Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

WebUIBench: A Comprehensive Benchmark for Evaluating Multimodal Large Language Models in WebUI-to-Code

T0 review · 3 major / 5 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 WebUIBench claims that current multimodal large language models cannot perform full front-end development as effectively as humans, and that a four-dimension evaluation of sub-capabilities predicts their WebUI-to-Code performance.

desk verdict Useful benchmark with a genuine contribution, but the central correlation validation is confounded by model scale and needs proper controls. read the letter →

arxiv 2506.07818 v1 pith:PL7UOMEJ submitted 2025-06-09 cs.CL

Zhiyu Lin , Zhengda Zhou , Zhiyuan Zhao , Tianrui Wan , Yilun Ma , Junyu Gao , Xuelong Li This is my paper
classification cs.CL
keywords WebUIBenchmultimodallargelanguagemodelsUI-to-CodeHTMLcodegenerationbenchmarkevaluationfront-endsoftwaredevelopmentvisualgroundingwebpageunderstanding
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

This paper introduces WebUIBench, a benchmark for evaluating how well multimodal large language models (MLLMs) can convert a webpage screenshot into working HTML code. The central claim is that this ability should not be measured by a single end-to-end score, but by four separate capability dimensions: perceiving visual elements, writing HTML code, understanding the correspondence between a webpage image and its code, and performing the full WebUI-to-Code task. The benchmark contains 21,793 question-answer pairs built from 719 real-world websites, and the authors evaluate 29 MLLMs on it. Their main finding is that no current model is close to human-level performance on complete front-end development, and that models show uneven strengths and weaknesses across the sub-capabilities. A sympathetic reader should care because the benchmark's structure is meant to show where the bottlenecks actually lie, so that future improvements target the right skills.

What carries the argument

The central machinery is the task taxonomy plus the automatic labeling pipeline. The taxonomy decomposes WebUI-to-Code into nine subtasks grouped under three sub-capability dimensions, and the labeling pipeline fills templated question-answer pairs using element information, CSS, and DOM structure extracted from real websites. For the WebUI-to-Code task itself, the evaluation uses a fine-grained scoring procedure: it simplifies the DOM tree to keep visually relevant elements, matches generated elements to ground-truth elements with the Hungarian algorithm using text and color similarity, and compares layout via a sketch representation of the page encoded by a visual backbone such as CLIP. This machinery is what lets the benchmark produce both sub-capability scores and an end-to-end generation score that correlates with human expert rankings.

What would settle it

Re-annotate a random sample of the 21,793 question-answer pairs with multiple independent human annotators; if a non-negligible fraction (for example, more than 10%) of answers are found to be wrong or ambiguous, and model rankings on the affected subtasks shift materially, the paper's comparative conclusions about model strengths and weaknesses would be undermined.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the four-way evaluation taxonomy—WebUI Perception, HTML Programming, WebUI-HTML Understanding, and WebUI-to-Code—produces scores that positively correlate with WebUI-to-Code performance across models. The authors interpret this correlation as validating the evaluation approach, because it shows the sub-capabilities are predictive of the end-to-end task. They also report specific capability gaps: most MLLMs perform poorly on visual grounding, especially bounding-box prediction; models below 40 billion parameters score near random chance on Webpage-HTML Matching; and small models often produce HTML that fails to compile. The paper concludes that the primary challenge is to enhance and balance sub-capabilities across different dimensions, rather than to scale a single monolithic skill.

Load-bearing premise

The automatically generated question-answer pairs are assumed to be correct and unambiguous, but their quality is checked only by sampling inspection with no measured inter-annotator agreement or reported error rate.

Editorial extensions

If this is right

  • If the positive correlation between sub-capability scores and WebUI-to-Code performance holds, then improving a model's perception, HTML programming, or cross-modal understanding should directly improve its end-to-end website generation ability.
  • The near-random performance of small models on Webpage-HTML Matching suggests that cross-modal reasoning between images and code is a distinct bottleneck that needs dedicated data and training.
  • The finding that small models generate non-compiling HTML even when they follow instructions implies that code compilation success should be a standard reporting metric in future WebUI-to-Code benchmarks.
  • The observation that models internally couple element and layout generation points toward a two-stage generation strategy, such as first producing a layout sketch and then filling in element content.
  • The preliminary experiment with the Tailwind framework indicates that using a CSS framework improves compilation success but can lower generation quality, so framework adoption must be evaluated on multiple axes.

