InfiniteWeb: Scalable Web Environment Synthesis for GUI Agent Training

• Existing benchmarks are manually built and contain only tens to hundreds of applications, while synthetic-task approaches remain confined to the same small set of benchmark environments.Source: §1 · p. 1
• Multi-page generation risks incompatible interfaces, data formats, and state management; functional bugs can compound across long tasks, while repetitive tasks and styles can encourage overfitting.Source: §1 · p. 2
• INFINITEWEB addresses consistency with a shared specification, correctness with task-centric test-driven development, and diversity with seed-matched tasks and design-image-guided websites.Source: §1 · p. 2
• Its verifiable task evaluators track task-related variables during agent runs, enabling dense reward signals for reinforcement learning.Source: §1 · p. 2
• +6.9% on OSWorld, +5.7% on Online-Mind2Web, and +3.9% on MobileWorld are training gains; INFINITEWEB surpasses advanced coding agents on WebGen-Bench in visual and functional quality.Source: §1 · p. 3

GUI Agent Benchmarks

• Benchmarks such as WebArena, OSWorld (Xie et al., 2024), and Mind2Web (Deng et al., 2023) require substantial manual construction, limiting scale and diversity.InfiniteWeb addresses this bottleneck by automatically generating complete functional environments, including websites, tasks, and evaluators.Source: §2 · pp. 3–4

LLM-based Code and Website Generation

• Unlike CodeT (Chen et al., 2022), which generates tests to filter incorrect code, InfiniteWeb verifies task-centric correctness so required functionality is bug-free.Focusing verification on functionality needed for specific user tasks makes the verification problem tractable.Source: §2 · p. 3

Synthetic Environment and Data Generation

• Unlike WebSailor-V2 (Li et al., 2025b), UI-Evol (Zhang et al., 2025), and AgentSynth (Xie et al., 2025a), which generate training data within existing environments, InfiniteWeb generates the environments themselves.This includes websites, tasks, and automatic evaluators.Source: §2 · pp. 3–4

3.1 Overview

• The pipeline uses Common Crawl website seeds and design images to generate functional websites, website-specific tasks, and evaluators through four stages, with backend and frontend generation in parallel.The stages are unified specification, task-centric backend, design-guided frontend, and evaluator generation.Source: §3.1 · p. 4

3.2 Unified Specification Stage

• Task-derived data models, page plans, and shared interfaces keep pages consistent while tasks vary by website seed and interface signatures omit system-managed parameters.The system generates seed-specific tasks, then derives shared specifications; parameters such as userId and sessionId are retrieved internally from localStorage.Source: §3.2 · pp. 4–5

3.3 Task-Centric Backend

• Task-centric correctness limits verification to task-relevant functionality, using generated data and parallel test/code generation followed by iterative fixes until tests pass or the iteration limit is reached.Generated test cases and implementations use the same data; placeholder resources can be replaced with context-relevant content via external APIs.Source: §3.3 · p. 5

3.4 Design-Guided Frontend

• A vision-language model extracts global visual styles from reference images, which guide page architecture and realization with shared styling, backend-connected UI elements, and localStorage-based data persistence.Page design defines content, interaction flows, URL-parameter routing, and responsive layouts; homepage initialization stores generated data in localStorage for cross-page persistence.Source: §3.4 · p. 5

3.5 Automatic Evaluator Generation

• Generated JavaScript evaluators combine application state with task-step instrumentation to assess partial completion and award dense reinforcement-learning rewards.Instrumentation records task progress in localStorage as corresponding functions execute, enabling credit for completed steps rather than only binary success or failure.Source: §3.5 · p. 6

4.1 Experimental Setup

• Evaluation uses WebGen-Bench task outcomes, GPT-5 pairwise design judgments, and Online-Mind2Web, OSWorld-Verified, and MobileWorld to assess web, desktop, and mobile transfer.Visual judgments compare full-page screenshots with reference designs across layout, color, typography and spacing, component structure, and aesthetics.Source: §4.1 · p. 7

