Pith. sign in

REVIEW 4 major objections 5 minor 54 references

The paper claims that UI design-to-code should use vector images instead of bitmaps, and demonstrates a 90B model fine-tuned on 354,000 SVG–markup pairs that rivals or beats commercial bitmap-based models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-05 05:40 UTC pith:OWPBEDW6

load-bearing objection Solid open engineering artifacts, but the vector-superiority claim rests on a benchmark the paper itself calls non-comparable. the 4 major comments →

arxiv 2509.05394 v1 pith:OWPBEDW6 submitted 2025-09-05 cs.SE cs.AI

Reverse Browser: Vector-Image-to-Code Generator

Zoltan Toth-Czifra This is my paper
classification cs.SE cs.AI
keywords vector image to codeSVG to HTML/CSSimage-to-UI generationUI code generationmulti-scale pixel similarityperceptual image quality assessmentLoRA fine-tuninglarge language models for code
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Reverse Browser sets out to overturn a default choice in image-to-code systems: the input image. The paper argues that a bitmap screenshot hides web-page structure inside a W×H×C pixel grid, while a vector image (SVG) spells out the same design as explicit primitives — boxes, coordinates, colors, text styles — much like the target HTML/CSS does. To make the case, the author builds large synthetic and public-web datasets of SVG–markup pairs, introduces a stricter multi-scale image-similarity metric, and fine-tunes an open 90B Llama 3.2 model with LoRA on a single GPU. On an adapted 900-page benchmark, the resulting model scores comparable EMS and clearly better normalized LPIPS than the best commercial multimodal models that consume bitmaps. The paper's conclusion is that vector images are a more promising input modality for high-fidelity UI-to-code generation.

Core claim

The central claim is that feeding a UI design to a code generator as a vector image (SVG) rather than a bitmap screenshot is the right move. Because an SVG is produced by laying out DOM elements, it already contains the structure the model must output: nested rectangles, text, colors, and positions. The paper builds about 354,000 SVG–HTML/CSS training pairs from public web pages, filters them with its new MSPS metric plus LPIPS, and fine-tunes Llama 3.2 90B (quantized, LoRA) in about 392 GPU-hours on one H100. Final test-set scores are 0.9530 MSPS and 0.3012 LPIPS. On an adapted version of the Image2Struct benchmark, the model reaches EMS 0.693 and normalized LPIPS 0.755, compared with best

What carries the argument

The load-bearing object is the SVG representation of a laid-out web page: each page is crawled and rendered headlessly, then converted into an SVG whose shapes match the DOM elements' positions, sizes, colors, and text, so the generation task becomes inverting layout arithmetic rather than perceiving layout from pixels. The second piece is Multi-Scale Pixel Similarity (MSPS), a new full-reference metric that averages pixel MSE at successive downscaling levels; it is stricter than perceptual metrics like LPIPS and is used to filter web-to-SVG conversions, monitor training, and rank samples. The third is the fine-tuned 90B decoder-only Llama 3.2 model trained with quantized LoRA on public-web

Load-bearing premise

The load-bearing premise is that Table 2's comparison — bitmaps on the original task versus SVGs on a modified protocol with JavaScript disabled, out-of-viewport DOM removed, and 18 of 900 pages discarded — still measures a genuine advantage for vector inputs.

What would settle it

Run a matched head-to-head on the same 900 web pages: same model family, same prompts, same sampling, same evaluation pipeline; one arm receives the bitmap screenshot, the other the SVG render, both from identical viewports with JavaScript enabled and without pruning the DOM. If the bitmap arm does not score lower on LPIPS/EMS (or at least equal), the central recommendation that vector images are better collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • UI design tools that already export SVG become drop-in front ends for code generation, and image-to-code benchmarks should add a vector-to-code track with normalized metrics.
  • High-fidelity image-to-code becomes reproducible on modest hardware: the whole training run fits on a single H100 under 400 hours, so the result is no longer gated to large industrial clusters.
  • Evaluation should pair a strict pixel-level metric like MSPS with a perceptual metric like LPIPS, because perceptual scoring alone can miss visual differences that matter for production UI.
  • Public-web crawls can supply the large, diverse SVG–markup pairs needed to overcome the low-complexity limitation of synthetic training sets used by earlier image-to-code work.
  • With vector inputs, the remaining hard problems shift from perception to generation: exact layout arithmetic, responsive breakpoints, and functional/interactive behavior.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the SVG advantage holds under a controlled head-to-head, a natural next step is joint training on paired bitmap+vector inputs — vector supplies structure, bitmap preserves textures and rendered effects that SVG conversion discards; the paper lists this idea only as future work.
  • Editorial inference: MSPS is a generic fidelity check for any structured-image generation task (diagrams, CAD drawings, game UI mockups), not just webpages, and could be reused for data filtering or as a reward signal.
  • Editorial inference: the comparison hints that part of frontier VLMs' poor Image2Struct performance is caused by the bitmap input modality itself, which could be tested by giving the same commercial models SVG-based inputs.
  • Editorial inference: reward hacking via 'SVG-in, SVG-out' is a direct risk of the vector-input design, which actually makes image-to-code a good candidate for verifiable-reward reinforcement learning where the image-similarity metric doubles as the reward.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces Reverse Browser, a vector-image-to-code system that converts SVG renderings of web pages into HTML/CSS markup. The author contributes a crawler that builds SVG-markup pairs from web layouts, several synthetic datasets and a large public web dataset, a new image quality metric (Multi-Scale Pixel Similarity, MSPS), and a fine-tuned Llama 3.2 90B model with LoRA. The model is evaluated on synthetic data and on a modified version of the Image2Struct benchmark, where its scores are contrasted with bitmap-input foundation models. The paper's headline conclusion is that vector images may be a better choice for image-to-UI than bitmaps.

