Training agents for front-end web development requires demonstrations of code editing, execution, browser inspection and revision, yet public trajectories and reusable task environments remain limited. We introduce WebUniverse, a pipeline that constructs verifiable tasks with verified development trajectories from open-source repositories and real web pages. The tasks cover text-grounded application development in standalone and iterative settings and vision-grounded reproduction across pages and viewports. We collect agent trajectories for these tasks and verify the resulting applications against task-specific functional or visual requirements. We release WebUniverse-Env, containing 15K executable task environments, and WebUniverse-Mix, containing the corresponding 15K development trajectories with verification records and scores. Fine-tuning Qwen3.5-35B-A3B on WebUniverse-Mix raises the Vision2Web average score from 51.7 to 68.3, WebGen-Bench accuracy from 34.5% to 63.9%, and Web-Bench pass@2 from 12.2% to 23.1%. The resulting model also outperforms a public-data baseline trained with the same configuration and sample budget across all reported metrics. Ablations show complementary benefits from text and visual trajectories and continued gains as training data scales from 5K to 15K trajectories.
Figures & tables
Table 1: Training resources for coding agents. Pair: code-supervision pairs; AT: agent trajectories; ENV: executable task environments. Runtime checks assess training solutions; public environments enable new attempts; records contain per-sample verification outcomes. : documented coverage or release; : partial coverage or release; : not established in the reviewed sources.
Figure 2: The WebUniverse construction pipeline. Independent sampling and trajectory refinement produce candidate trajectories, which undergo task-specific verification and quality filtering to form WebUniverse-Mix . Task inputs, starting workspaces and verification criteria form WebUniverse-Env .
Figure 3: Composition of WebUniverse-Mix (15K trajectories). Rings encode task family, variant and collection mode, from inside to outside. The outer ring distinguishes independent sampling (light gray) from trajectory refinement (dark gray). Iterative counts represent trajectories for individual tasks.
Table 2: Main benchmark results; Vision2Web reports Level 1 scores. The + rows fine-tune Qwen3.5-35B-A3B on 15K samples from StepFun-Mix or WebUniverse-Mix . Func., Data and Design denote functional, data-display and design-validation accuracy; DPSK denotes DeepSeek and PA denotes pass. WebGen-Bench and Web-Bench report percentages; Web-Bench Avg. averages PA@1 and PA@2. Higher is better. Bold WebUniverse-Mix scores exceed at least one of the first four reference models; underlined reference scores are below ours.
Table 3: Data composition ablation of WebUniverse-Mix using the same checkpoint and training configuration. Metrics follow Table 2 ; bold and underlined values indicate the best and second-best scores, respectively.
Figure 4: Results by training-data size. Open markers at 0 denote Qwen3.5-35B-A3B; 15K denotes Qwen3.5-WebUniverse. Panels use different y-axis ranges with nonzero origins. Higher is better.
Figure 5: Qualitative examples: (a) corresponding lower-page sections on Vision2Web’s tefcu task; (b) phone-operator applications and recorded checks on WebGen-Bench task 000020.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
UI stack
Language
State library
Styling
Build tool
React 18
TypeScript
—
Tailwind
Vite
Vue 3
TypeScript
—
Tailwind
Vite
React 18
TypeScript
Jotai
Plain CSS
Vite
React 18
TypeScript
MobX
Plain CSS
Vite
React 18
TypeScript
Redux Toolkit
Plain CSS
Vite
React 18
TypeScript
Zustand
Plain CSS
Vite
Appendix
Table 4: Fifteen project scaffolds for detailed tasks, arranged in nine groups. Each row is one scaffold; — denotes no additional state-management library.
Path
Contents
spec.json
Task and step statements, categories, source records, fixed checks, starting-state information, repository commit and image digest.
start/
Saved initial project files, when needed.
environment/
Reference screenshots, assets and multi-page navigation information.
verifier/task.json
Visual targets, viewports and reference-image mapping.
solution/<step>/
Reference implementation for each step.
manifest.json
SHA-256 hashes of the package files.
Appendix
Table 5: Task-package files. Optional directories are included when required by the task.