Reading between the lines

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

  • A testable extension of this work would be to train or fine-tune a model specifically on the nine subtasks and then measure whether its WebUI-to-Code score improves; a positive result would confirm that the sub-capabilities are causally load-bearing rather than merely correlated.
  • The correlation between sub-capability scores and WebUI-to-Code performance might be partly confounded by model scale, so a direct comparison within the same model family across parameter sizes would isolate capability balance from raw capacity.
  • The benchmark's QA pairs are generated from templates and DOM data, which means the perception tasks may be easier than real-world visual ambiguity; a supplement using manually crafted or adversarially modified screenshots could probe this gap.
  • The proposed two-stage generation hypothesis could be evaluated directly by forcing models to output a layout sketch before filling in element content, and comparing the result to end-to-end generation on the same WebUI-to-Code samples.
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

3 major / 5 minor

Summary. The paper introduces WebUIBench, a benchmark for evaluating multimodal large language models (MLLMs) on WebUI-to-Code, organized along four capability dimensions (WebUI Perception, HTML Programming, WebUI-HTML Understanding, and WebUI-to-Code) with nine subtasks and 21,793 question-answer pairs derived from 719 real-world websites. The authors evaluate 29 open- and closed-source MLLMs, report performance across all subtasks, and propose a two-level automatic metric for WebUI-to-Code comprising coarse-grained visual similarity and fine-grained element- and layout-level evaluation. The central claims are that current MLLMs cannot perform the complete front-end development process as effectively as humans and that a positive correlation between sub-capability scores and WebUI-to-Code performance validates the proposed evaluation taxonomy.

Significance. If the benchmark and its validation hold, WebUIBench is a useful community resource: it decomposes WebUI-to-Code into interpretable sub-capabilities, uses real-world websites rather than synthetic data, and provides detailed annotation templates and an automatic metric design. The human-correlation study (reported average 0.83 over three experts) is genuine external evidence for the WebUI-to-Code metric, and the open-source release on GitHub supports reproducibility. However, the paper's two validation pillars—QA label quality and the sub-capability correlation argument—are currently not measured rigorously enough to support the broad conclusions about model strengths and weaknesses. The benchmark itself is likely to be valuable even after these gaps are addressed.

major comments (3)
  1. [Section 3.2] The paper states that automatic labeling is performed 'in multiple batches' with 'sampling inspection' for each task, but it reports no inter-annotator agreement, no error-rate measurement, and no detailed manual verification protocol. All model rankings in Tables 3, 4, and 6, and hence the comparative conclusions about MLLM sub-capabilities, depend on the correctness and unambiguity of these automatically generated QA pairs. The authors should provide a measured estimate of label quality, for example by having human annotators re-label a random sample of each task and reporting agreement statistics or a precise error rate, together with the number of samples inspected per batch.
  2. [Section 5.2.2 and Figure 8] The central claim that the 'positive correlation' between WebUI-to-Code performance and sub-capability performance 'validates our evaluation approach' is not established by the reported analysis. No correlation coefficients, significance tests, or confidence intervals are given for Figure 8. Moreover, the 29 evaluated models range from 2B to 78B parameters, and model scale correlates strongly with nearly every score in Table 3; the apparent correlation may thus reflect size-driven ability rather than the specific relevance of the proposed sub-capability taxonomy. The example of NVLM-D-72B is suggestive but not a systematic control. The authors should report correlation coefficients with significance tests and either partial correlations controlling for parameter count or within-size-tier analyses; otherwise the validation claim is unsupported.
  3. [Section 4.2] The human-centric evaluation is the main external validation of the WebUI-to-Code metric, but the paper does not specify which correlation coefficient is used (e.g., Spearman's rho, Kendall's tau), how many webpage pairs each expert ranked, or whether any significance or interval estimate was computed. The statement 'correlation>0.8' is also imprecise given that the average is 0.83. The authors should state the exact coefficient and its uncertainty so that the claimed agreement with human experts can be properly evaluated.
minor comments (5)
  1. [Table 3 and Table 6] There are typographical errors in the table headers and model names, such as 'Evalutaion' in Table 3 and 'Cluad' and 'Ovsi' in Table 6; these should be corrected.
  2. [Table 4] The column labeled 'Accuracy' with a downward arrow is ambiguous; it should be clearly named 'Instruction-following failure rate' or similar, and the meaning of the '♯Samples' column should be defined in the caption.
  3. [Figure 8] The text version of Figure 8 contains unreadable encoded glyphs for axis labels and legend entries; the published figure should be a clean vector graphic with clearly readable labels.
  4. [Appendix A.3, Algorithm 2] The pseudocode for the webpage segmentation algorithm contains formatting artifacts, including a stray '1;' on the input line and unclear subscript notation; the pseudocode should be reformatted for readability.
  5. [Section 3.3 and Limitations] The dataset distribution is highly imbalanced, with some subtasks such as font-style having only 0.67% of samples; the authors should report per-task sample sizes with confidence intervals or otherwise note that scores on low-count tasks will be statistically noisy.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; benchmark data, scoring, and human-validation steps are independent of the evaluated models.