Significance. If the comparative claim were valid, this would be a practically important result: it would suggest that the standard bitmap-input paradigm for image-to-code should be revisited. The engineering contributions are concrete and well-documented: the crawler, datasets, metric implementation, and model weights are all released under permissive licenses, which is valuable for a relatively unexplored research direction. However, the central comparison in Table 2 is confounded by task and protocol differences that the author himself concedes in Section 6.1. The paper is transparent about many limitations, but the Section 7 claim of vector superiority goes beyond what the evidence supports. The most defensible contribution is the open-source infrastructure and the demonstration that a single trained model can produce plausible code from SVG inputs, not the comparative superiority claim.

major comments (4)
  1. [Section 7 / Section 6.1 / Appendix A.2] The central claim in Section 7 ('vector images may be a better choice for image-to-UI than bitmaps') rests on Table 2. As Section 6.1 itself states, 'the scores are not comparable because they measure the performance on different tasks.' Appendix A.2 makes the differences concrete: the Reverse Browser pipeline disabled JavaScript, removed DOM elements outside the viewport before building the SVG, supplied image assets to the generated pages, and evaluated on 882 pages after discarding 18 pages that prevented SVG conversion or rendering. The bitmap FM scores are taken from the unmodified Image2Struct leaderboard, where models must infer the full rendered 1920x1080 viewport including JS-dependent and viewport-relative content. The RB LPIPS* value (0.755 vs. 0.383 best FM) and EMS score (0.693 vs. 0.735) are therefore confounded by target difficulty and evaluation protocol. This gap does no
  2. [Section 5.3 / Section 7] The phrase 'relatively small model trained by an individual with a limited budget' is not supported by the reported setup. The final model is a 90B-parameter Llama 3.2 fine-tuned with LoRA, trained for 392 hours on a single H100 GPU (3.64e20 FLOPs). That is a substantial, though not industrial, compute budget. More importantly, no matched bitmap-input model was trained on the same pages with the same data and budget, so Table 2 cannot isolate input modality as the causal variable. A controlled comparison—for example, training or adapting a bitmap-input model on the same 882-page benchmark with the same evaluation protocol, or at least running a bitmap FM on the simplified target pages—is needed to support the Section 7 claim.
  3. [Sections 3.2, 4.2, 5.3] MSPS is introduced in this paper and used as the quality gate for accepting or rejecting HTML-to-SVG conversions in the training dataset, as well as an evaluation metric. Because MSPS is a new, unvalidated metric, the training distribution may be silently biased toward pages that score well under this particular metric. The paper reports a moderate correlation with LPIPS (Pearson r ≈ -0.62) but not a human-study validation of MSPS or an analysis of which pages are discarded by the filter. At minimum, report the MSPS distribution of rejected and accepted pages and check whether the final model's test-set scores are sensitive to the filter threshold.
  4. [Section 6.1 / Table 2] Table 2 combines two different success-rate definitions: 'Compilation success' as defined by Image2Struct and 'Training and inference success' influenced by different failure modes. The Reverse Browser row reports 0.981 for the latter, while the FM rows report the former. This makes even the successful-compilation comparison ambiguous. Please report the same success metric for all models, or clearly define the difference in the table and in the text.
minor comments (5)
  1. [Section 4.2, Eq. (3)] The variables H_i and W_i are used in the equation but not defined; the definition of N in Eq. (4) uses H1 and W1, while H and W in Eq. (3) are ambiguous. Please clarify with consistent notation.
  2. [Section 6.1, Table 2] The table header is malformed ('EMS (vector)LPIPS*') and the meaning of the '(vector)' columns is unclear until the reader reaches the text. Add proper column separation and explain the notation in the caption.
  3. [Section 5.2] Training details such as batch size, learning rate, optimizer, and LoRA rank are not given in the text and are only available in the external repository. Add the key hyperparameters in the paper for reproducibility.
  4. [Section 2.1] The sentence 'According to the author, the results are of low quality' is informal and vague. Replace it with a citation to a concrete evaluation claim.
  5. [Throughout] There are typographical and consistency issues: 'LLaV A' should be 'LLaVA', 'Image2struct' capitalization is inconsistent, the title has 'V ector' spacing, and the reference list includes nonstandard retrieval-date parentheticals. A copy-edit pass is needed.

