Pith. sign in

REVIEW 5 major objections 5 minor 2 cited by

Human-Agent Collaborative Paper-to-Page Crafting

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read AutoPage claims a multi-agent, coarse-to-fine pipeline can turn a paper PDF into a polished interactive project page in under 15 minutes for less than $0.1, beating one-shot LLM generation.

desk verdict A new task and a useful first benchmark, but the quantitative case rests on swapped VLM rubrics and inconsistent numbers — send it out, but only with major revision. read the letter →

arxiv 2510.19600 v2 pith:VKXHZBJ5 submitted 2025-10-22 cs.SE cs.AIcs.CL

Qianli Ma , Siyu Wang , Yilin Chen , Yinhao Tang , Yixiang Yang , Chang Guo , Bingjie Gao , Zhening Xing
show 2 more authors
Yanan Sun Zhipeng Zhang
This is my paper · ORCID
classification cs.SEcs.AIcs.CL
keywords AutoPagepaper-to-pagegenerationmulti-agentLLMsystemsprojectwebpageBenchhallucinationverificationVLM-as-Judgeevaluationscientificcommunication
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

The paper tries to establish that building a scientific project webpage is not a one-shot generation problem but a hierarchical process, and that its multi-agent system, AutoPage, can automate it almost for free. AutoPage splits paper-to-page into narrative planning, multimodal content generation, and interactive rendering, with checker agents that verify each stage against the source PDF and optional human checkpoints that let authors steer the result. The paper reports that AutoPage lifts content and visual quality above end-to-end LLM baselines across the PageBench benchmark and wins a forced-choice human preference study. If the claims hold, researchers could publish accessible, interactive summaries of their papers in minutes and at negligible cost.

What carries the argument

The load-bearing mechanism is AutoPage's coarse-to-fine, verification-gated pipeline. Rather than generating a whole HTML page in one pass, the system first parses the PDF into an asset library, then plans a section-level narrative, then writes text before selecting visuals, then matches a templated design and renders HTML. The 'Checker' agents are the mechanism that enforces grounding: a Content Checker verifies consistency between generated prose and paired figures, and an HTML Checker inspects the rendered page for layout flaws such as oversized images, missing containers, or theme clashes; each checker feeds suggestions back into a revision loop. Optional human checkpoints sit at the con

What would settle it

Have independent human raters blindly score a sample of AutoPage and baseline pages on the same 10-point rubric used in the study, then compute correlation between those ratings and PageBench's automated metrics; if the metrics fail to track human preference, the system's central quality claim is not established.

Watch

Extended reading notes

Core claim

The central claim is that a structured, collaborative pipeline—rather than a monolithic LLM prompt—is the right way to convert an academic paper into an interactive project webpage. AutoPage implements this as three phases: a parser and planner build a narrative blueprint; a text-first generator produces section prose and then selects the most relevant figures and tables; a template matcher and HTML generator render the page with interactive elements. At each phase, a dedicated checker agent compares the generated output against the source paper and issues revision feedback, and optional human-in-the-loop checkpoints allow authors to add, delete, or restyle content. The paper also introduces

Load-bearing premise

The load-bearing premise is that the author-defined PageBench metrics and VLM-as-Judge scores are valid proxies for real content fidelity and visual quality; if those instruments are not calibrated to human judgment, the reported quality gains and cost-efficiency conclusions lose their support.

Editorial extensions

If this is right

  • Researchers can auto-publish project pages for nearly any paper at negligible cost, making interactive dissemination the default rather than a manual chore.
  • Since AutoPage wraps arbitrary end-to-end models, it upgrades the quality of whatever backbone a team already uses, with the largest gains for weaker open-weight models.
  • The PageBench test set and metrics give the first shared baseline for paper-to-page systems, so subsequent work can be compared quantitatively.
  • Because each stage is verified against the source document, the system is designed to catch hallucinations before they reach the final page, which supports factual fidelity of the output.

Reading between the lines

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

  • A testable extension: run the same checker-and-revise loop on poster or slide generation from the same paper; if the gains come from coarse-to-fine planning and verification rather than page-specific templates, similar lifts should appear.
  • The template library of 87 archetypes means output variety is bounded by the library; an editorial bet is that the next bottleneck will be template coverage, not generation quality.
  • The compression-aware accuracy metric rewards strong summarization, so it may over-credit systems that delete nuances; a human-study design that probes recall of small but important details would test whether 'concise' and 'faithful' are actually aligned.
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

5 major / 5 minor

Summary. The paper proposes AutoPage, a multi-agent system that converts a paper PDF into an interactive project webpage through three phases: narrative planning, multimodal content generation, and interactive page rendering, with verifier agents at each phase and optional human checkpoints. The authors also introduce PageBench, a benchmark of ~1,500 real project pages, from which a 100-page test set and an 87-template library are constructed, together with a suite of content and visual quality metrics. The headline claims are that AutoPage improves both content and visual quality over end-to-end baselines such as GPT-4o-mini and Gemini-2.5-Flash, that it is a 'powerful enhancer' for such models, and that it can generate a page in under 15 minutes for under $0.1.

