acceptodds
Under review as a conference paper at ICLR 2027

MockWeb-Bench: Evaluating Hybrid Agents through Website Reconstruction

Abstract

Reconstructing a complete website from a running reference requires discovering its appearance and behavior across pages, interaction states, and screen sizes, then reproducing both in an independent implementation. Unlike fixed screenshots or recorded interactions, a running reference lets agents decide what to inspect as they build. We introduce MockWeb-Bench, which evaluates hybrid graphical user interface (GUI) and coding agents on 50 offline reference websites with 4,101 manually verified evaluation pages and paired desktop and mobile captures. Each task provides an entry URL, a starter workspace, and a six-hour initial budget; the list of scored pages stays hidden. Functional tests are validated on the reference and frozen before candidate generation, then applied to each reconstruction to measure behavioral fidelity. A fixed vision-language model (VLM) checks candidate screenshots against visual requirements extracted from reference screenshots, while separate checks assess quality and structure. Across ten model–agent configurations,GPT-6 Astra with Codex achieves the highest mean final score of 56.26%. All ten model-agent systems score lower on interaction tests than on general content and navigation tests. Beyond final scores, trajectory case studies reveal both useful strategies and attempts to take shortcuts. Agents write code to compute pixel differences between reference and candidate screenshots, using the results to make reconstructed pages look much closer to the reference. Some systems also try to read protected files or copy reference code, showing why evaluations need to restrict access to reference files and check submissions for copying. MockWeb-Bench supports studying both the websites hybrid agents build and the strategies they use to build them. Our code and dataset are available at https://anonymous.4open.science/r/mockweb-bench-208C/README.md.

Primary area:
datasets and benchmarks
Keywords:
Coding Agents, Computer Use Agents, Website Reconstruction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_3RvUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces MockWeb-Bench, a benchmark where agents try to reconstruct a working website. The benchmark consists of 50 offline reference websites and 4,101 evaluation pages, with desktop and mobile views. It evaluates the results on functionality, visual similarity, code quality, and structure. The authors test ten model-agent configurations and analyze their behavior.

Critical strengths
  1. Agents can browse working websites while rebuilding them, checking how the pages look and how different features work. This tests both browsing and coding abilities, rather than just copying screenshots.

  2. The evaluation is designed carefully. It checks whether reconstructed websites actually work, alongside their visual similarity. The authors also validate the functional tests on the original websites and evaluate ten model-agent configurations.

Critical weaknesses
  1. The paper emphasizes interactive reference exploration as its main contribution, but provides limited evidence of what additional insights this setting offers compared to existing benchmarks. It also does not directly measure its added value, for example by comparing reconstruction performance with and without interactive exploration.

  2. The authors check that the generated tests work on the original websites, but they do not show how much of each website's functionality the tests cover. Important interactions could still be missed.

Questions for authors
  1. What additional insights does interactive reference exploration provide over existing benchmarks?

  2. How much of the website's important functionality is covered by the final tests?

  3. Since each system was tested only once per website, how reliable are the rankings, especially when the top systems have very similar scores?

3 comments

Overall Response

We appreciate the reviewer’s constructive feedback on the task setting and evaluation. Concretely, we have (i) clarified why recreation provides a verifiable task for studying sustained GUI–CLI coordination and highlighted the existing trajectory findings, (ii) added a reference-first component audit of 60 target pages from 30 websites, documenting behavioral checks for 44 of 48 reference-confirmed core groups, and (iii) included the repeated-run protocol and close-configuration comparison for GLM-5.3 and Claude Opus 5. These explanations, audit criteria, and comparison tables are now documented in the manuscript and appendix.

Response to W1 & Q1: Insights into hybrid capabilities

We appreciate the opportunity to clarify the hybrid capability this task is designed to study. Our objective is to analyze sustained GUI–CLI coordination. Many real-world agent tasks, including office workflows, data processing, and application use, combine GUI observation and interaction with CLI file operations, program execution, and automation. We study how agents gather information, choose GUI or CLI actions, and use observations and execution feedback to guide subsequent work.

We use website recreation as a verifiable task for this analysis: the running reference supplies explicit ground truth for appearance and supported frontend behavior. The reference anchors evaluation of the implementation, while development trajectories show how GUI exploration and CLI work are coordinated. Compared with the fixed target materials in Vision2Web and WebCompass, our setting lets agents choose which pages, visual details, responsive states, and interactions to inspect throughout development.