Circularity Check

2 steps flagged

Minor circularity from internal MSPS filter/evaluation and a non-comparable Image2Struct comparison; the core model training is otherwise empirical and not circular.

specific steps
  1. other [Section 3.2 + Section 4.2 + Section 5.3]
    "To filter imperfect HTML-to-SVG conversion results, a new, stricter metric was developed... It is called Multi-Scale Pixel Similarity (MSPS). ... The items were then filtered by their similarity score, comparing the bitmap of the webpage with the bitmap rendering of the resulting vector image... The final test set scores are: 0.9530 MSPS and 0.3012 LPIPS."

    The same metric introduced in this paper (MSPS) is used both as the acceptance filter when building the training/evaluation datasets and as the reported accuracy metric for the final model. The reported MSPS score therefore measures performance on data that was already selected to score highly on MSPS, so the metric is internal to the data-construction loop rather than an independent external benchmark. This is a minor circularity because the central vector-superiority claim rests on LPIPS/EMS against Image2Struct, not on MSPS.

  2. other [Section 6.1 + Section 7 + Appendix A.2]
    "the scores are not comparable because they measure the performance on different tasks ... Reverse Browser's renderer explicitly disables JavaScript in the crawled pages, and the model is not trained to generate JavaScript ... When crawling, I removed all DOM elements outside the viewport and captured SVG with only the remaining elements."

    Section 7 interprets the Table 2 scores as evidence that vector images may be better than bitmaps, but the vector pipeline's inputs are produced by the author's own crawler under simplified conditions: JavaScript disabled, out-of-viewport DOM removed, and 18 of 900 pages discarded. The bitmap FMs are scored on the original Image2Struct task, which includes JS-dependent and viewport-relative content. The score gap is therefore partly a consequence of the easier task definition built into the SVG construction, not an isolated effect of input modality. The paper honestly discloses this non-comparability, but the conclusion still relies on the comparison.

full rationale

There is no fitted-parameter circularity and no self-citation chain: the central result is an empirical model trained on large datasets, and the model's ability to map SVG to markup is demonstrated on synthetic data and on a held-out test set. The two flagged issues are minor and internal rather than fatal. First, MSPS is introduced by this paper and used both to filter the dataset and to report final accuracy, so MSPS numbers are not an independent external benchmark. Second, the Image2Struct comparison is explicitly acknowledged by the paper as non-comparable because the vector task uses the author's crawler with JavaScript disabled and viewport-only DOM, whereas the bitmap models face the full rendered page. These issues confound the vector-superiority interpretation but do not reduce the core derivation to its own inputs. The paper's central contribution remains an open-weights model and datasets, not a mathematical derivation, so the appropriate circularity score is low.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on engineering choices and domain assumptions, not on a mathematical derivation. The only newly named item, MSPS, is a metric rather than a postulated entity. The listed free parameters are data-processing and inference choices that affect results but are disclosed in the paper.

free parameters (3)
  • Training sequence length cutoff = 12,000 tokens
    Set by single-GPU memory limits and used to filter the 'Large' public web dataset before training the 90B model (Section 5.3).
  • SVG conversion acceptance threshold = not stated precisely; uses MSPS filter
    The crawler keeps only HTML-to-SVG conversions whose bitmap rendering is similar enough according to MSPS (Section 3.2); the exact threshold is not reported.
  • Best-of-N inference parameters for the qualitative demo = N=50, temperature=0.5, top-p=0.9
    Used for the use-case demonstration in Figure 6 and discussed in Section 6.2; the benchmark scores use temperature 0.0 without best-of-N.
axioms (4)
  • domain assumption The browser-layout-to-SVG conversion preserves all visual information needed to reconstruct the page.
    Section 3.2 constructs SVG from element attributes (position, size, color, text style). Text, embedded images, and visual effects are not proven to be lossless.
  • domain assumption A transformer fine-tuned on SVG-markup pairs can learn the inverse of the layout arithmetic.
    Section 5.2 frames the task as inverse layout computation and relies on proof-of-concept experiments on synthetic data to justify scaling up.
  • domain assumption MSPS and LPIPS scores reflect design fidelity for both filtering and final evaluation.
    Section 4 uses these metrics as ground truth for conversion filtering, training monitoring, and final evaluation, without a human study validating the thresholds.
  • domain assumption Web pages sampled from Common Crawl represent real-world UI complexity.
    Section 3.2 uses Common Crawl URLs for the large dataset; Section 6.2 admits the resulting model is not commercially viable, so the complexity coverage remains incomplete.

