Pith. sign in

REVIEW 2 major objections 6 minor 50 references

VISCA: Inferring Component Abstractions for Automated End-to-End Testing

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read VISCA's component abstraction lifts LLM end-to-end test coverage to 92%, surpassing the 79% reported for the AUTOE2E baseline.

desk verdict VISCA is a well-engineered pipeline for turning webpages into component abstractions for LLM-based E2E testing, but the headline 16% coverage gain over AUTOE2E rests on a baseline number that was never re-run in the same harness. read the letter →

arxiv 2506.04161 v1 pith:5TITJLX3 submitted 2025-06-04 cs.SE

Parsa Alian , Martin Tang , Ali Mesbah This is my paper
classification cs.SE
keywords end-to-endtestingLLMtestgenerationUIcomponentabstractionwebpagesegmentationfeatureinferenceDOMtreemultimodalvocabulary
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

VISCA claims that the main obstacle to LLM-based end-to-end test generation is not the LLM itself but the context it receives: isolated HTML action snippets are too narrow, and full-page HTML is too broad. The paper proposes replacing both with a component abstraction of the page, built from a fixed vocabulary of 50 UI component templates, and feeding that abstraction to the LLM for feature inference and test generation. On the eight applications of the E2EBench suite, VISCA reports an average feature coverage of 92%, a 16-percentage-point improvement over the 79% reported for AUTOE2E, with equal-or-better coverage on every application. The paper also claims that the underlying abstraction faithfully preserves the original page: about 91.5% of abstracted components fully match ground-truth UI components in both structure and content.

What carries the argument

The load-bearing object is the component abstraction itself: a DOM-derived hierarchy whose nodes are drawn from a curated vocabulary of 50 common UI templates. It is produced by a four-stage pipeline: visibility-based pruning of redundant single-child nodes; a candidate-segment selection algorithm that maximizes the potential function $\Psi_n = \log\left(\frac{\text{size}(n)}{1 + \sum_i \text{dist}(n, \text{sib}_i(n))}\right)$ using tree edit distance; a top-down multimodal LLM classification into Containers, Lists, and Components with ancestor context feeding each description; and a transformation step that maps Component segments onto the 50-template vocabulary, representing Lists by a single sample child. This abstraction is what replaces action HTML in the downstream feature-inference and test-generation step.

What would settle it

Re-run AUTOE2E on E2EBench under exactly the same conditions as VISCA, using the same multimodal LLM at temperature 0 and the same feature definitions; if the baseline then reaches or exceeds 92%, the claimed 16% improvement is an artifact of differing setups rather than of the component abstraction.

Watch

Extended reading notes

Core claim

VISCA's central claim is that a webpage's optimal context for LLM-based end-to-end test generation is the UI component, not the raw HTML action snippet and not the whole page. To support this, VISCA builds a hierarchical component abstraction: it prunes the DOM, selects candidate segments using a potential function that rewards large subtrees similar to their siblings, classifies each candidate as a Container, List, or Component through multimodal LLM analysis, and transforms Component segments into a vocabulary of 50 common templates such as Card, Navbar, and Tab. The resulting abstraction replaces isolated action HTML in the feature-inference and test-generation step, and the paper reports that this yields 92% average feature coverage against the 79% reported for AUTOE2E, alongside a segmentation precision of 83.4% versus 65.7% for the VIPS baseline.

Load-bearing premise

The load-bearing premise is that AUTOE2E's reported 79% feature coverage is directly comparable to VISCA's 92%, which requires that the two were measured under the same LLM, temperature, execution environment, and feature definitions.