full rationale

WebUIBench's derivation chain is self-contained: the QA pairs are generated from collected DOM trees, element metadata, and screenshots (Section 3.2 and Appendix A.4), independently of any MLLM output; the WebUI-to-Code metric is defined by an explicit DOM-simplification, Hungarian-matching, and visual-similarity procedure (Section 4.1); and the metric itself is anchored to an external criterion via human-expert rank correlation (Table 2). The only place that could look self-referential is Section 5.2.2, where an observed positive correlation between sub-capability scores and WebUI-to-Code scores is said to 'validate our evaluation approach.' Both variables are computed inside the same benchmark, so the correlation is internal evidence rather than an independent confirmation, and the absence of correlation coefficients or scale controls makes it weak evidence. However, this is an interpretive and validity limitation, not a circular reduction: the correlation is not a fitted parameter, no score is defined in terms of another score, and no load-bearing conclusion is obtained by renaming an input as a prediction. There are no self-citations used as premises. Accordingly, no circular step meets the evidentiary bar set by the analysis rules.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The central design rests on standard similarity metrics and on the domain assumption that source-code-derived metadata accurately reflects rendered webpages.

free parameters (3)
  • Dice similarity threshold δ
    Algorithm 1 uses a manually chosen threshold to decide when sibling DOM nodes are redundant and should be merged or removed. It affects the simplified DOM tree and therefore the set of QA pairs generated from each webpage.
  • Slice height Hs and minimum slice height Hmin
    The webpage segmentation algorithm in Algorithm 2 uses hand-chosen slice height parameters. They determine the granularity of the webpage slices and the element sets used for QA generation.
  • 4x4 grid for coarse visual grounding
    The coarse-grained visual grounding task divides screenshots into a 4x4 grid and computes grid numbers from element coordinates. This design choice defines the answer space and affects task difficulty.
assumptions (3)
  • domain assumption DOM tree and CSS values faithfully represent the visually rendered webpage
    All automatic labels (element tags, text colors, background colors, bounding boxes, OCR content) are derived from source code rather than from pixel rendering. If a page renders differently due to dynamic content, web fonts, or browser inconsistencies, the generated labels could be wrong. Invoked throughout Section 3.1 and 3.2.
  • domain assumption Human expert rankings are a valid gold standard for webpage generation quality
    The metric validation in Section 4.2 assumes that three front-end experts ranking 100 generated pages produce a trustworthy ordering, and that this ordering is what the automated metric should approximate. The paper does not measure inter-expert reliability in detail.
  • standard math Sørensen-Dice similarity, CLIP cosine similarity, and CIEDE2000 are appropriate proxies for text, layout, and color fidelity
    These are standard similarity and color-difference measures from the literature, used in Section 4.1 for scoring. They are well-established tools, but the specific thresholds and weighting are design choices.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WebUIBench: A Comprehensive Benchmark for Evaluating Multimodal Large Language Models in WebUI-to-Code." pith.science (2026). https://pith.science/paper/PL7UOMEJ

@misc{pith2026250607818,
  author       = {Pith},
  title        = {Pith review of: WebUIBench: A Comprehensive Benchmark for Evaluating Multimodal Large Language Models in WebUI-to-Code},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PL7UOMEJ}},
  note         = {Machine review of arXiv:2506.07818}
}
read the original abstract

With the rapid advancement of Generative AI technology, Multimodal Large Language Models(MLLMs) have the potential to act as AI software engineers capable of executing complex web application development. Considering that the model requires a confluence of multidimensional sub-capabilities to address the challenges of various development phases, constructing a multi-view evaluation framework is crucial for accurately guiding the enhancement of development efficiency. However, existing benchmarks usually fail to provide an assessment of sub-capabilities and focus solely on webpage generation outcomes. In this work, we draw inspiration from the principles of software engineering and further propose WebUIBench, a benchmark systematically designed to evaluate MLLMs in four key areas: WebUI Perception, HTML Programming,WebUI-HTML Understanding, and WebUI-to-Code. WebUIBench comprises 21K high-quality question-answer pairs derived from over 0.7K real-world websites. The extensive evaluation of 29 mainstream MLLMs uncovers the skill characteristics and various weakness that models encountered during the development process.