pith-pipeline@v1.4.0-alltime-deepseek-medium · 14379 in / 11528 out tokens · 120880 ms · 2026-08-05T05:40:37.022237+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Reverse Browser: Vector-Image-to-Code Generator." pith.science (2026). https://pith.science/paper/OWPBEDW6

@misc{pith2026250905394,
  author       = {Pith},
  title        = {Pith review of: Reverse Browser: Vector-Image-to-Code Generator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OWPBEDW6}},
  note         = {Machine review of arXiv:2509.05394}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Automating the conversion of user interface design into code (image-to-code or image-to-UI) is an active area of software engineering research. However, the state-of-the-art solutions do not achieve high fidelity to the original design, as evidenced by benchmarks. In this work, I approach the problem differently: I use vector images instead of bitmaps as model input. I create several large datasets for training machine learning models. I evaluate the available array of Image Quality Assessment (IQA) algorithms and introduce a new, multi-scale metric. I then train a large open-weights model and discuss its limitations.

Figures

Figures reproduced from arXiv: 2509.05394 by Zoltan Toth-Czifra.

Figure 1
Figure 1. Figure 1: Conceptual figure of rendering a web page into pairs of vector images and markup. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of Multi-Scale Pixel Similarity (MSPS). [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: LongT5 pre-training and fine-tuning on the "Color+Text+Size" dataset. The upper figure [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Llama 3.2 90B fine-tuning on the "Large" dataset. The left-side chart shows the evolution [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: A few randomly selected examples from the model’s output on the Image2Struct web pages [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The best model output by LPIPS out of 50 samples for each design step, with accuracy [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: The best model output by LPIPS out of 25 samples for each design step without filtering [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Examples of web pages included in each simple synthetic dataset. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 17 canonical work pages · 4 internal anchors

  1. [1]

    Image2Struct: Benchmarking Structure Extraction for Vision- Language Models

    Josselin Somerville Roberts et al. Image2Struct: Benchmarking Structure Extraction for Vision- Language Models. arXiv:2410.22456 [cs]. Oct. 2024. DOI: 10.48550/arXiv.2410.22456. URL: http://arxiv.org/abs/2410.22456 (retrieved 5/4/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.22456 2024

  2. [2]

    WebCode2M: A Real-World Dataset for Code Generation from Webpage Designs

    Yi Gui et al. “WebCode2M: A Real-World Dataset for Code Generation from Webpage Designs”. In: Proceedings of the ACM on Web Conference 2025. arXiv:2404.06369 [cs]. Apr. 2025, pp. 1834–1845. DOI: 10.1145/3696410.3714889. URL: http://arxiv.org/abs/ 2404.06369 (retrieved 7/5/2025)

    Pith/arXiv arXiv 2025

  3. [3]

    2024 Stack Overflow Developer Survey

    Stack Overflow. 2024 Stack Overflow Developer Survey. Publisher: Stack Overflow, Inc. 2024. URL: https://survey.stackoverflow.co/2024/ (retrieved 3/5/2025)

    work page 2024

  4. [4]

    Attention Is All You Need

    Ashish Vaswani et al. Attention Is All You Need . arXiv:1706.03762 [cs]. June 2017. DOI: 10.48550/arXiv.1706.03762 . URL: http://arxiv.org/abs/1706.03762 (retrieved 28/2/2025). 10https://huggingface.co/datasets/stanford-crfm/image2struct-webpage-v1 11https://jekyllrb.com/ 12https://github.com/tcz/rb-experiments/blob/main/helm-generations/vllm-image2struct...

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.03762 2017

  5. [5]

    Weisz et al

    Justin D. Weisz et al. Examining the Use and Impact of an AI Code Assistant on Developer Productivity and Experience in the Enterprise. arXiv:2412.06603 [cs]. Mar. 2025. DOI: 10. 48550 / arXiv . 2412 . 06603. URL: http : / / arxiv . org / abs / 2412 . 06603(retrieved 5/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.06603 2025

  6. [6]

    Web Hypertext Application Technology Working Group.HTML Living Standard. Aug. 2025. URL: https://html.spec.whatwg.org/

    work page 2025

  7. [7]

