cs.CLMar 11, 2026
SaveSafe and Scalable Web Agent Learning via Recreated Websites
Organizations: School of Interactive Computing, Georgia Institute of Technology
Abstract
Training autonomous web agents is fundamentally limited by the environments they learn from: real-world websites are unsafe to explore, hard to reset, and rarely provide verifiable feedback. We propose VeriEnv, a framework that treats language models as environment creators, automatically cloning real-world websites into fully executable, verifiable synthetic environments. By exposing controlled internal access via a Python SDK, VeriEnv enables agents to self-generate tasks with deterministic, programmatically verifiable rewards, eliminating reliance on heuristic or LLM-based judges. This design decouples agent learning from unsafe real-world interaction while enabling scalable self-evolution through environment expansion. Through experiments on web agent benchmarks, we show that agents trained with VeriEnv generalize to unseen websites, achieve site-specific mastery through self-evolving training, and benefit from scaling the number of training environments. Code and resources will be released at https://github.com/kyle8581/VeriEnv upon acceptance.
Figures & tables
Figure 1: Comparison between the traditional self-evolution paradigm and our verifiable environment framework. (a) In traditional settings, agents interact directly with real-world environments and rely on unvalidated synthetic tasks and non-verifiable, LLM-based reward signals, leading to unsafe exploration and unreliable learning. (b) In contrast, VeriEnv clones real-world websites into synthetic environments with full internal access, enabling safe exploration, validated task generation, and deterministic, verifiable reward signals for stable and scalable agent learning.
Figure 2: Overview of VeriEnv . VeriEnv first clones a real website into a fully instrumented synthetic environment (code , database , and a Python SDK ) via coding agent, then uses task and judge generators to produce tasks at varying difficulty and verify both tasks and judges by interacting with the website and database through the SDK, yielding deterministic, verified rewards for agent learning.
Figure 3: Example of a verifiable task with executable validation in a synthetic recipe website ( i.e. , cloned from apartments.com).
| Statistic | Value |
| Number of websites | 149 |
| Number of tasks per website | 49.5 |
| Total number of tasks | 7,400 |
| Easy tasks | 2,972 (40.2%) |
| Medium tasks | 2,900 (39.2%) |
| Hard tasks | 1,528 (20.6%) |
Table 1: Statistics of constructed synthetic environments and generated tasks.
| Dataset / Benchmark | # Websites | # Tasks | Browser Interaction | Verifiable Judge | Scalable Task Gen. |
| WebArena | 5 | 812 | ✓ | ✓ | ✗ |
| WorkArena | 1 | 33 | ✓ | ✓ | ✗ |
| WebVoyager | 15 | 643 | ✓ | ✗ | ✗ |
| Mind2Web | 137 | 2,350 | ✗ | ✓ | ✗ |
| Mind2Web-Online | 136 | 300 | ✓ | ✗ | ✗ |
| Mind2Web-Live | 137 | 542 | ✓ | ✗ | ✗ |
Table 2: Comparison with existing web agent datasets and benchmarks. VeriEnv uniquely enables verifiable evaluation and scalable task generation through executable synthetic environments.
| Metric | Result |
| Environment quality | |
| Functional correctness (avg.) | 90% |
| Signup | 94% |
| Login | 95% |
| Search | 81% |
| Filter | 88% |
Table 3: Human evaluation of the generated websites and tasks.
| Method | Shopping | CMS | GitLab | Map | Total | ||
| GPT-4o-mini ( Hurst et al., 2024 ) * | 21.74 | 22.86 | 19.05 | 34.38 | 19.35 | 23.64 | – |
| GPT-4o ( Hurst et al., 2024 ) * | 23.91 | 31.43 | 28.57 | 56.25 | 19.35 | 31.52 | – |
| Qwen3-4B | 3.77 | 6.67 | 4.17 | 13.89 | 14.29 | 7.88 | – |
| +Synatra ( Ou et al., 2024 ) | 0.00 | 0.00 | 12.50 | 8.33 | 0.00 | 3.64 | |
| +ADP ( Song et al., 2025 ) | 4.35 | 5.71 | 9.52 | 3.13 | 9.68 | 6.06 | |
| + VeriEnv (Ours) | 4.35 | 20.00 | 23.81 | 12.50 | 16.13 | 13.94 | +6.06 |
Table 4: WebArena-Lite evaluation results across different websites. Methods annotated with * use numbers reported by Chae et al. (2025) .
| Method | Easy | Medium | Hard | Total | |
| Browser-Use-GPT-4o ( Browser-Use Contributors, 2024 ) * | 55.40 | 26.60 | 8.10 | 30.00 | – |
| Claude-3.5-Sonnet ( Anthropic, 2025 ) * | 56.60 | 26.60 | 6.80 | 28.80 | – |
| Qwen3-4B | 26.32 | 9.41 | 11.63 | 13.18 | – |
| +Synatra ( Ou et al., 2024 ) | 35.09 | 5.88 | 9.30 | 14.55 | |
| +ADP ( Song et al., 2025 ) | 26.32 | 7.06 | 6.98 | 11.36 | |
| + VeriEnv (Ours) | 29.82 | 23.53 | 6.98 | 20.45 | +7.27 |
Table 5: Mind2Web-Online results across difficulty levels. Methods annotated with * use numbers reported by Xue et al. (2025) .