Our analyses reveal three findings:

  1. Visual fidelity can coexist with missing behavior: 16 of 35 uncapped outputs with Visual ≥80% and recorded interaction results score below 50% on interactions.
  2. Coding continues during refinement: across eight eligible Claude Code configurations, a median 26.4% of source-write calls occur after source size first reaches 90% of its final value.
  3. Agents use code to quantify GUI feedback: Astra's recorded RGB MAE across eight mobile routes falls from 22.40 to 8.09 during refinement.

This evaluation purpose is explicit in the Introduction and task definition, and the analysis section presents these findings and their implications for hybrid coordination.

Response to W2 & Q2: Coverage of important website functionality

Thank you for highlighting coverage of important interactive components. We conducted a reference-first component audit of 60 target pages from 30 websites spanning all ten categories. Core components were identified independently of the tests by their role in primary navigation, content access, or principal frontend operations. Shared widgets and sibling tabs or FAQ entries are grouped, while their instances and behavior variants remain explicit. Coverage requires a test to operate the component and check its observable response.

The existing fixed suite checks the defining operations of most reference-confirmed core component groups in this audit:

Core component coverage in the audited sampleCount
Reference-confirmed core component groups48
Groups with behavioral checks in the fixed suite44 / 48 (91.7%)
Groups without those defining-operation checks4 / 48 (8.3%)

The four localized omissions concern command copying, installation-method navigation, a product-information tab group, and a FAQ accordion. The existing evaluation jointly assesses content and navigation, tested interaction outcomes, visual fidelity, quality, and structure, providing a broad assessment of overall reconstruction alongside this substantial core-component coverage. The appendix now reports the audit design, coverage results, and their interpretation.

Response to Q3: Reliability of close rankings

We agree that close scores warrant careful examination of run-to-run variability. We selected GLM-5.3 and Claude Opus 5, whose reported mean scores are close, for an additional repeated-run comparison. Each configuration was independently evaluated three times on each of the same 50 websites, including the original evaluation. Each run started from a fresh workspace, with the same task settings, tools, and time budget.

ConfigurationOriginal Final (%)Three-run mean Final (%)Mean within-website SD (pp)
GLM-5.353.2754.193.15
Claude Opus 552.7552.202.62

The paired mean difference between the per-website three-run averages is 1.99 percentage points, with a 95% confidence interval of [-5.32, 6.90]. The repeated-run means remain close to the original results, while the interval includes zero. We therefore characterize these configurations as having comparable performance and report their observed mean scores without claiming a statistically resolved ordering. The appendix now includes the repeated-run protocol, comparison table, and aggregate-score stability analysis.

AuthorsOpenReview ↗

Kind Reminder: Your Additional Thoughts on Submission 11898 Are Welcome

Dear Reviewer 3RvU,

Thank you very much for your valuable comments and for taking the time to consider our rebuttal. We understand the discussion period can be demanding on your schedule, and we would welcome the chance to clarify any remaining points before it closes on Nov 18.

We hope our responses have addressed your questions and concerns so far. If anything remains unclear, or if you have further concerns, we would be very glad to follow up and will do our best to clarify promptly.

Thank you again for your feedback; it has genuinely helped us improve the paper.

Warm regards, The Authors of Submission 11898

Reviewer_C1jnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  1. https://huggingface.co/datasets/Qwen/RecreationBench/viewer/web is identical to https://huggingface.co/datasets/submission11898/MockWeb-Bench/tree/main/v3_relicensed_260825-assets

  2. https://anonymous.4open.science/r/mockweb-bench-208C/mockweb_public/native_core/evaluation/scrape_gate.py is identical to https://github.com/QwenLM/RecreationWorld/blob/main/scripts/web/evaluation/scrape_gate.py

However, Submission11898 does not cite https://huggingface.co/datasets/Qwen/RecreationBench or RecreationWorld (https://arxiv.org/abs/2609.22000) or mention RecreationWorld. As this is a benchmark paper, adopting a subset of an existing paper's data and code raises questions about novelty.