    Cascading Style Sheets Level 2 Revision 2 (CSS 2.2) Specification

    Bert Bos. Cascading Style Sheets Level 2 Revision 2 (CSS 2.2) Specification. Apr. 2016. URL: https://www.w3.org/TR/CSS22/

    work page 2016

  8. [8]

    pix2code: Generating Code from a Graphical User Interface Screenshot

    Tony Beltramelli. pix2code: Generating Code from a Graphical User Interface Screenshot. arXiv:1705.07962 [cs]. Sept. 2017. DOI: 10 . 48550 / arXiv . 1705 . 07962. URL: http : //arxiv.org/abs/1705.07962 (retrieved 28/2/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.1705.07962 2017

  9. [9]

    Frontend Diffusion: Empowering Self-Representation of Junior Researchers and Designers Through Agentic Workflows

    Zijian Ding et al. Frontend Diffusion: Empowering Self-Representation of Junior Researchers and Designers Through Agentic Workflows . arXiv:2502.03788 [cs]. Feb. 2025. DOI: 10 . 48550 / arXiv . 2502 . 03788. URL: http : / / arxiv . org / abs / 2502 . 03788(retrieved 5/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.03788 2025

  10. [10]

    Sketch2code: Generating a website from a paper mockup

    Alex Robinson. Sketch2code: Generating a website from a paper mockup. arXiv:1905.13750 [cs]. May 2019. DOI: 10.48550/arXiv.1905.13750 . URL: http://arxiv.org/abs/ 1905.13750 (retrieved 28/2/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.1905.13750 1905

  11. [11]

    Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Under- standing

    Kenton Lee et al. Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Under- standing. arXiv:2210.03347 [cs]. June 2023. DOI: 10.48550/arXiv.2210.03347 . URL: http://arxiv.org/abs/2210.03347 (retrieved 28/2/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2210.03347 2023

  12. [12]

    OCR-free Document Understanding Transformer

    Geewook Kim et al. OCR-free Document Understanding Transformer. arXiv:2111.15664 [cs]. Oct. 2022. DOI: 10.48550/arXiv.2111.15664 . URL: http://arxiv.org/abs/2111. 15664 (retrieved 28/2/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2111.15664 2022

  13. [13]

    Learning UI-to-Code Reverse Generator Using Visual Critic Without Rendering

    Davit Soselia, Khalid Saifullah, and Tianyi Zhou. Learning UI-to-Code Reverse Generator Using Visual Critic Without Rendering. arXiv:2305.14637 [cs]. Nov. 2023. DOI: 10.48550/ arXiv.2305.14637. URL: http://arxiv.org/abs/2305.14637 (retrieved 28/2/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2305.14637 2023

  14. [14]

    Language Models are Unsupervised Multitask Learners

    Alec Radford et al. Language Models are Unsupervised Multitask Learners. Feb. 2019. URL: https : / / cdn . openai . com / better - language - models / language _ models _ are _ unsupervised_multitask_learners.pdf

    work page 2019

  15. [15]

    LLaMA: Open and Efficient Foundation Language Models

    Hugo Touvron et al. LLaMA: Open and Efficient Foundation Language Models . arXiv:2302.13971 [cs]. Feb. 2023. DOI: 10 . 48550 / arXiv . 2302 . 13971. URL: http : //arxiv.org/abs/2302.13971 (retrieved 5/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2302.13971 2023

  16. [16]

    UICopilot: Automating UI Synthesis via Hierarchical Code Generation from Webpage Designs

    Yi Gui et al. “UICopilot: Automating UI Synthesis via Hierarchical Code Generation from Webpage Designs”. en. In: Proceedings of the ACM on Web Conference 2025. Sydney NSW Australia: ACM, Apr. 2025, pp. 1846–1855.DOI: 10.1145/3696410.3714891. URL: https: //dl.acm.org/doi/10.1145/3696410.3714891 (retrieved 5/5/2025)

    work page arXiv 2025

  17. [17]

    GPT-4 Technical Report

    OpenAI et al. GPT-4 Technical Report. arXiv:2303.08774 [cs]. Mar. 2024. DOI: 10.48550/ arXiv.2303.08774. URL: http://arxiv.org/abs/2303.08774 (retrieved 5/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.08774 2024

  18. [18]

    Scalable Vector Graphics (SVG) 2

    Amelia Bellamy-Royds et al. Scalable Vector Graphics (SVG) 2. URL: https://www.w3. org/TR/SVG2/ (retrieved 3/3/2025)

    work page 2025

  19. [19]

    Applying Probability Measures to Abstract Languages

    T.L. Booth and R.A. Thompson. “Applying Probability Measures to Abstract Languages”. In: IEEE Transactions on Computers C-22.5 (May 1973), pp. 442–450. DOI: 10.1109/T- C.1973.223746. URL: http://ieeexplore.ieee.org/document/1672339/ (retrieved 11/3/2025)

    arXiv 1973

  20. [20]

