	21	wm_train-00002-of-00012.jsonl	SWE-agent-trajectories	terminal/shell, code/IDE	6,661
	24	wm_full_sft.jsonl	Neulab/agenttuning_db	code/IDE, API	527
	13	wm_full_sft.jsonl	Neulab/agenttuning_os	terminal/shell	195

Appendix

Table 14: Overview of the Training Set in Cross-Environment Data (training set).

Category	Rank	Dataset Name	Project Source	Environment Type	Samples
A. Code & Development	27	wm_train-00001-of-00001.2.jsonl	SWE-agent-trajectories	terminal/shell, code	6,660
A. Code & Development	25	wm_mbpp_before.jsonl	AgentBank/mbpp_before	code/IDE	707
B. GUI & Desktop	2	wm_agentnet_ubuntu_5k.jsonl	AgentNet	GUI, application	5,000
C. Game & Simulation	16	wm_rearrange.jsonl	AgentBank/rearrange	game, GUI/desktop	299
C. Game & Simulation	29	wm_alfworld_sft.jsonl	Agent-ETO	game	3,119
D. API & Services	19	wm_db.jsonl	AgentInstruct	API, database	538

Appendix

Table 15: Overview of the Training Set in Cross-Environment Data (test set).

Figure 7: Impact of Reasoning Activation on Output Token Length. The plot compares the average tokens of answer between the Real-World Transition Modeling baseline (grey dashed line) and the Reasoning Activation stage (blue solid line). The introduction of reasoning data leads to a

\sim

49.4% reduction in output length, indicating a shift from verbose raw state prediction to a more concise and structured simulation pattern.

Panel of Figure 8 — Figure 8: Template for the WebWorld.

Panel of Figure 9 — Figure 9: Prompt for the Web Agent.

Panel of Figure 10 — Figure 10: Prompt for the Value Model to evaluate task completion (inference-time lookahead search).

Panel of Figure 11 — Figure 11: Generating abstract goal from seed goals (Agent Data Synthesis, Stage 1)

Panel of Figure 12 — Figure 12: Instantiating a specific goal from exploration traces (Agent Data Synthesis, Stage 2).

Panel of Figure 13 — Figure 13: Prompt for Factuality Score (WebWorld-Bench).

Panel of Figure 14 — Figure 14: Prompt for Web Turing Test (WebWorld-Bench).

Panel of Figure 15 — Figure 15: Prompt for Level 2 Data Collection: Self-proposed Task.

Panel of Figure 16 — Figure 16: Prompt for Level 2 Data Collection: Long-horizon Dependency.

Panel of Figure 17 — Figure 17: Prompt for Level 2 Data Collection: Composite Action Interaction.

Panel of Figure 18 — Figure 18: Prompt for Level 2 Data Collection: Curiosity Discovery.

Explore similar work

Jun 23, 2026

Qwen-AgentWorld: Language World Models for General Agents

A world model predicts environment dynamics based on current observations and actions, serving as a core cognitive mechanism for reasoning and planning. In this work, we investigate how world modeling based on language models can further push the boundaries of general agents. (i) We first focus on building foundation models for agentic environment simulation. We introduce Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, the first language world models capable of simulating agentic environments covering 7 domains via long chain-of-thought reasoning. Leveraging more than 10M environment interaction trajectories of 7 domains in real-world environments, we develop Qwen-AgentWorld through a three-stage training pipeline: CPT injects general-purpose world modeling capabilities from the state transition dynamics and augmented professional corpora, SFT activates next-state-prediction reasoning, and RL sharpens simulation fidelity through a tailored framework with hybrid rubric-and-rule rewards. To evaluate language world models, we present AgentWorldBench, a comprehensive benchmark constructed from real-world interactions of 5 frontier models on 9 established benchmarks. Empirical results demonstrate that Qwen-AgentWorld significantly outperforms existing frontier models. (ii) Beyond foundation models, we further investigate two complementary paradigms through which world modeling enhances general agents. First, as a decoupled environment simulator, Qwen-AgentWorld supports scalable and controllable simulation of thousands of real-world environments for agentic RL, yielding gains that surpass real-environment training alone. Second, as a unified agent foundation model, world-model training acts as a highly effective warm-up that improves downstream performance across 7 agentic benchmarks. Code: https://github.com/QwenLM/Qwen-AgentWorld