According to Raw=0.40 sfunc + 0.40 svis + 0.10 squal + 0.10sstruct. (Equation 1, this paper's page 6), the weights are 0.40, 0.40, 0.10, and 0.10; however, according to https://anonymous.4open.science/r/mockweb-bench-208C/mockweb_public/controller.py (line 38), the weights are _WEIGHTS = {"functional": 0.35, "visual": 0.35, "quality": 0.15, "structural": 0.15}. These two weights are not aligned.

Critical strengths

MockWeb-Bench is identical to the subset of https://huggingface.co/datasets/Qwen/RecreationBench

Critical weaknesses

MockWeb-Bench is identical to the subset of https://huggingface.co/datasets/Qwen/RecreationBench

Questions for authors

N/A

Reviewer_8pvAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MockWeb-Bench evaluates hybrid GUI and coding agents through 50 reference websites, with 4,101 evaluation pages and desktop/mobile views. Agents must discover and implement the reference website's appearance and supported frontend behavior. Evaluation combines functional tests, VLM-as-judge, quality and structural checks, and copying penalty. The benchmark tests ten model–agent configurations, and the best mean final score is 56.26%. The paper also analyzes development trajectories and visual refinement strategies.

Critical strengths
  1. The reference introduces a meaningful requirement-discovery component. Allowing the development agent to choose which pages and states to inspect creates an opportunity to study how observation, implementation, and verification are coordinated under a finite budget.

  2. The benchmark construction and supporting diagnostics are substantial. The collection provides deep reconstruction targets with desktop/mobile views. Validating tests on the references before freezing them, auditing copying, examining trajectories, and reporting evaluation sensitivity and human agreement are useful steps toward an inspectable benchmark.

Critical weaknesses
  1. Overlap with existing benchmarks: Vision2Web and WebCompass already evaluate interactive frontend generation with functional and visual assessment. A controlled comparison is missing to show what additional capability this setting measures.

  2. Limited amount of website and page diversity: The benchmark covers 50 websites and 4,101 pages, but the diversity of reconstruction scenarios is not established. It does not quantify page-template redundancy or coverage of distinct layouts and interaction patterns.

  3. Insufficient validation of scoring and judging: a) The paper specifies weights of 0.40/0.40/0.10/0.10, while the attached code use 0.35/0.35/0.15/0.15. b) The sensitivity study supports ranking stability, but does not validate the weights or Quality/Structural heuristics. c) Visual evaluation relies on Qwen3.7-Plus, with human validation covering only 48 image pairs from 12 sites and two configurations.

Questions for authors
  1. What capabilities does MockWeb-Bench evaluate beyond Vision2Web and WebCompass, and what evidence isolates the contribution of executable-reference exploration?

  2. Can you characterize the distinct layouts and interaction patterns covered by the 4,101 pages, including template reuse within and across the 50 websites?

  3. What evidence supports the weights and Quality/Structural heuristics? Which weighting scheme and code version produced the reported results, given the difference between the paper and the public release?

4 comments

Overall Response

Thank you for the careful reading and specific suggestions on the task design, diversity, and scoring. Concretely, we have (i) clarified the distinction from Vision2Web and WebCompass and how agents autonomously discover requirements through reference exploration, (ii) quantified layout diversity and repetition across all 4,101 target pages and characterized the interaction patterns represented in the tests, and (iii) corrected the public aggregation defaults and added the human-rating criteria, automatic–human score comparisons, and constrained weight-fit description. The manuscript and appendix now include these clarifications and supporting analyses.

Response to W1 & Q1: Capabilities beyond existing benchmarks

We appreciate this comparison with closely related benchmarks. The evaluation emphasis differs: Vision2Web and WebCompass assess website generation or reconstruction from supplied target materials, while MockWeb-Bench uses recreation to evaluate sustained coordination between GUI exploration and CLI development.

This focus is reflected in the task inputs. Vision2Web provides prototypes, requirements, and assets; WebCompass provides text specifications, screenshots, or interaction recordings. These materials describe the target in advance. MockWeb-Bench provides an entry URL to an executable reference and asks agents to discover the visual and behavioral requirements through their own exploration, while keeping the scored-page inventory hidden.