Figure 4: Site-specific self-evolving training within a cloned synthetic environment. Agents are trained on a fixed target website using automatically generated tasks and verifiable reward signals.
Figure 5: Analysis on the scaling effect of the number of websites.
Figure 6: Comparison of task ambiguity and evaluation reliability in PAE ( Zhou et al., 2025b ) and VeriEnv .
Figure 7: Primary failure reasons for websites excluded and bug types in error report.
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Distribution of website implementation time for constructing synthetic environments using a coding agent. Each bar shows the number of websites grouped by implementation duration, categorized as fast (¡45 minutes), medium (45–90 minutes), and slow (¿90 minutes). The distribution indicates that most websites can be reconstructed within a moderate time budget, with a long tail corresponding to more complex implementations.
| Website | Original URL | Part 1 | Part 2 | Part 3 |
| CarMax | https://www.carmax.com/ | Fig. 18 | Fig. 19 | Fig. 20 |
| CVS | https://www.cvs.com/ | Fig. 21 | Fig. 22 | Fig. 23 |
| Eventbrite | https://www.eventbrite.com/ | Fig. 24 | Fig. 25 | Fig. 26 |
| Google Finance | https://www.google.com/finance/ | Fig. 27 | Fig. 28 | Fig. 29 |
Table 6: Cloned synthetic website screenshots. Each entry links to the corresponding figure for Parts 1–3.
Figure 9: Website filtering flow for benchmark construction: 136 candidate websites are progressively filtered based on server script availability and task generation success, resulting in 97 validated benchmark websites.
| Website | Task Instruction | Judge Criteria | Diff. |
|---|---|---|---|
| adoptapet | Filter to cats and tell me how many results you get. | must_include: “2” | easy |
| airbnb | On the home page, what categories are available in the category row? Please list all of them. | must_include_all: “OMG!”, “Lakefront”, “Amazing pools”, … (11 items) | easy |
| allrecipes | Open the Ingredients A–Z directory and click the letter A. Tell me the first five ingredient names listed under A. | must_include: “Apple Cider Vinegar”, “Avocado”, “albacore tuna”, “alfalfa”, “almond oil” | easy |
| apartments | In Columbus, OH, sort listings by price (low to high). What is the first listing’s name and minimum monthly price? | must_include: “The Charles at Bexley”, “2938” | medium |
| bestbuy | Sign in, save two specific products, open Saved Items and tell me how many total saved items. | exact_match: “2” | easy |
| budget | Find the Red Ball Parking Garage location in New York, NY. Tell me the location code and hours. | must_include: “NYC1”, “Mon-Sun 7:00 AM - 10:00 PM” | easy |
Table 7: Example tasks and evaluation criteria from VeriEnv environments.
Figure 10: Prompt used for website reconstruction by the coding agent.
Figure 11: Prompts used for website implementation, bug reporting, and debugging by the coding agent.
Figure 12: Prompt used to generate verifiable task instructions and executable judges using the Python SDK.
Figure 13: A browser snapshot attached to the bug report ‘bug_report_2026-01-07_23-36-45’.
Figure 14: Bug report example highlighting a missing homepage section discovered during automated traversal.
Figure 15: Bug report example showing incomplete long-form content on an informational page relative to the reference.
Figure 16: Bug report example illustrating a desktop–mobile layout mismatch detected across viewports.
Figure 17: Structured debug report after patching, summarizing fixes and verification results.
Figure 18: Screenshot from a cloned CarMax website (Part 1).
Figure 19: Screenshot from a cloned CarMax website (Part 2).
Figure 20: Screenshot from a cloned CarMax website (Part 3).
Figure 21: Screenshot from a cloned CVS website (Part 1).
Figure 22: Screenshot from a cloned CVS website (Part 2).
Figure 23: Screenshot from a cloned CVS website (Part 3).
Figure 24: Screenshot from a cloned eventbrite website (Part 1).
Figure 25: Screenshot from a cloned eventbrite website (Part 2).
Figure 26: Screenshot from a cloned eventbrite website (Part 3).
Figure 27: Screenshot from a cloned Google Finance website (Part 1).
Figure 28: Screenshot from a cloned Google Finance website (Part 2).
Figure 29: Screenshot from a cloned Google Finance website (Part 3).
More in this topic
Web Agents
Web agents use language and vision models to navigate websites, gather information, and complete tasks through interactive interfaces.
See all 174 papers