    What Makes My Model Perplexed? A Linguistic Investigation on Neural Language Models Perplexity

    Alessio Miaschi et al. “What Makes My Model Perplexed? A Linguistic Investigation on Neural Language Models Perplexity”. en. In: Proceedings of Deep Learning Inside Out (DeeLIO): The 2nd Workshop on Knowledge Extraction and Integration for Deep Learning Architectures. Online: Association for Computational Linguistics, 2021, pp. 40–47. DOI: 10.18653/v1/ 20...

    doi:10.18653/v1/ 2021

  21. [21]

    Common Crawl Dataset

    Common Crawl. Common Crawl Dataset. 2024. URL: https://commoncrawl.org

    work page 2024

  22. [22]

    Yi Tay et al.Long Range Arena: A Benchmark for Efficient Transformers. en. arXiv:2011.04006 [cs]. Nov. 2020. DOI: 10.48550/arXiv.2011.04006 . URL: http://arxiv.org/abs/ 2011.04006 (retrieved 4/3/2025). 16

    Pith/arXiv arXiv doi:10.48550/arxiv.2011.04006 2011

  23. [23]

    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

    Charlie Snell et al. Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. arXiv:2408.03314 [cs]. Aug. 2024. DOI: 10.48550/arXiv.2408. 03314. URL: http://arxiv.org/abs/2408.03314 (retrieved 3/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.03314 2024

  24. [24]

    BLEU: a method for automatic evaluation of machine translation

    Kishore Papineni et al. “BLEU: a method for automatic evaluation of machine translation”. en. In: Proceedings of the 40th Annual Meeting on Association for Computational Linguistics - ACL ’02. Philadelphia, Pennsylvania: Association for Computational Linguistics, 2001, p. 311. DOI: 10.3115/1073083.1073135 . URL: http://portal.acm.org/citation.cfm? doid=10...

    arXiv 2001

  25. [25]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang et al. “Image quality assessment: from error visibility to structural similarity”. In: IEEE Transactions on Image Processing13.4 (Apr. 2004), pp. 600–612. DOI: 10.1109/TIP. 2003.819861. URL: https://ieeexplore.ieee.org/document/1284395/ (retrieved 5/4/2025)

    arXiv 2004

  26. [26]

    Learning Transferable Visual Models From Natural Language Supervision

    Alec Radford et al. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020 [cs]. Feb. 2021. DOI: 10 . 48550 / arXiv . 2103 . 00020. URL: http : //arxiv.org/abs/2103.00020 (retrieved 8/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2103.00020 2021

  27. [27]

    A metric for distributions with applications to image databases

    Y . Rubner, C. Tomasi, and L.J. Guibas. “A metric for distributions with applications to image databases”. In: Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271). Bombay, India: Narosa Publishing House, 1998, pp. 59–66. DOI: 10.1109/ICCV.1998. 710701. URL: http://ieeexplore.ieee.org/document/710701/ (retrieved 5/4/2025)

    doi:10.1109/iccv.1998 1998

  28. [28]

    TOPIQ: A Top-down Approach from Semantics to Distortions for Image Quality Assessment

    Chaofeng Chen et al. TOPIQ: A Top-down Approach from Semantics to Distortions for Image Quality Assessment. arXiv:2308.03060 [cs]. Aug. 2023.DOI: 10.48550/arXiv.2308.03060. URL: http://arxiv.org/abs/2308.03060 (retrieved 2/4/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2308.03060 2023

  29. [29]

    Attentions Help CNNs See Better: Attention-based Hybrid Image Quality Assessment Network

    Shanshan Lao et al. Attentions Help CNNs See Better: Attention-based Hybrid Image Quality Assessment Network. arXiv:2204.10485 [cs]. Apr. 2022.DOI: 10.48550/arXiv.2204.10485. URL: http://arxiv.org/abs/2204.10485 (retrieved 2/4/2025)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2204.10485 2022

  30. [30]

    PieAPP: Perceptual Image-Error Assessment through Pairwise Preference

    Ekta Prashnani et al.PieAPP: Perceptual Image-Error Assessment through Pairwise Preference. arXiv:1806.02067 [cs]. June 2018. DOI: 10 . 48550 / arXiv . 1806 . 02067. URL: http : //arxiv.org/abs/1806.02067 (retrieved 2/4/2025)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1806.02067 2018

  31. [31]

    The Unreasonable Effectiveness of Deep Features as a Perceptual Metric

    Richard Zhang et al. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. arXiv:1801.03924 [cs]. Apr. 2018. DOI: 10 . 48550 / arXiv . 1801 . 03924. URL: http://arxiv.org/abs/1801.03924 (retrieved 2/4/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.1801.03924 2018