Editorial extensions

If this is right

  • If the component abstraction is the right contextual unit, LLM-based feature inference should improve on web applications beyond the eight E2EBench subjects, since the vocabulary is framework-agnostic.
  • The high reported fidelity of the abstraction (91.5% full structural and content match, 98.2% completeness) suggests the same component model could serve other downstream tasks such as documentation generation, accessibility analysis, or UI code synthesis.
  • Because VISCA matches or beats AUTOE2E on every benchmark application, the paper's design implies that the gain comes from the context representation itself rather than from a more powerful feature-inference prompt.
  • The 50-template vocabulary acts as a reusable descriptive language for webpages, meaning pages not built with component frameworks can still be analyzed at component granularity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the 16-point coverage gap could shrink if AUTOE2E were re-run under the same LLM, temperature, and environment settings as VISCA; the paper compares against the 79% value reported in the original AUTOE2E paper rather than a same-run baseline.
  • An untested implication is that the component vocabulary, not the segmentation step, drives the improvement; an ablation that feeds VISCA's raw segments into the same feature-inference framework without transformation would isolate which stage contributes the gain.
  • VISCA's segmentation shows much lower recall than VIPS (23.3% versus 60.8%), so on pages without repeated, sibling-similar structures the method may produce fewer usable component candidates and the coverage advantage could diminish.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. VISCA (Visual-Semantic Component Abstractor) is a method for transforming webpages into a hierarchical component abstraction and using that abstraction as the context for LLM-based end-to-end test generation. The pipeline (Section III) prunes the DOM, applies a heuristic potential function (Eq. 1) that favors large, sibling-similar subtrees to extract candidate segments, classifies candidates as Containers/Lists/Components using a multimodal LLM with chain-of-thought prompting, extracts titles and natural-language descriptions, and maps Components onto a hand-curated vocabulary of 50 common templates drawn from five UI libraries. The resulting abstraction replaces the raw action HTML used by AUTOE2E as input to feature inference and test-case generation. Evaluations report 91.5% structural and 91.5% content accuracy for component abstraction on 26 showcase pages from three UI libraries (RQ1); PB3 precision of 83.4% versus VIPS's 65.7% for segmentation, at lower recall and F1 (RQ2); and 92% average feature coverage on the E2EBench benchmark versus a reported 79% for AUTOE2E, with global precision 0.40, recall 0.83, and F1 0.54 (RQ3). The paper's central claim is that component-level context yields more accurate feature inference and higher feature coverage than raw HTML snippets.

Significance. The paper's central idea is credible and timely: LLM context, not just the model, determines feature-inference quality in automated E2E testing, and the right granularity of abstraction may lie at the level of UI components. If the controlled comparison bears out, VISCA would be a useful advance over AUTOE2E and related approaches, and the 50-template vocabulary plus the segmentation heuristics could serve as a reusable basis for richer UI test generation. Strengths include a complete, end-to-end pipeline evaluated on a standard benchmark (E2EBench, 8 apps, 645 generated tests); transparent reporting of the precision/recall trade-off in segmentation; explicit documentation of temperature and reasoning-mode settings; and human validation of the LLM-judged RQ1 labels. The main reservations are experimental: the headline 16% coverage gain is measured against a previously reported number rather than a same-harness re-run, and the RQ1 ground truth shares its vocabulary source with the abstraction's template set. These issues are fixable with additional experiments, and the current evidence already supports a plausible, though not yet conclusive, case for the method's effectiveness.