4.2 Website Functional Correctness

• 85.6% overall score leads the WebGen-Bench baselines; INFINITEWEB also achieves 80.9% in Functional Testing and 82.8% in Design Validation.The benchmark evaluates task completion on generated websites and reports Passed, Partial, and Overall scores; design images are supplied to all methods.Source: §4.2 · p. 7

4.3 LLM-as-Judge Visual Quality

• INFINITEWEB achieves 69–85% win rates against Codex, Claude-Code, and Bolt.diy in pairwise visual-quality comparisons.Human evaluation agrees with automated assessments 91% of the time.Source: Figure 6 · p. 7, §4.3 · p. 8

4.4 Effectiveness for Agent Training

• +6.9% on OSWorld (24.5% →31.4%), +5.7% on Online-Mind2Web (23.0% →28.7%), and +3.9% on MobileWorld (6.4% →10.3%) after training on generated environments.Improvements span out-of-domain desktop, in-domain web, and mobile transfer; the authors report that data quality, not quantity, drives improvement, and gains scale with training data.Source: §4.4 · p. 8

4.5 Generated Environment Quality

• On InfiniteWeb, agents score 2–3× lower than on OSWorld and require longer successful-task trajectories; the Agent S2–UI-TARS score gap is 6.7 points versus 2.8 on OSWorld.These results indicate that InfiniteWeb tasks are more challenging and more discriminative of agent capability.Source: Table 3 · p. 8, §4.5 · p. 8

4.6 Ablation Studies

• Removing TCTDD lowers the overall score by 5.0 points, while replacing GPT-5 with GPT-4.1 lowers it by 8.2 points (85.6 →77.4).Without TCTDD, the method still scores 80.6%, comparable to Codex; with GPT-4.1, it still exceeds Claude-Code with GPT-5 (75.8%).Source: §4.6 · p. 9
• Dense rewards make 767 tasks discriminative for GRPO learning versus 174 with binary rewards, a 4.4× increase.Partial credit for intermediate progress expands the effective training signal.Source: Figure 8 · p. 8, §4.6 · p. 9

4.7 Generation Efficiency

• Generating one website costs approximately $1.93 and takes a median of approximately 20 minutes with the reported API configuration.Generation averages 0.36M input and 0.34M output tokens; independent websites can be generated in parallel, and runtime depends on API response speed and rate limits.Source: §4.7 · p. 9

5 Conclusion

• InfiniteWeb targets functional web environments for GUI-agent training, addressing consistency, correctness, and diversity through unified interface design, task-centric test-driven development, and website-seed variation with design-image guidance.The authors present the system as a response to the challenges of constructing realistic, functional websites with interconnected pages.Source: §5 · p. 9
• InfiniteWeb surpasses commercial coding agents in website construction, and experiments show advantages for training GUI agents.The authors release the system and generated datasets to support research on more capable and generalizable GUI agents.Source: §5 · p. 9

Limitations

• Current tasks operate within individual websites, leaving coordinated cross-website tasks such as price comparisons or cross-source aggregation uncovered.The authors identify these as important real-world scenarios for future work.Source: Limitations · p. 9
• Single-website training improves OSWorld multi-application task performance from 3.8% to 9.7%, suggesting some cross-application transfer.Source: Limitations · p. 9
• Despite responsive layouts and +3.9% transfer to MobileWorld, InfiniteWeb currently generates web environments only.Source: Limitations · p. 9
• Generating native mobile application environments for mobile GUI-agent training remains future work.Source: Limitations · p. 9
• Website generation requires multi-stage LLM calls for task generation, architecture design, code generation, and test validation.Source: Limitations · p. 9
• Median generation takes approximately 20 minutes per website and costs $1.93 per site using GPT-5 batch pricing.Source: Limitations · p. 9
• Generation is orders of magnitude cheaper than OSWorld's approximately 1,800 person-hours for 369 tasks, accounts for 6.9% of training time, and parallel execution produces 600 websites in approximately 125 minutes.Further reducing per-website cost and latency remains an engineering direction.Source: Limitations · p. 9