  32. [32]

    Image Quality Assessment: Unifying Structure and Texture Similarity

    Keyan Ding et al. “Image Quality Assessment: Unifying Structure and Texture Similarity”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence(2020). arXiv:2004.07728 [cs], pp. 1–1. DOI: 10.1109/TPAMI.2020.3045810. URL: http://arxiv.org/abs/2004. 07728 (retrieved 2/4/2025)

    Pith/arXiv arXiv 2020

  33. [33]

    Deep Neural Networks for No-Reference and Full-Reference Image Quality Assessment

    Sebastian Bosse et al. “Deep Neural Networks for No-Reference and Full-Reference Image Quality Assessment”. In: IEEE Transactions on Image Processing 27.1 (Jan. 2018). Con- ference Name: IEEE Transactions on Image Processing, pp. 206–219. DOI: 10.1109/TIP. 2017.2760518. URL: https://ieeexplore.ieee.org/document/8063957 (retrieved 2/4/2025)

    arXiv 2018

  34. [34]

    Learning Conditional Knowledge Distillation for Degraded-Reference Image Quality Assessment

    Heliang Zheng et al. Learning Conditional Knowledge Distillation for Degraded-Reference Image Quality Assessment. arXiv:2108.07948 [eess]. Aug. 2021. DOI: 10.48550/arXiv. 2108.07948. URL: http://arxiv.org/abs/2108.07948 (retrieved 2/4/2025)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2108.07948 2021

  35. [35]

    FSIM: A Feature Similarity Index for Image Quality Assessment

    Lin Zhang et al. “FSIM: A Feature Similarity Index for Image Quality Assessment”. In: IEEE Transactions on Image Processing 20.8 (Aug. 2011), pp. 2378–2386. DOI: 10.1109/TIP. 2011.2109730. URL: http://ieeexplore.ieee.org/document/5705575/ (retrieved 2/4/2025)

    arXiv 2011

  36. [36]

    Multi-Scale Structural Similarity for Image Quality Assessment

    Zhou Wang, Eero P. Simoncelli, and Alan C. Bovik. “Multi-Scale Structural Similarity for Image Quality Assessment”. In: Proceedings of the 37th IEEE Asilomar Conference on Signals, Systems and Computers. Pacific Grove, CA: IEEE, Nov. 2003, pp. 1398–1402. URL: https://ieeexplore.ieee.org/document/1292216

    arXiv 2003

  37. [37]

    Complex Wavelet Structural Similarity: A New Image Similarity Index

    M.P. Sampat et al. “Complex Wavelet Structural Similarity: A New Image Similarity Index”. In: IEEE Transactions on Image Processing 18.11 (Nov. 2009), pp. 2385–2401. DOI: 10. 1109/TIP.2009.2025923. URL: http://ieeexplore.ieee.org/document/5109651/ (retrieved 2/4/2025). 17

    work page arXiv 2009

  38. [38]

    Image information and visual quality

    H.R. Sheikh and A.C. Bovik. “Image information and visual quality”. In: IEEE Transactions on Image Processing15.2 (Feb. 2006), pp. 430–444. DOI: 10.1109/TIP.2005.859378. URL: https://ieeexplore.ieee.org/document/1576816 (retrieved 2/4/2025)

    work page arXiv 2006

  39. [39]

    Gradient Magnitude Similarity Deviation: A Highly Efficient Perceptual Image Quality Index

    Wufeng Xue et al. “Gradient Magnitude Similarity Deviation: A Highly Efficient Percep- tual Image Quality Index”. In: IEEE Transactions on Image Processing 23.2 (Feb. 2014). arXiv:1308.3052 [cs], pp. 684–695. DOI: 10 . 1109 / TIP . 2013 . 2293423. URL: http : //arxiv.org/abs/1308.3052 (retrieved 2/4/2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  40. [40]

    Perceptual image quality assessment using a normalized Laplacian pyramid

    Valero Laparra et al. “Perceptual image quality assessment using a normalized Laplacian pyramid”. In: Electronic Imaging 28.16 (Feb. 2016), pp. 1–6. DOI: 10.2352/ISSN.2470- 1173.2016.16.HVEI-103 . URL: https://library.imaging.org/ei/articles/28/ 16/art00008 (retrieved 2/4/2025)

    arXiv 2016

  41. [41]

    VSI: A Visual Saliency-Induced Index for Perceptual Image Quality Assessment

    Lin Zhang, Ying Shen, and Hongyu Li. “VSI: A Visual Saliency-Induced Index for Perceptual Image Quality Assessment”. In: IEEE Transactions on Image Processing23.10 (Oct. 2014), pp. 4270–4281. DOI: 10.1109/TIP.2014.2346028 . URL: https://ieeexplore.ieee. org/document/6873260/ (retrieved 2/4/2025)

    work page arXiv 2014

  42. [42]