major comments (2)
  1. [IV-C (Figure 7, Table III)] The headline claim of a 16% improvement over AUTOE2E is not a controlled comparison. The paper compares VISCA's measured 92% average feature coverage against 'the reported 79% for AUTOE2E' (Section IV-C, Figure 7), but it never states that AUTOE2E was re-run under identical conditions in the same harness; the two numbers may differ in LLM version, temperature, prompting scheme, execution environment, benchmark application versions, and feature ground truth. The confounding is material because VISCA generates 645 test cases across the eight applications (Table III), while the paper does not report AUTOE2E's test-generation budget for comparison; a larger inference budget can raise recall mechanically. The same table shows VISCA's global precision of 0.40 and F1 of 0.54 versus AUTOE2E's reported F1 of 0.62, so the coverage gain is accompanied by a large increase in false-positive feature inferences, which a coverage-only headline does not convey. To support the causal claim that the component abstraction, rather than the model or the budget, drives the improvement, please re-run AUTOE2E under the same conditions (same LLM, temperature, apps, and feature lists) and include an ablation of VISCA that replaces the component abstraction with raw action HTML while holding the rest of the pipeline fixed.
  2. [IV-A (Table I)] The RQ1 evaluation of the 91.5% structural/content accuracy is partially circular and is likely to overstate generalizability. The 50-template vocabulary was curated from the five libraries in Table I (Section III-A), and the 26 ground-truth subjects are showcase pages from three of those same libraries (Material UI, Shadcn, Bootstrap), with library-specific component types serving as the ground truth. The abstraction is therefore rewarded for recognizing a vocabulary that was induced from the very population being tested; this measures vocabulary fit rather than the ability to abstract arbitrary pages, which is the paper's motivating scenario (the Amazon cart example). In addition, the match categories are assigned by Gemini 2.5 Flash in reasoning mode, the same model family that generated the abstractions, with author validation reported but no inter-rater agreement statistic. Please re-run RQ1 with held-out libraries (for example, Ant Design or Chakra UI from Table I, or a library not used in vocabulary induction) and report agreement statistics on the LLM-judged labels. The threat-to-validity paragraph on subject representativeness discusses page redundancy but does not address this vocabulary overlap.
minor comments (6)
  1. [IV-B / Abstract] The abstract-level claim that VISCA 'outperforms the state-of-the-art VIPS algorithm by 26.9% in this regard' holds only for precision (PB3 83.4% vs 65.7%); VISCA's recall (RB3 23.3%) and F1 (FB3 27.2%) are substantially below VIPS's best values (60.8% and 48.5%). Since Section IV-B and Figure 6 present these numbers transparently, the issue is presentation; please qualify the claim in the Abstract and Introduction as a precision-only improvement, or add the downstream justification from Table II there.
  2. [Table III / Figure 7] The application name appears as 'EverTraduora' in Table III but as 'Traduora' in Figure 7 and in the text of Section IV-C; the baseline name also alternates between 'AUTOE2E' and 'AutoE2E' (Figure 7). Please standardize both.
  3. [III-A2 (Algorithm 2)] In Algorithm 2, lines 7-10, all children of a node classified as a List are assigned the Component class without LLM verification, so a misclassification of a Container as a List propagates to the entire subtree; moreover, the recursion stops at any Component (line 11), so nested Components or Lists inside a Component are never classified, which is in tension with the composition relation shown in Figure 4. The paper does not report the three-way classification accuracy of the LLM classifier, leaving this error-propagation path unmeasured; please report classification accuracy and briefly justify the forced List-to-Component assignment.
  4. [III-A3 / IV-A] The contextual titles and natural-language descriptions (e.g., Listing 3) are the claimed source of the semantic enrichment that improves feature inference, yet RQ1 evaluates only structural accuracy and content preservation of Components; there is no evaluation of whether the extracted titles and descriptions match developer-assigned component names or page semantics. A small human-graded assessment of the extracted context would directly test the paper's central motivation.
  5. [III-A4] When transforming a List, VISCA generates the abstraction for a single representative child and reuses it for all items; please state explicitly how distinct per-item features (e.g., deleting item 1 versus item 2, or editing different list entries) are handled in feature inference and coverage counting, since collapsing all items to one representative could merge genuinely distinct features or, conversely, cause per-item test cases to be generated from a single abstraction.
  6. [Availability] No code or data artifact is provided, and the 50-template vocabulary is not listed in the paper; for reproducibility of the RQ3 run (645 test cases) and the RQ1 corpus (26 pages, 386 components), please provide the vocabulary list and an artifact link, or add an appendix containing the template names and the subject-page list.

Circularity Check

2 steps flagged · score 3.0 of 10

RQ1's ground-truth abstractions come from the same Table I libraries that define VISCA's 50-template vocabulary, and the same Gemini 2.5 Flash both generates and grades the abstractions; Table II self-measures segmentation via VISCA's own classifier.

  1. self definitional [Section IV-A (RQ1); template curation in Section III.A; Table I]
    "To evaluate the accuracy of our component abstraction module, the core of VISCA, we need to compare its generated component-based abstraction against a ground-truth abstraction that also utilizes our defined common templates. ... Among the five UI libraries we initially considered (Table I), three provided open-source demonstration pages suitable for our analysis: Material UI, Shadcn, and Bootstrap. These templates were systematically derived by analyzing widely-used UI design libraries (Table I)."