Agents decide which pages to visit, what to inspect in screenshots, which layouts and content to examine, and which responsive states or interactions to explore. They gather and revisit evidence as needed, use it to guide CLI implementation, and verify the rebuilt website through GUI interaction. Recreation gives this hybrid capability explicit ground truth through the reference's appearance and supported frontend behavior. The Introduction and task definition state this evaluation focus and autonomous requirement-discovery process explicitly.

Response to W2 & Q2: Layout diversity and template reuse

We appreciate the request to quantify layout diversity and repetition. MockWeb-Bench covers diverse website scenarios, page layouts, and interaction patterns.

We compare page structure and desktop/mobile layouts across all 4,101 pages together, and automatically cluster them into layout-similarity categories at a distance cutoff of 0.10.

Dataset diversity characterizationResult
Distinct source websites / website categories50 / 10
Quality-verified target pages4,101
Automatic layout-similarity categories across all pages2,380
Single-page layout categories / all layout categories1,812 / 2,380 (76.1%)
Interaction patterns represented in released checks12 overlapping categories

The global category count and single-page category share characterize detailed layout variation under this comparison rule. Across sites, broad hero or article layouts can recur with different proportions, typography, and mobile arrangements.

We report interaction-pattern coverage by the number of websites with corresponding checks. A website is counted once per pattern when at least one check pairs a user action with an expected navigation outcome, content change, or control-state change. The six patterns most widely represented in the released checks are:

Pattern represented in released checksWebsites / 50
Page/route navigation50
Menus, dropdowns, and drawers38
Responsive/mobile controls36
Disclosures and expandable content24
Input values and choices23
Tab and panel switching20

Patterns can overlap within a website. The appendix now describes the layout comparison rule and reports the global diversity statistics and the six most widely represented interaction patterns.

Response to W3 & Q3: Weights, scoring measurements, and human validation

Thank you for pointing out the aggregation-weight discrepancy and asking for broader score validation. The initial public release inadvertently retained an earlier set of default aggregation weights. All reported scores were computed with 0.40/0.40/0.10/0.10, consistent with Eq. (1). We have updated the public defaults to match the weights in Eq. (1).

The allocation reflects the task's priorities: Functional and Visual receive 40% each because recreation targets both behavioral and visual fidelity. Quality and Structural receive 10% each as complementary criteria for basic page quality/accessibility and preservation of content/navigation organization.

To assess whether this weight allocation is consistent with human judgments, we include an additional human evaluation of 30 reference–candidate page pairs, sampled across ten website categories and all ten model–agent combinations evaluated in this paper. Three raters assess the same 30 pairs, using matched views and reference-supported interactions. Each independently rates the four dimensions and assigns a direct overall-quality score on the same scale. The overall score is not calculated from the component ratings. The appendix describes a constrained least-squares model using the four mean human component ratings to predict the mean overall rating. Here, sfuncs_{\mathrm{func}}, sviss_{\mathrm{vis}}, squals_{\mathrm{qual}}, and sstructs_{\mathrm{struct}} denote the three-rater means. The model uses one weight vector across all 30 pairs, with nonnegative coefficients summing to one:

Overall≈0.37×sfunc+0.48×svis+0.09×squal+0.06×sstruct.\mathrm{Overall} \approx 0.37 \times s_{\mathrm{func}} + 0.48 \times s_{\mathrm{vis}} + 0.09 \times s_{\mathrm{qual}} + 0.06 \times s_{\mathrm{struct}}.

The fitted allocation is consistent with our design rationale: behavioral and visual fidelity receive the primary emphasis, while quality and structural fidelity provide complementary criteria. This supports the reasonableness of the weight allocation used in the paper.

The same additional evaluation also assesses agreement between the automatic component scores and human judgments. The existing visual pilot gives Spearman ρ=0.812\rho = 0.812 [0.634, 0.916] against mean human visual-fidelity ratings.

The human Quality criteria cover readability, labels and control usability, keyboard accessibility, and visible errors. The Structural criteria cover preservation of the reference's content hierarchy, navigation, main regions, and responsive organization. The appendix documents comparisons with the corresponding automatic scores for all four dimensions:

For each page pair and dimension, the analysis averages the three human ratings and calculates Spearman correlation with the automatic score for the same page and evaluated interactions.

ScoreSpearman ρ\rho [95% CI]
Functional0.823 [0.711, 0.894]
Visual0.818 [0.733, 0.886]
Quality0.929 [0.820, 0.974]
Structural0.934 [0.846, 0.968]

We have added the sampling protocol, rating criteria, correlation table, and constrained-fit description to the appendix.