Figures

Figures reproduced from arXiv: 2506.07818 by the authors.

Figure 1
Figure 1. Evaluation taxonomy of WebUIBench. Recent works[25, 32, 4] have evaluated MLLMs and reached a consensus that MLLMs struggle to generate complex websites, revealing a significant gap between solutions and practical applications. Therefore, it is essential to identify the challenges across various development stages and evaluate the corresponding sub-capabilities of models. How￾ever, current benchmarks[11, 25] typical… view at source ↗
Figure 2
Figure 2. Task examples in the WebUI benchmark, from the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Question-Answer distribution of WebUIBench [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Schematic diagram of fine-grained WebUI-to-Code task evaluation process. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Visual grounding task examples (GPT-4o and InternVL2.5-78B). [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Results of grid number prediction (coarse [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Positive correlations between WebUI-to-Code performance and sub-capability performance. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Examples of generated webpage by Qwen2-VL-72B and GPT4o on complex webpage and slices. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Samples of Element Classification [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Samples of Attribute Perception [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Samples of OCR in the Webpage [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Samples of Visual Grounding(fine granularity). [PITH_FULL_IMAGE:figures/full_fig_p013_13.png]
Figure 14
Figure 14. Figure 14: Samples of Visual Grounding(coarse granularity). [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]
Figure 15
Figure 15. Figure 15: Samples of Code Error Correction [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]
Figure 16
Figure 16. Figure 16: Samples of Code Function Editing [PITH_FULL_IMAGE:figures/full_fig_p014_16.png]
Figure 17
Figure 17. Figure 17: Comparison of webpage generation effects by different models on complex webpages and slices. [PITH_FULL_IMAGE:figures/full_fig_p018_17.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Pattern over Pixels: Measuring Pattern Completion Bias in Multimodal Code Generation

    cs.SE 2026-08 conditional novelty 6.0 of 10

    On a new 1,440-screenshot fill-in-the-blank benchmark, five frontier multimodal models default to repeated UI patterns instead of visual deviations, with mean bias rates from 69.8% on card widths to 80.2% on font sizes.

Reference graph

Works this paper leans on

35 extracted references · 16 canonical work pages · cited by 1 Pith paper

  1. [2]

    Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. 2024. Phi-3 technical report: A highly ca- pable language model locally on your phone.arXiv preprint arXiv:2404.14219

    arXiv 2024

  2. [3]

    The claude 3 model family: Opus, sonnet, haiku

    Anthropic. The claude 3 model family: Opus, sonnet, haiku

  3. [4]

    Tony Beltramelli. 2018. pix2code: Generating code from a graphical user interface screenshot. InPro- ceedings of the ACM SIGCHI symposium on engi- neering interactive computing systems, pages 1–6

    work page 2018

  4. [5]

    Jorge Biolchini, Paula Gomes Mian, Ana Can- dida Cruz Natali, and Guilherme Horta Travassos

    work page

  5. [6]

    Liguo Chen, Qi Guo, Hongrui Jia, Zhengran Zeng, Xin Wang, Yijiang Xu, Jian Wu, Yidong Wang, Qing Gao, Jindong Wang, et al. 2024. A survey on evalu- ating large language models in code generation tasks. arXiv preprint arXiv:2408.16498

    arXiv 2024

  6. [7]

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. 2024. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling.arXiv preprint arXiv:2412.05271

    arXiv 2024

  7. [8]

    Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuolin Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping

    work page

  8. [9]

    Enrique Dehaerne, Bappaditya Dey, Sandip Halder, Stefan De Gendt, and Wannes Meert. 2022. Code generation using machine learning: A systematic review.Ieee Access, 10:82434–82455

    work page 2022

Show all 35 references
  1. [10]

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools.arXiv preprint arXiv:2406.12793

    2024 arXiv

  2. [11]

    Hongcheng Guo, Wei Zhang, Junhao Chen, Yaonan Gu, Jian Yang, Junjia Du, Binyuan Hui, Tianyu Liu, Jianxin Ma, Chang Zhou, et al. 2024. Iw-bench: Evaluating large multimodal models for converting image-to-web.arXiv preprint arXiv:2409.18980

    2024

  3. [12]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card.arXiv preprint arXiv:2410.21276

    2024 arXiv

  4. [13]

    Hugo Laurençon, Léo Tronchon, and Victor Sanh

  5. [14]

    Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexi- ang Hu, Fangyu Liu, Julian Martin Eisenschlos, Ur- vashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. 2023. Pix2struct: Screen- shot parsing as pretraining for visual language under- standing. InInternationa...