    RQ1's correctness target is defined as a ground-truth abstraction that also utilizes VISCA's own defined common templates, and that ground truth is taken from pages built with the same Table I libraries (Material UI, Shadcn, Bootstrap) from which the 50-template vocabulary itself was systematically derived. The answer key and the output space thus share a single provenance; a vocabulary curated from different libraries could plausibly lower the full-match rates. The judge is also Gemini 2.5 Flash, the same LLM family that generates the abstractions, with author re-checking as mitigation.

  2. other [Section IV-B, Table II; classification stage in Section III-A2]
    "To assess how well we achieve this goal, we evaluate the proportion of segments produced by our heuristic that are subsequently identified as Components during our multimodal LLM classification stage (Section III-A2)."

    The success metric for the segmentation heuristic is the share of its own output segments that VISCA's own multimodal-LLM classification stage (the same stage that consumes these segments and labels them Container/List/Component) calls Component. This measures internal agreement inside the VISCA pipeline rather than an external property of the segments, so the 74.5% Component rate is self-referential. The headline RQ2 numbers (PB3 83.4%, +26.9% over VIPS) are computed on the externally human-annotated dataset [33], so this self-measure is illustrative rather than load-bearing for that headline.

full rationale

VISCA's central RQ3 claim — 92% average feature coverage vs. the reported 79% for AUTOE2E (Figure 7) — is measured by running VISCA on the public E2EBench benchmark with the feature-coverage metric adopted from [11]; it is an independent measurement, not a quantity derived from fitted parameters or from VISCA's own prior outputs. No equation in the paper reduces to another: the potential function Psi_n (Eq. 1) has no parameters fitted to the evaluation set, and the test-generation framework is adapted from AUTOE2E rather than derived from VISCA's own results. The 16%-improvement framing rests on a number reported in the authors' own prior ICSE paper [11] without a same-harness re-run, but that is a comparability/validity risk for the correctness pass, not a circular reduction: the 79% value is externally falsifiable on the same public benchmark, and per the review rules a falsifiable external result counts as independent support. The partial circularity is confined to two supporting evaluations. RQ1 defines its ground truth as an abstraction that uses VISCA's own common templates, derives those templates from the same Table I libraries that furnish the subject pages, and grades matches with the same Gemini 2.5 Flash family that produces the abstractions. Table II similarly measures segmentation success by the share of VISCA's own segments that its own classification LLM labels Component. Neither is a by-construction equality — the pipeline demonstrably errs (2.9% no match; PB3 83.4% with RB3 23.3%) — so the result is partial, not total, circularity. Overall, the headline feature-coverage result is externally grounded while two supporting evaluations are partially self-referential; score 3.

Assumptions & free parameters 0 free parameters · 5 assumptions · 1 invented entities

The central claim rests on design assumptions about the expressiveness of the template vocabulary, the heuristic's ability to find component-like segments, and the reliability of multimodal LLM classification and evaluation. No numeric free parameters are fitted; the key choices are categorical and are listed as axioms.

assumptions (5)
  • domain assumption A predefined vocabulary of 50 common UI templates, curated from five open-source UI libraries, can express the structure and function of virtually any webpage.
    Section III.A posits that common templates can serve as a descriptive vocabulary for pages built with other technologies, e.g., Amazon's cart page.
  • ad hoc to paper Nodes with larger subtree size and higher structural similarity to siblings are likely component instances.
    Section III-A1, Eq. (1) and Algorithm 1: this heuristic underlies candidate segment extraction and is not independently validated before use.
  • domain assumption A multimodal LLM can reliably classify candidate segments into Container, List, or Component and map Component segments to the predefined vocabulary.
    Sections III-A2 and III-A4: the entire abstraction pipeline relies on this capability, and it is evaluated only on 26 library-built pages.
  • domain assumption LLM-based judgment (Gemini 2.5 Flash in reasoning mode) plus author validation reliably measures structural and content match between abstraction and ground truth.
    Section IV-A methodology: match outcomes are assigned by an LLM and then validated by authors; no fully independent metric is used.
  • domain assumption E2EBench feature annotations and the feature coverage metric are an appropriate and reliable measure of test generation quality.
    Section IV-C: coverage and correct-feature counts rely on the benchmark's feature definitions, and precision and recall are computed from them.
invented entities (1)
  • VISCA component abstraction (hierarchical structure over a hand-curated 50-template UI vocabulary)
    purpose: Replace raw HTML or whole-page context for LLM-based feature inference and E2E test generation.
    The abstraction is a software representation defined by the paper. It is compared to public library components in RQ1, but no artifact, code, or dataset is released to allow independent falsification outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VISCA: Inferring Component Abstractions for Automated End-to-End Testing." pith.science (2026). https://pith.science/paper/5TITJLX3