AuthorsOpenReview ↗

Kind Reminder: Your Additional Thoughts on Submission 11898 Are Welcome

Dear Reviewer 8pvA,

Thank you very much for your valuable comments and for taking the time to consider our rebuttal. We understand the discussion period can be demanding on your schedule, and we would welcome the chance to clarify any remaining points before it closes on Nov 18.

We hope our responses have addressed your questions and concerns so far. If anything remains unclear, or if you have further concerns, we would be very glad to follow up and will do our best to clarify promptly.

Thank you again for your feedback; it has genuinely helped us improve the paper.

Warm regards, The Authors of Submission 11898

Dear Reviewer 8pvA,

The discussion period ends tomorrow. We would be grateful if you could let us know whether our rebuttal has addressed your main concerns. We would be happy to clarify any remaining points.

Thank you for your consideration.

Best regards,

The Authors of Submission 11898

Reviewer_TUgAOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

his paper introduces MockWeb-Bench, a benchmark in which agents autonomously explore a continuously running reference website without access to its source code and reconstruct its pages, interaction states, and responsive layouts. It therefore evaluates agents’ capabilities in GUI exploration, requirement discovery, code implementation, and iterative verification. The benchmark comprises 50 complete websites, 4,101 manually verified evaluation pages, and paired desktop and mobile views. The specific pages used for scoring are hidden from the agents, reducing the risk that they implement only the publicly known test cases. The benchmark assesses agents’ ability to gather and reproduce information related to functionality, visual appearance, code quality, and page structure.

Critical strengths
  1. The task setting closely reflects real-world usage and development workflows. Agents must autonomously explore a continuously running reference website without access to its source code. This setting provides a comprehensive assessment of agents’ capabilities in GUI interaction, code implementation, and iterative debugging, while also offering practical value for real-world website construction and reconstruction tasks.

  2. The benchmark provides a comprehensive evaluation framework. It assesses functional correctness, visual fidelity, code and accessibility quality, and page structure.

  3. MockWeb-Bench offers broad evaluation coverage and detailed trajectory analysis. It includes 50 complete websites and 4,101 manually verified evaluation pages, covering multiple pages, interaction states, and screen sizes. The paper also analyzes the complete execution trajectories of various model–agent configurations, characterizes how different systems allocate their effort among exploration, coding, and verification, and identifies potential shortcut behaviors, such as copying reference code and attempting to access protected files. These findings provide useful guidance for designing more reliable agent benchmarks in future work.

Critical weaknesses
  1. The paper does not provide sufficient evidence that continuous access to a running reference website, together with the agent’s ability to autonomously select what to inspect, is more effective than using screenshots, videos, or a fixed set of reference pages. Additional comparative experiments under these different input conditions would help validate the benefits of the proposed task setting.

  2. The human validation of the visual evaluation is based on a limited sample and covers relatively few website categories. The human validity of the overall score therefore remains unclear. The current human study validates only the visual score and does not examine whether the functional, structural, and quality scores, or the weighted Final score combining all four dimensions, align with human judgments of overall reconstruction quality.

Questions for authors

See the weakness

4 comments

Overall Response

We are grateful for the reviewer’s careful assessment of the task setting and evaluation. Concretely, we have (i) clarified how the running reference supports autonomous requirement discovery and sustained GUI–CLI coordination, (ii) added the 30-page-pair human-rating protocol and comparisons for the component scores and overall reconstruction quality, and (iii) documented a constrained fit relating the human component ratings to independently assigned overall judgments to examine the weight allocation. These explanations, rating criteria, and comparison tables are now included in the manuscript and appendix.

Response to W1: Why a running reference is part of the hybrid task

Thank you for asking about the role of continuous reference access. The evaluation emphasis differs: Vision2Web and WebCompass assess website generation or reconstruction from supplied target materials, while MockWeb-Bench uses recreation to evaluate sustained coordination between GUI exploration and CLI development.

This focus is reflected in the task inputs. Vision2Web provides prototypes, requirements, and assets; WebCompass provides text specifications, screenshots, or interaction recordings. These materials describe the target in advance. MockWeb-Bench provides an entry URL to an executable reference and asks agents to discover the visual and behavioral requirements through their own exploration, while keeping the scored-page inventory hidden.