Technology
Environments: initial workspaces
Trajectories: observed code
Text-grounded
Vision-grounded
Text-grounded
Vision-grounded
React
4,154 (39.6%)
549 (12.2%)
6,859 (65.3%)
549 (12.2%)
vanilla
1,613 (15.4%)
1,286 (28.6%)
3,014 (28.7%)
3,951 (87.8%)
Vue
527 (5.0%)
0 (0.0%)
527 (5.0%)
0 (0.0%)
Angular
60 (0.6%)
0 (0.0%)
60 (0.6%)
0 (0.0%)
Svelte
40 (0.4%)
0 (0.0%)
40 (0.4%)
0 (0.0%)
Appendix
Table 6: Technologies in environment initial workspaces and code observed in trajectories: 15K WebUniverse-Env entries and 15K WebUniverse-Mix trajectories.
Family
≤ 10
11–20
21–50
51–100
> 100
Total
Text-grounded
496
4,786
5,038
176
4
10,500
Vision-grounded
0
4
860
3,299
337
4,500
Total
496
4,790
5,898
3,475
341
15K
Appendix
Table 7: Distribution of tool calls per trajectory.
Specification
Setting
Count
High-level
Standalone
4,500
High-level
Iterative
1,000
Detailed
Standalone
1,855
Detailed
Iterative
3,145
Total
10,500
Appendix
Table 8: Text-grounded trajectories and environments by specification level and development setting. Counts are identical for both.
Verifier
Accuracy
Precision
Recall
F1
Browser scripts
94.5
98.0
91.6
94.7
GUI agent
95.5
97.0
94.2
95.6
Visual judge
85.0
87.0
83.7
85.3
Appendix
Table 9: Verifier performance on 600 manually audited attempts. Values are percentages. Each verifier is evaluated on 100 originally accepted and 100 originally rejected attempts, with human judgments as the reference.
Table 10: Supplementary Vision2Web Level 2 results. The + row fine-tunes Qwen3.5-35B-A3B on WebUniverse-Mix . Composite averages the visual and functional scores within each run; each column reports the median across three runs. Higher is better; bold marks the higher score in each column.
Figure 7: Complete desktop pages for the tefcu example in Vision2Web Level 1. The reference and outputs from Qwen3.5-35B-A3B and Qwen3.5-WebUniverse are shown with their aspect ratios preserved.
Behavior metric
Vision2Web Level 1
WebGen-Bench
Web-Bench
Base
Tuned
Base
Tuned
Base
Tuned
Evaluated units ( n )
100
100
101
101
275
305
Pre-edit inspection
5.0
77.0
1.0
0.0
73.8
85.2
Post-edit verification
90.0
100.0
50.5
95.0
72.7
75.7
Error-triggered revision
0.0
1.0
45.5
73.3
1.1
2.0
Server-launch timeout
98.0
43.0
51.5
27.7
0.0
1.0
Appendix
Table 11: Behavior occurrence rates (%) in evaluation traces. Base: Qwen3.5-35B-A3B; Tuned: Qwen3.5-WebUniverse. Vision2Web and WebGen-Bench use task-level units with retries combined; Web-Bench uses individual attempts. Tuned-model traces are from supplementary runs. Metrics are defined in the text.
Region score
Base
Tuned
0
12.0
2.9
0.25
16.1
5.2
0.5
27.3
22.6
0.75
40.8
50.5
1
3.8
18.8
Evaluated regions ( n )
2,709
2,739
Appendix
Table 12: Vision2Web Level 1 region-score distributions (%). Base: Qwen3.5-35B-A3B; Tuned: Qwen3.5-WebUniverse. Percentages are computed within each model’s evaluated regions. The tuned-model run is supplementary to Table 2 .
Failure category
Diagnostic criterion
Tasks
Share (%)
Browser runtime failure
Blank page during browser evaluation.
13
72.2
Service-port mismatch
Service runs on a port other than the required port.
4
22.2
Execution loop
Repeated tool calls; no launch script delivered before the time limit.
1
5.6
Appendix
Table 13: Failure categories among 18 tasks with zero visual and functional scores in the inspected Qwen3.5-WebUniverse Vision2Web Level 2 run. Shares use these 18 tasks as the denominator, out of 66 evaluated tasks.