@misc{pith2026250604161,
  author       = {Pith},
  title        = {Pith review of: VISCA: Inferring Component Abstractions for Automated End-to-End Testing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5TITJLX3}},
  note         = {Machine review of arXiv:2506.04161}
}
read the original abstract

Providing optimal contextual input presents a significant challenge for automated end-to-end (E2E) test generation using large language models (LLMs), a limitation that current approaches inadequately address. This paper introduces Visual-Semantic Component Abstractor (VISCA), a novel method that transforms webpages into a hierarchical, semantically rich component abstraction. VISCA starts by partitioning webpages into candidate segments utilizing a novel heuristic-based segmentation method. These candidate segments subsequently undergo classification and contextual information extraction via multimodal LLM-driven analysis, facilitating their abstraction into a predefined vocabulary of user interface (UI) components. This component-centric abstraction offers a more effective contextual basis than prior approaches, enabling more accurate feature inference and robust E2E test case generation. Our evaluations demonstrate that the test cases generated by VISCA achieve an average feature coverage of 92%, exceeding the performance of the state-of-the-art LLM-based E2E test generation method by 16%.

Figures

Figures reproduced from arXiv: 2506.04161 by the authors.

Figure 1
Figure 1. Amazon’s shopping cart page 1 <span class="size-small action-delete"> 2 <input value="Delete" class="color-link"/> 3 </span> Listing 1: Delete action’s HTML from Amazon’s cart page Based on this action’s HTML alone, the precise target of the “Delete” is ambiguous. This snippet lacks the surrounding information necessary to determine its purpose. Furthermore, since multiple items are in the cart, their respective “De… view at source ↗
Figure 2
Figure 2. Amazon’s cart page segmented by VIPS [15] [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of VISCA We posit that an effective and balanced level of granularity for context can be found at the level of UI Components. Web application front-ends are built using HTML, CSS, and JavaScript. However, traditional development relying solely on these technologies often struggled with modularity, as HTML lacks inherent mechanisms for encapsulating the view and behavior of distinct application elements. Com… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Segment types’ composition relations of a multimodal LLM. This step determines the semantic role of each segment and identifies those suitable for trans￾formation into our target component abstraction. For each candidate segment, the multimodal LLM is provided with the…
Figure 6
Figure 6. Figure 6: Comparison of B-Cubed scores for VIPS and V [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Feature Coverage of VISCA and AUTOE2E on E2EBENCH (i.e., the proportion of all actual features across E2EBENCH that VISCA successfully identified). These aggregate scores yield a global F1-score of 0.54. Furthermore, a direct comparison of feature coverage achieved by …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages

  1. [1]

    Crawling ajax-based web applications through dynamic analysis of user interface state changes,

    A. Mesbah, A. Van Deursen, and S. Lenselink, “Crawling ajax-based web applications through dynamic analysis of user interface state changes,” ACM Transactions on the Web (TWEB), vol. 6, no. 1, pp. 1–30, 2012

    work page 2012

  2. [2]

    Diversity-based web test generation,

    M. Biagiola, A. Stocco, F. Ricca, and P. Tonella, “Diversity-based web test generation,” inProceedings of the 2019 27th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, 2019, pp. 142–153. 10

    work page 2019

  3. [3]

    Fragment-Based Test Generation for Web Apps,

    R. K. Yandrapally and A. Mesbah, “Fragment-Based Test Generation for Web Apps,”IEEE Transactions on Software Engineering, vol. 49, no. 3, pp. 1086–1101, 2023

    work page 2023

  4. [4]

    Dependency-Aware Web Test Generation,

    M. Biagiola, A. Stocco, F. Ricca, and P. Tonella, “Dependency-Aware Web Test Generation,” in2020 IEEE 13th International Conference on Software Testing, Validation and Verification (ICST), 2020, pp. 175–185

    work page 2020

  5. [5]