Agents decide which pages to visit, what to inspect in screenshots, which layouts and content to examine, and which responsive states or interactions to explore. They gather and revisit evidence as needed, use it to guide CLI implementation, and verify the rebuilt website through GUI interaction. Recreation gives this hybrid capability explicit ground truth through the reference's appearance and supported frontend behavior. This task motivation and autonomous requirement-discovery process are stated explicitly in the Introduction and task definition.

Response to W2: Human validation of component and overall scores

We agree that human validation should address both the component scores and overall reconstruction quality. The existing visual pilot gives Spearman ρ=0.812\rho = 0.812 [0.634, 0.916] against mean human visual-fidelity ratings.

We include an additional evaluation of 30 reference–candidate page pairs across ten website categories and all ten evaluated model–agent combinations. Three raters assess the same pairs, using matched views and reference-supported interactions. Each independently assigns four component ratings and a direct overall-quality rating on the same scale. The overall rating is not calculated from the components.

Human Quality ratings assess readability, labels and control usability, keyboard accessibility, and visible errors. Human Structural ratings assess preservation of the reference's content hierarchy, navigation, main regions, and responsive organization. For each page and dimension, we average the three ratings and compare them with the corresponding automatic score for the same page and evaluated interactions. We also compare the automatic page-level composite using the paper weights with the mean directly assigned overall rating.

ScoreSpearman ρ\rho [95% CI]
Functional0.823 [0.711, 0.894]
Visual0.818 [0.733, 0.886]
Quality0.929 [0.820, 0.974]
Structural0.934 [0.846, 0.968]
Overall0.914 [0.842, 0.951]

To examine the weight allocation, we use the same human component and overall ratings in a constrained least-squares fit. In this fit, sfuncs_{\mathrm{func}}, sviss_{\mathrm{vis}}, squals_{\mathrm{qual}}, and sstructs_{\mathrm{struct}} denote the three-rater human means, and the nonnegative coefficients sum to one. The fitted relation is expressed as:

Overall≈0.37×sfunc+0.48×svis+0.09×squal+0.06×sstruct.\mathrm{Overall} \approx 0.37 \times s_{\mathrm{func}} + 0.48 \times s_{\mathrm{vis}} + 0.09 \times s_{\mathrm{qual}} + 0.06 \times s_{\mathrm{struct}}.

This analysis checks whether the relative emphasis in human judgments is consistent with the paper's functional/visual priorities. We have added the sampling protocol, rating criteria, component and overall-score comparisons, and weight-fit description to the appendix.

AuthorsOpenReview ↗

Kind Reminder: Your Additional Thoughts on Paper Submission 11898 Are Welcome

Dear Reviewer TUgA,

Thank you very much for your valuable comments and for taking the time to consider our rebuttal. We understand the discussion period can be demanding on your schedule, and we would welcome the chance to clarify any remaining points before it closes on Nov 18.

We hope our responses have addressed your questions and concerns so far. If anything remains unclear, or if you have further concerns, we would be very glad to follow up and will do our best to clarify promptly.

Thank you again for your feedback; it has genuinely helped us improve the paper.

Warm regards, The Authors of Submission 11898

I share Reviewer 8pvA’s concern that Overlap with existing benchmarks:. I have therefore lowered my score by one point.

Dear Reviewer TUgA,

Thank you for the follow-up. We would like to clarify the specific capability distinction underlying our contribution.

Table 1 compares the target inputs: VisionWeb and WebCompass provide prototypes, specifications, screenshots, or recordings. MockWeb-Bench provides an entry URL to an executable reference while withholding the scored-page inventory. Agents must decide which pages, responsive states, and interactions to inspect, and coordinate this information gathering with ongoing implementation and verification.

Our contribution is to evaluate active requirement discovery and sustained GUI-CLI coordination against an executable reference. Our trajectory analyses also examine this coordination in practice, including agents writing screenshot-comparison programs to guide visual refinement (Section 5.3 and Appendix F.2).

Could you clarify which aspect of this requirement-discovery setting you consider already covered by Vision2Web or WebCompass, and whether your remaining concern is specifically the controlled input comparison requested in W1? This would help us address the precise basis of your novelty concern.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.