    Allan G. Weber. The USC-SIPI Image Database: Version 6. Tech. rep. 432. Los Angeles, CA: Signal and Image Processing Institute, University of Southern California, Feb. 2018. URL: https://sipi.usc.edu/database/SIPI_Database.pdf

    work page 2018

  43. [43]

    Transformers Can Do Arithmetic with the Right Embeddings

    Sean McLeish et al. Transformers Can Do Arithmetic with the Right Embeddings . arXiv:2405.17399 [cs]. Dec. 2024. DOI: 10 . 48550 / arXiv . 2405 . 17399. URL: http : //arxiv.org/abs/2405.17399 (retrieved 5/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.17399 2024

  44. [44]

    Teaching Arithmetic to Small Transformers

    Nayoung Lee et al. Teaching Arithmetic to Small Transformers. arXiv:2307.03381 [cs]. July

    Pith/arXiv arXiv

  45. [45]

    Pre-trained Large Language Models Use Fourier Features to Compute Addition

    Tianyi Zhou et al. Pre-trained Large Language Models Use Fourier Features to Compute Addition. arXiv:2406.03445 [cs]. June 2024. DOI: 10.48550/arXiv.2406.03445 . URL: http://arxiv.org/abs/2406.03445 (retrieved 5/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2406.03445 2024

  46. [46]

    Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer

    Colin Raffel et al. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. arXiv:1910.10683 [cs]. Sept. 2023. DOI: 10.48550/arXiv.1910.10683. URL: http://arxiv.org/abs/1910.10683 (retrieved 14/4/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.1910.10683 1910

  47. [47]

    LongT5: Efficient Text-To-Text Transformer for Long Sequences

    Mandy Guo et al. LongT5: Efficient Text-To-Text Transformer for Long Sequences . arXiv:2112.07916 [cs]. May 2022. DOI: 10 . 48550 / arXiv . 2112 . 07916. URL: http : //arxiv.org/abs/2112.07916 (retrieved 16/4/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2112.07916 2022

  48. [48]

    The Llama 3 Herd of Models

    Aaron Grattafiori et al. The Llama 3 Herd of Models. arXiv:2407.21783 [cs]. Nov. 2024. DOI: 10.48550/arXiv.2407.21783 . URL: http://arxiv.org/abs/2407.21783 (retrieved 16/4/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2407.21783 2024

  49. [49]

    Hu et al.LoRA: Low-Rank Adaptation of Large Language Models

    Edward J. Hu et al.LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685 [cs]. Oct. 2021. DOI: 10.48550/arXiv.2106.09685 . URL: http://arxiv.org/abs/ 2106.09685 (retrieved 30/4/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.09685 2021

  50. [50]

    arXiv:2503.23829 [cs]

    Yi Su et al.Crossing the Reward Bridge: Expanding RL with Verifiable Rewards Across Diverse Domains. arXiv:2503.23829 [cs]. Apr. 2025. DOI: 10.48550/arXiv.2503.23829 . URL: http://arxiv.org/abs/2503.23829 (retrieved 8/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2503.23829 2025

  51. [51]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao et al. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300 [cs]. Apr. 2024. DOI: 10.48550/arXiv.2402.03300. URL: http://arxiv.org/abs/2402.03300 (retrieved 2/9/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.03300 2024

  52. [52]

    GTPO and GRPO-S: Token and Sequence-Level Reward Shaping with Policy Entropy

    Hongze Tan and Jianfei Pan. GTPO and GRPO-S: Token and Sequence-Level Reward Shaping with Policy Entropy. arXiv:2508.04349 [cs]. Aug. 2025.DOI: 10.48550/arXiv.2508.04349. URL: http://arxiv.org/abs/2508.04349 (retrieved 2/9/2025)

    work page doi:10.48550/arxiv.2508.04349 2025

  53. [53]

    CAPO: Towards Enhancing LLM Reasoning through Verifiable Generative Credit Assignment

    Guofu Xie et al. CAPO: Towards Enhancing LLM Reasoning through Verifiable Generative Credit Assignment. arXiv:2508.02298 [cs]. Aug. 2025. DOI: 10.48550/arXiv.2508.02298. URL: http://arxiv.org/abs/2508.02298 (retrieved 2/9/2025). 18

    work page doi:10.48550/arxiv.2508.02298 2025

  54. [2023]

    URL: http://arxiv.org/abs/2307.03381 (retrieved 5/5/2025)

    DOI: 10.48550/arXiv.2307.03381. URL: http://arxiv.org/abs/2307.03381 (retrieved 5/5/2025)

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.03381 2025