    A Reinforcement Learning Approach to Generating Test Cases for Web Applications,

    X. Chang, Z. Liang, Y . Zhang, L. Cui, Z. Long, G. Wu, Y . Gao, W. Chen, J. Wei, and T. Huang, “A Reinforcement Learning Approach to Generating Test Cases for Web Applications,” in2023 IEEE/ACM International Conference on Automation of Software Test (AST), 2023, pp. 13–23

    work page 2023

  6. [6]

    Test intention guided llm-based unit test generation,

    Z. Nan, Z. Guo, K. Liu, and X. Xia, “Test intention guided llm-based unit test generation,” in2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 2025, pp. 779–779

    work page 2025

  7. [7]

    Rug: Turbo llm for rust unit test generation,

    X. Cheng, F. Sang, Y . Zhai, X. Zhang, and T. Kim, “Rug: Turbo llm for rust unit test generation,” in2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 2025, pp. 634–634

    2025

  8. [8]

    What you see is what you get: Attention-based self-guided automatic unit test generation,

    X. Yin, C. Ni, X. Xu, and X. Yang, “What you see is what you get: Attention-based self-guided automatic unit test generation,”arXiv preprint arXiv:2412.00828, 2024

    arXiv 2024

Show all 50 references
  1. [9]

    A multi-agent approach for rest api testing with semantic graphs and llm-driven inputs,

    M. Kim, T. Stennett, S. Sinha, and A. Orso, “A multi-agent approach for rest api testing with semantic graphs and llm-driven inputs,”arXiv preprint arXiv:2411.07098, 2024

    2024 arXiv

  2. [10]

    Automated generation of accessibility test reports from recorded user transcripts,

    S. F. Huq, M. Tafreshipour, K. Kalcevich, and S. Malek, “Automated generation of accessibility test reports from recorded user transcripts,” in 2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 2024, pp. 534–546

    2025

  3. [11]

    Feature- driven end-to-end test generation,

    P. Alian, N. Nashid, M. Shahbandeh, T. Shabani, and A. Mesbah, “Feature- driven end-to-end test generation,” in2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 2025, pp. 678–678

    2025

  4. [12]

    “React,” https://react.dev/, 2025, accessed: 2025-04-10

    2025

  5. [13]

    “Vuejs,” https://vuejs.org/, 2025, accessed: 2025-04-10

    2025

  6. [14]

    “Svelte,” https://svelte.dev/, 2025, accessed: 2025-04-10

    2025

  7. [15]

    Vips: a vision-based page segmentation algorithm,

    D. Cai, S. Yu, J.-R. Wen, and W.-Y . Ma, “Vips: a vision-based page segmentation algorithm,” 2003

    2003

  8. [16]

    Web page dom node characterization and its application to page segmentation,

    G. Vineel, “Web page dom node characterization and its application to page segmentation,” in2009 IEEE International Conference on Internet Multimedia Services Architecture and Applications (IMSAA). IEEE, 2009, pp. 1–6

    2009

  9. [17]

    Repetition-based web page segmentation by detecting tag patterns for small-screen devices,

    J. Kang, J. Yang, and J. Choi, “Repetition-based web page segmentation by detecting tag patterns for small-screen devices,”IEEE Transactions on Consumer Electronics, vol. 56, no. 2, pp. 980–986, 2010

    2010

  10. [18]

    Site-independent template-block detection,

    A. Kołcz and W.-t. Yih, “Site-independent template-block detection,” in European Conference on Principles of Data Mining and Knowledge Discovery. Springer, 2007, pp. 152–163

    2007

  11. [19]

    Fully convolutional neural networks for newspaper article segmentation,

    B. Meier, T. Stadelmann, J. Stampfli, M. Arnold, and M. Cieliebak, “Fully convolutional neural networks for newspaper article segmentation,” in2017 14th IAPR International conference on document analysis and recognition (ICDAR), vol. 1. IEEE, 2017, pp. 414–419

    2017

  12. [20]

    Towards an improved vision-based web page segmentation algorithm,

    M. Cormer, R. Mann, K. Moffatt, and R. Cohen, “Towards an improved vision-based web page segmentation algorithm,” in2017 14th Conference on Computer and Robot Vision (CRV). IEEE, 2017, pp. 345–352