    2023

  6. [15]

    Unlocking the conversion of web screenshots into html code with the websight dataset.arXiv preprint arXiv:2403.09029

    arXiv

  7. [16]

    Lin Li, Guikun Chen, Hanrong Shi, Jun Xiao, and Long Chen. 2024. A survey on multimodal bench- marks: In the era of large ai models.arXiv preprint arXiv:2409.18142

    2024 arXiv

  8. [17]

    Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. 2023. Seed-bench: Bench- marking multimodal llms with generative comprehen- sion.arXiv preprint arXiv:2307.16125

    2023 arXiv

  9. [18]

    Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Wei- hua Luo, Kaifu Zhang, and Han-Jia Ye. 2024. Ovis: Structural embedding alignment for multimodal large language model.arXiv:2405.20797

    2024 arXiv

  10. [19]

    Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. 2025. Mm- bench: Is your multi-modal model an all-around player? InEuropean conference on computer vi- sion, pages 216–233. Springer

    2025

  11. [20]

    Tuan Anh Nguyen and Christoph Csallner. 2015. Reverse engineering mobile application user inter- faces with remaui (t). In2015 30th IEEE/ACM In- ternational Conference on Automated Software Engi- neering (ASE), pages 248–259. IEEE

    2015

  12. [21]

    M Ronnier Luo, Guihua Cui, and Bryan Rigg. 2001. The development of the cie 2000 colour-difference formula: Ciede2000.Color Research & Applica- tion: Endorsed by Inter-Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Associatio...

    2001

  13. [22]

    Generating code from a graphical user interface screenshot [electronic re- source].arXiv preprint arXiv:1705.07962

    Tony Beltramelli Pix2code. Generating code from a graphical user interface screenshot [electronic re- source].arXiv preprint arXiv:1705.07962

    arXiv

  14. [23]

    Wujian Peng, Lingchen Meng, Yitong Chen, Yi- weng Xie, Yang Liu, Tao Gui, Hang Xu, Xipeng Qiu, Zuxuan Wu, and Yu-Gang Jiang. 2024. Inst-it: Boosting multimodal instance understanding via ex- plicit visual prompt instruction tuning.arXiv preprint arXiv:2412.03565

    2024

  15. [24]

    Jiho Shin and Jaechang Nam. 2021. A survey of au- tomatic code generation from natural language.Jour- nal of Information Processing Systems, 17(3):537– 555

    2021

  16. [25]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual mod- els from natural language supervision. InInterna- tional conference on machine lea...

    2021

  17. [26]

    Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al

  18. [27]

    Chenglei Si, Yanzhe Zhang, Zhengyuan Yang, Ruibo Liu, and Diyi Yang. 2024. Design2code: How far are we from automating front-end engineering? arXiv preprint arXiv:2403.03163

    2024 arXiv

  19. [28]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elic- its reasoning in large language models.Advances in neural information processing systems, 35:24824– 24837

    2022

  20. [29]

    Gemini 1.5: Unlocking multimodal under- standing across millions of tokens of context.arXiv preprint arXiv:2403.05530

    arXiv

  21. [30]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayi- heng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-vl: Enhancing vision-language mo...

    2024 arXiv

  22. [31]

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. 2024. Mmmu: A massive multi-discipline multimodal un- derstanding and reasoning benchmark for expert agi. InProceedings of the IEEE/CVF Conferen...

    2024

  23. [32]

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. 2024. Minicpm-v: A gpt-4v level mllm on your phone.arXiv preprint arXiv:2408.01800

    2024 arXiv

  24. [33]

    Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. 2024. Yi: Open foundation models by 01. ai.arXiv preprint arXiv:2403.04652

    2024 arXiv

  25. [35]

    closing tags

    Sukmin Yun, Haokun Lin, Rusiru Thushara, Mo- hammad Qazim Bhat, Yongxin Wang, Zutao Jiang, Mingkai Deng, Jinhong Wang, Tianhua Tao, Junbo Li, et al. 2024. Web2code: A large-scale webpage- to-code dataset and evaluation framework for multi- modal llms.arXiv preprint arXiv:2406....

    2024 arXiv

  26. [2005]

    Systematic review in software engineer- ing.System engineering and computer science department COPPE/UFRJ, Technical Report ES, 679(05):45

  27. [2024]

    arXiv preprint

    Nvlm: Open frontier-class multimodal llms. arXiv preprint

Pith tools

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