Yuxin Zuo, Zikai Xiao, Li Sheng +30

Qwen Team

LLM World Models Agentic RL World Model Learning World Model Evaluation

Ask PDF

Apr 9, 2026

Structured Distillation of Web Agent Capabilities Enables Generalization

Frontier LLMs can navigate complex websites, but their cost and reliance on third-party APIs make local deployment impractical. We introduce Agent-as-Annotators, a framework that structures synthetic trajectory generation for web agents by analogy to human annotation roles, replacing the Task Designer, Annotator, and Supervisor with modular LLM components. Using Gemini 3 Pro as teacher, we generate 3,000 trajectories across six web environments and fine-tune a 9B-parameter student with pure supervised learning on the 2,322 that pass quality filtering. The resulting model achieves 41.5% on WebArena, surpassing closed-source models such as Claude 3.5 Sonnet (36.0%) and GPT-4o (31.5%) under the same evaluation protocol, and nearly doubling the previous best open-weight result (Go-Browse, 21.7%). Capabilities transfer to unseen environments, with an 18.2 percentage point gain on WorkArena L1 (an enterprise platform never seen during training) and consistent improvements across three additional benchmarks. Ablations confirm that each pipeline component contributes meaningfully, with Judge filtering, evaluation hints, and reasoning traces each accounting for measurable gains. These results demonstrate that structured trajectory synthesis from a single frontier teacher is sufficient to produce competitive, locally deployable web agents. Project page: https://agent-as-annotators.github.io

Xing Han Lù, Siva Reddy

Mila – Quebec AI Institute McGill University

LLM Agent Training Web Agents LLM Agent Trajectory Synthesis Language Model Distillation

Ask PDF

Apr 9, 2026

MolmoWeb: Open Visual Web Agent and Open Data for the Open Web

Web agents--autonomous systems that navigate and execute tasks on the web on behalf of users--have the potential to transform how people interact with the digital world. However, the most capable web agents today rely on proprietary models with undisclosed training data and recipes, limiting scientific understanding, reproducibility, and community-driven progress. We believe agents for the open web should be built in the open. To this end, we introduce (1) MolmoWebMix, a large and diverse mixture of browser task demonstrations and web-GUI perception data and (2) MolmoWeb, a family of fully open multimodal web agents. Specifically, MolmoWebMix combines over 100K synthetic task trajectories from multiple complementary generation pipelines with 30K+ human demonstrations, atomic web-skill trajectories, and GUI perception data, including referring expression grounding and screenshot question answering. MolmoWeb agents operate as instruction-conditioned visual-language action policies: given a task instruction and a webpage screenshot, they predict the next browser action, requiring no access to HTML, accessibility trees, or specialized APIs. Available in 4B and 8B size, on browser-use benchmarks like WebVoyager, Online-Mind2Web, and DeepShop, MolmoWeb agents achieve state-of-the-art results outperforming similar scale open-weight-only models such as Fara-7B, UI-Tars-1.5-7B, and Holo1-7B. MolmoWeb-8B also surpasses set-of-marks (SoM) agents built on much larger closed frontier models like GPT-4o. We further demonstrate consistent gains through test-time scaling via parallel rollouts with best-of-N selection, achieving 94.7% and 60.5% pass@4 (compared to 78.2% and 35.3% pass@1) on WebVoyager and Online-Mind2Web respectively. We will release model checkpoints, training data, code, and a unified evaluation harness to enable reproducibility and accelerate open research on web agents.

Tanmay Gupta, Piper Wolters, Zixian Ma +13

Allen Institute for AI · University of Washington · UNC-Chapel Hill

Web Agents LLM Agent Training Tool-Using Vision-Language Agents

Ask PDF