    2017

  13. [21]

    Mmdetection: Open mmlab detection toolbox and benchmark. arxiv 2019,

    K. Chen, J. Wang, J. Pang, Y . Cao, Y . Xiong, X. Li, S. Sun, W. Feng, Z. Liu, J. Xuet al., “Mmdetection: Open mmlab detection toolbox and benchmark. arxiv 2019,”arXiv preprint arXiv:1906.07155, 1906

    2019 arXiv

  14. [22]

    Page segmentation using visual adjacency analysis,

    M. Bajammal and A. Mesbah, “Page segmentation using visual adjacency analysis,”arXiv preprint arXiv:2112.11975, 2021

    2021 arXiv

  15. [23]

    An empir- ical comparison of web page segmentation algorithms,

    J. Kiesel, L. Meyer, F. Kneist, B. Stein, and M. Potthast, “An empir- ical comparison of web page segmentation algorithms,” inEuropean Conference on Information Retrieval. Springer, 2021, pp. 62–74

    2021

  16. [24]

    Bootstrap,

    “Bootstrap,” https://getbootstrap.com/, 2025, accessed: 2025-04-10

    2025

  17. [25]

    Material ui,

    “Material ui,” https://mui.com/, 2025, accessed: 2025-04-10

    2025

  18. [26]

    Ant design,

    “Ant design,” https://ant.design/, 2025, accessed: 2025-04-10

    2025

  19. [27]

    “Shadcn,” https://ui.shadcn.com/, 2025, accessed: 2025-04-10

    2025

  20. [28]

    Chakra ui,

    “Chakra ui,” https://chakra-ui.com/, 2025, accessed: 2025-04-10

    2025

  21. [29]

    Generating reusable web components from mockups,

    M. Bajammal, D. Mazinanian, and A. Mesbah, “Generating reusable web components from mockups,” inProceedings of the 33rd ACM/IEEE International Conference on Automated Software Engineering, 2018, pp. 601–611

    2018

  22. [30]

    Simple fast algorithms for the editing distance between trees and related problems,

    K. Zhang and D. Shasha, “Simple fast algorithms for the editing distance between trees and related problems,”SIAM journal on computing, vol. 18, no. 6, pp. 1245–1262, 1989

    1989

  23. [31]

    Mongodb atlas,

    “Mongodb atlas,” https://www.mongodb.com, 2024, accessed: 2025-04- 10

    2024

  24. [32]

    Selenium,

    “Selenium,” https://www.selenium.dev, 2024, accessed: 2025-04-10

    2024

  25. [33]

    A quantitative comparison of semantic web page segmenta- tion algorithms,

    R. Kreuzer, “A quantitative comparison of semantic web page segmenta- tion algorithms,” Master’s thesis, 2013

    2013

  26. [34]

    Web page segmentation revisited: evaluation framework and dataset,

    J. Kiesel, F. Kneist, L. Meyer, K. Komlossy, B. Stein, and M. Potthast, “Web page segmentation revisited: evaluation framework and dataset,” in Proceedings of the 29th ACM International Conference on Information & Knowledge Management, 2020, pp. 3047–3054

    2020

  27. [35]

    A comparison of extrinsic clustering evaluation metrics based on formal constraints,

    E. Amig ´o, J. Gonzalo, J. Artiles, and F. Verdejo, “A comparison of extrinsic clustering evaluation metrics based on formal constraints,” Information retrieval, vol. 12, pp. 461–486, 2009

    2009

  28. [36]

    Extracting logical hierarchical structure of html documents based on headings,

    T. Manabe and K. Tajima, “Extracting logical hierarchical structure of html documents based on headings,”Proceedings of the VLDB Endowment, vol. 8, no. 12, pp. 1606–1617, 2015

    2015

  29. [37]

    Dynamic web page segmentation based on detecting reappearance and layout of tag patterns for small screen devices,

    K. Rajkumar and V . Kalaivani, “Dynamic web page segmentation based on detecting reappearance and layout of tag patterns for small screen devices,” in2012 International Conference on Recent Trends in Information Technology. IEEE, 2012, pp. 508–513

    2012

  30. [38]