A.1 Cross-Domain Transfer Analysis

• Website training develops domain-agnostic exploration persistence, workflow completeness, and loop avoidance that also apply to image editors, office suites, and system utilities.Diverse interaction patterns and multi-step transactions in website environments train these transferable behaviors.Source: §A.1 · pp. 11–12
• In the Arabic language-change task, the baseline stopped after 5 steps, while the trained model persisted for 15 steps and successfully found and selected Arabic.The trained model tried typing the language name, scrolling, and clearing and retrying after its initial approach failed.Source: §A.1 · p. 11
• For Spotify installation, the baseline stopped after 4 steps without installing, while the trained model completed the 13-step flow through authentication, installation, and verification.The trained model searched, clicked Install, entered a password, waited for installation, and verified completion.Source: §A.1 · pp. 11–12
• In the Thunderbird attachment task, the baseline entered an 11-step loop after attaching the file, while the trained model completed the task cleanly in 5 steps.The baseline had attached the file in 4 steps before repeatedly opening and canceling the file picker.Source: §A.1 · p. 11

A.2 Automatic Evaluator Generation

• Weighted evaluator checkpoints track required actions, verify created records, and confirm workflow completion, giving proportional partial credit while rejecting shortcuts such as direct localStorage manipulation.The weighted sum provides denser GRPO training signals than sparse 0/1 rewards, including when an agent initiates but does not complete a task.Source: §A.2 · p. 12

A.3 TCTDD Validation and Auto-Fix

• Most generated websites require 1–3 iterations to pass all tests, with a maximum of 8 iterations allowed.The TCTDD loop detects a failing test, uses an LLM to analyze and fix the implementation, then re-validates until tests pass.Source: §A.3 · pp. 12–13, Table 4 · p. 13

B Data Collection and Implementation

• Common Crawl pages were rendered in a headless browser; screenshots became design images, and an LLM generated website seeds from their visual content.Pages violating robots.txt or containing illegal content were filtered out.Source: B Data Collection and Implementation · p. 13
• Generation used 8–10 tasks per website, at most 12 pages, and up to 8 TCTDD validation iterations.The configuration also set temperature to 0.7 and maximum output to 32,000 tokens.Source: B Data Collection and Implementation · p. 13
• UI-TARS-1.5-7B was post-trained with GRPO using 128 parallel environments, 8 trajectories per task, and episodes capped at 15 steps.Training used learning rate 1e-6, AdamW with bf16, gradient clipping 1.0, global batch size 16, one PPO epoch, clip ratio 0.2–0.3, discount factor γ=0.95, and sampling temperature 1.0.Source: B Data Collection and Implementation · p. 13

C Experimental Details and Results

• For direct-prompting baselines, GPT-5 with high reasoning effort received a prompt specifying the website seed, functionality, technical requirements, and code standards.The requirements included up to 12 pages, localStorage, and a reference design image.Source: C Experimental Details and Results · p. 13, §C.1 · p. 13

C.1 WebGen-Bench Results

• InfiniteWeb significantly outperformed Bolt.diy, Claude-Code, and Codex on WebGen-Bench across three runs, with Welch’s t-tests supporting each comparison.The reported tests were t=14.81, p<0.001 versus Bolt.diy; t=6.33, p<0.01 versus Claude-Code; and t=6.57, p<0.05 versus Codex.Source: §C.1 · p. 13
• Both WebGen-Bench ablations significantly degraded results: using GPT-4.1 instead of GPT-5 and removing TCTDD validation.The reported tests were t=7.70, p<0.01 for the model change and t=2.82, p<0.05 for removing validation.Source: §C.1 · p. 13

C.2 Online-Mind2Web Results

• Welch’s t-tests against baseline report significant differences at 200 tasks (t=2.96, p<0.05), 400 tasks (t=4.47, p<0.05), and 600 tasks (t=6.58, p<0.01).Table 7 reports Online-Mind2Web results by difficulty, with standard deviation over three runs; the supplied passages do not give its cell values.Source: Table 7 · p. 14, §C.2 · p. 14