    A densitometric approach to web page segmentation,

    C. Kohlsch ¨utter and W. Nejdl, “A densitometric approach to web page segmentation,” inProceedings of the 17th ACM conference on Information and knowledge management, 2008, pp. 1173–1182

    2008

  31. [39]

    A systematic literature review on automatic website generation,

    T. Kaluarachchi and M. Wickramasinghe, “A systematic literature review on automatic website generation,”Journal of Computer Languages, vol. 75, p. 101202, 2023

    2023

  32. [40]

    pix2code: Generating code from a graphical user interface screenshot,

    T. Beltramelli, “pix2code: Generating code from a graphical user interface screenshot,” inProceedings of the ACM SIGCHI symposium on engineering interactive computing systems, 2018, pp. 1–6

    2018

  33. [41]

    Controllable gui exploration,

    A. Garg, Y . Jiang, and A. Oulasvirta, “Controllable gui exploration,” arXiv preprint arXiv:2502.03330, 2025

    2025

  34. [42]

    Prototype2code: End-to-end front-end code generation from ui design prototypes,

    S. Xiao, Y . Chen, J. Li, L. Chen, L. Sun, and T. Zhou, “Prototype2code: End-to-end front-end code generation from ui design prototypes,” in International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, vol. 88353. American Soc...

    2024

  35. [43]

    An empirical evaluation of using large language models for automated unit test generation,

    M. Sch ¨afer, S. Nadi, A. Eghbali, and F. Tip, “An empirical evaluation of using large language models for automated unit test generation,”IEEE Transactions on Software Engineering, vol. 50, no. 1, pp. 85–105, 2024

    2024

  36. [44]

    Chatunitest: A framework for llm-based test generation,

    Y . Chen, Z. Hu, C. Zhi, J. Han, S. Deng, and J. Yin, “Chatunitest: A framework for llm-based test generation,” inCompanion Proceedings of the 32nd ACM International Conference on the Foundations of Software Engineering, 2024, pp. 572–576

    2024

  37. [45]

    Axnav: Replaying accessibility tests from natural language,

    M. Taeb, A. Swearngin, E. Schoop, R. Cheng, Y . Jiang, and J. Nichols, “Axnav: Replaying accessibility tests from natural language,” inProceedings of the CHI Conference on Human Factors in Computing Systems, ser. CHI ’24. New York, NY , USA: Association for Computing Machinery...

    2024

  38. [46]

    Semantic constraint inference for web form test generation,

    P. Alian, N. Nashid, M. Shahbandeh, and A. Mesbah, “Semantic constraint inference for web form test generation,” inProceedings of the 33rd ACM SIGSOFT International Symposium on Software Testing and Analysis, 2024, pp. 932–944

    2024

  39. [47]

    Chatting with gpt-3 for zero-shot human-like mobile automated gui testing. arxiv 2023,

    Z. Liu, C. Chen, J. Wang, M. Chen, B. Wu, X. Che, D. Wang, and Q. Wang, “Chatting with gpt-3 for zero-shot human-like mobile automated gui testing. arxiv 2023,”arXiv preprint arXiv:2305.09434, 2023

    2023 arXiv

  40. [48]

    Make llm a testing expert: Bringing human-like interaction to mobile gui testing via functionality-aware decisions,

    ——, “Make llm a testing expert: Bringing human-like interaction to mobile gui testing via functionality-aware decisions,” inProceedings of the IEEE/ACM 46th International Conference on Software Engineering, 2024, pp. 1–13

    2024

  41. [49]

    Xuat-copilot: Multi-agent collaborative system for automated user acceptance testing with large language model,

    Z. Wang, W. Wang, Z. Li, L. Wang, C. Yi, X. Xu, L. Cao, H. Su, S. Chen, and J. Zhou, “Xuat-copilot: Multi-agent collaborative system for automated user acceptance testing with large language model,”arXiv preprint arXiv:2401.02705, 2024

    2024 arXiv

  42. [50]

    Scenario-driven and context-aware automated accessibility testing for android apps,

    Y . Zhang, S. Chen, X. Xie, Z. Liu, and L. Fan, “Scenario-driven and context-aware automated accessibility testing for android apps,” in2025 IEEE/ACM 47th International Conference on Software Engineering (ICSE). IEEE Computer Society, 2025, pp. 630–630. 11

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.