C.4 Comparison with Claude-Code-Generated Websites

• The comparison controls training data quantity and setup: both website sources use 600 tasks, GPT-5 with high reasoning effort for generation, and three-run averages.Both conditions use the same 600-task GRPO setup.Source: §C.4 · p. 14

C.3 Appearance Win Rate

• The appearance comparison measures how often InfiniteWeb-generated websites are judged visually closer to the reference design image, with results averaged over three runs.The supplied passages describe the win-rate measure but do not provide its numerical results.Source: §C.3 · p. 14, Table 8 · p. 15

C.4 Comparison with Claude-Code-Generated Websites

• With the same 600-task GRPO setup, InfiniteWeb achieves significant gains across all three benchmarks, whereas Claude-Code-generated websites yield marginal or negative improvements.Table 9 compares the two website sources using identical GRPO training; the supplied passages do not give its cell values.Source: §C.4 · p. 14, Table 9 · p. 15
• InfiniteWeb supplies each generation step only directly relevant specifications, while Claude-Code carries the full codebase in context.The paper identifies context management as the source of the quality gap.Source: §C.4 · pp. 14–15
• Claude-Code websites have 4.9 pages per website versus InfiniteWeb’s 12.6, alongside cross-page inconsistencies and less complete business logic as context grows.The passage links these limitations to accumulating the full codebase in context.Source: §C.4 · p. 15

D Human Evaluation for Visual Quality

• The study randomly sampled 100 comparison cases across InfiniteWeb-Codex, InfiniteWeb-Claude, and InfiniteWeb-Bolt, with approximately equal representation.Source: D Human Evaluation for Visual Quality · p. 15
• Evaluators compared a reference design image with two website screenshots and selected the implementation that more closely matched the reference.Source: D Human Evaluation for Visual Quality · p. 15
• 91% agreement with automated GPT-5 evaluations indicates that the visual-quality assessment aligns well with human judgments.Disagreements primarily involved subtle cases where both implementations were reasonably close to the reference design.Source: D Human Evaluation for Visual Quality · p. 15

E Human Verification of Task and Evaluator Quality

• The human study sampled 100 tasks and assessed task clarity and executability alongside automatic evaluator correctness.Evaluators judged whether tasks could be executed on the generated website and whether completion was correctly detected.Source: E Human Verification of Task and… · p. 15
• 95 of 100 sampled tasks passed human verification, supporting the quality of generated tasks and their automatic evaluators.Human evaluators checked whether each task was clear and executable on its website and whether its evaluator correctly determined completion.Source: E Human Verification of Task and… · p. 15
• Only 1.5% of generated websites remained unfixed after the maximum 8 TCTDD iterations.The paper presents this validation-loop statistic as evidence of the iterative test-driven approach's effectiveness in ensuring functional correctness.Source: E Human Verification of Task and… · p. 15

F Prompts

• The prompt appendix includes templates for baseline website generation, task generation, architecture design, data-model extraction, interface design, and parameter-hiding wrappers.Source: Figure 13 · p. 16, Figure 14 · p. 17, Figure 15 · p. 17
• Further templates cover complete website architecture with navigation, page functionality and components, design-image analysis, component layouts, and page frameworks.The page-framework prompt generates headers and footers from a design image.Source: Figure 19 · p. 21, Figure 20 · p. 22, Figure 21 · p. 23
• Implementation prompts address HTML with integrated UI JavaScript, CSS styling, realistic website data, business logic, integration tests, and task-completion evaluators.Source: Figure 24 · p. 24, Figure 25 · p. 25, Figure 26 · p. 25
• The final prompt set covers instrumentation requirements, instrumented code with tracking variables, and evaluators that use instrumentation support.Source: Figure 30 · p. 27, Figure 31 · p. 28, Figure 32 · p. 28