Significance. If the quantitative evaluation were sound, this would be a useful contribution to automated research communication: the task is well-motivated, the coarse-to-fine architecture with dedicated checkers is a plausible design, and the cross-pairing leakage-control protocol in Sec. 5.1 is thoughtful. The user study's forced-choice ranking also provides some evidence of human preference for the final pages. However, the paper's central quantitative claims rest on PageBench metrics whose validity is not established. The VLM-as-Judge rubrics in Appendix H contain a concrete internal inconsistency, several reported numbers in Tab. 1 and Tab. 3 do not match the accompanying text or equations, and no statistical or human-calibration evidence is provided for the metric suite. These issues affect the load-bearing 'powerful enhancer' and cost/efficiency claims, so the contribution is currently conditional on a substantial revision.

major comments (5)
  1. [Sec. 4.2.2 / Appendix H] The VLM-as-Judge prompts and scoring rubrics appear to be interchanged. The 'Element Quality Judge' system prompt instructs scoring only formula/image presentation and text relevance, explicitly ignoring layout, but its 3-5 point rubric describes page layout, white space, image sizes, and visual rhythm. Conversely, the 'Layout Quality Judge' system prompt says to focus only on layout and ignore formula clarity and image relevance, while its 3-5 point rubric describes formulas displayed correctly and images' relevance to paragraphs. If the judges follow the scoring rubric rather than the instruction, then the Tab. 1 metrics 'Visual Content Accuracy' and 'Layout and Cohesion' do not measure what their names claim. Please correct the mismatch, re-run the evaluation, and report the version actually used.
  2. [Sec. 4.2 / Tab. 1 / Sec. 5.3] None of the PageBench metrics is calibrated against human judgment. Visual scores are VLM-generated five-point ratings; content scores use cosine similarity, PPL, and QA pipelines with author-chosen models. Tab. 1 reports single means without error bars, confidence intervals, or significance tests, so the claim that AutoPage 'significantly elevates' content and visual quality is not statistically supported. The user study in Sec. 5.3 captures only overall preference, not the individual content/visual dimensions used in Tab. 1, so it cannot validate those metrics. Please add human-rating correlation per metric, inter-rater agreement, and paired significance tests (e.g., bootstrap or Wilcoxon) over the 100 test pages.
  3. [Sec. B.3 / Eq. (4) / Tab. 3] The Compression-Aware Information Accuracy values are inconsistent with the stated formula S_final = A * ln(C). For GPT-OSS-120B, Tab. 3 reports raw overall accuracy 0.640 and compression 10.833, which gives 0.640*ln(10.833)=1.525, not the reported 1.719. Using the overall accuracy 0.748 gives 1.782, also not 1.719. Similar mismatches occur for Llama-3.1-70B (reported 1.270 vs ~1.08 from 0.320*ln(29.594)). Additionally, the QA adjudication is underspecified: 100 QA pairs are generated by GPT-o3 and six answering models are used, but it is not stated how 'correctly answered' is determined. Please clarify the procedure, recalculate the table, and ensure the numbers satisfy Eq. (4).
  4. [Sec. 5.2 / Tab. 1] The main-results narrative contains factual errors that undermine the reading of the table. The text states that AutoPage-GPT4o-mini 'surpasses the baseline GPT4o-mini across all evaluated metrics,' but Tab. 1 shows readability worsens (10.047 to 10.819, where lower is better). It also states that AutoPage-Gemini-2.5-Flash's Compression-Aware Information Accuracy 'jumps from 1.276 to 1.941,' but Tab. 1 lists AutoPage-Gemini-2.5-Flash at 1.591; 1.941 is AutoPage-GPT4o-mini's value. These misstatements must be corrected so that the claims in the text match the numbers in Tab. 1.
  5. [Abstract / Sec. 6] The headline efficiency claim is inconsistent between sections. The Abstract and Sec. 1 state 'under 15 minutes for less than $0.1,' while Sec. 6 reports costs 'between just $0.06 and $0.20' and turnaround times 'ranging from 4 to 20 minutes depending on the chosen model.' Please specify the exact model/configuration that achieves the abstract's numbers, or revise the abstract to reflect the reported range.
minor comments (5)
  1. [Sec. 5.4] The section title 'Qualitive Study' should be 'Qualitative Study.'
  2. [Sec. 5.3] The last sentence of Sec. 5.3 refers to 'PageAgent's scores'; this should be 'AutoPage's scores.'
  3. [Sec. 6] The discussion refers to 'llama-3.1-80B,' but the model evaluated in Tab. 1 is Llama-3.1-70B. Please fix the name.
  4. [Sec. 5.1 / Tab. 1] The open-source Qwen model is called 'Qwen3-235B' in Sec. 5.1 and 'Qwen3-235B-A22B' in Tab. 1; please unify the notation.
  5. [Tab. 3] Table 3 is difficult to parse: the column grouping is ambiguous, and the labels 'D-Avg', 'U-Avg', and 'Open-Source/Close-Source' should be defined in the caption. A cleaner presentation would help readers verify the compression-aware calculation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: AutoPage's gains are empirical measurements, not predictions forced by fitted inputs; the user study supplies independent support.

full rationale

The paper contains no derivation chain in which a reported result is equivalent to an input by construction. AutoPage is described procedurally (parser, planner, generators, checkers) rather than fitted to PageBench; PageBench's test set and template library are constructed from real project pages and explicitly cross-pair paper content with templates from different papers ('each model must generate a webpage for a given paper using a template derived from a completely different paper's project website'), which avoids the main leakage route. The metrics in Sec. 4.2 are defined before and independently of AutoPage's outputs; even though the authors designed them, that makes them potentially uncalibrated, not circular. The central enhancer claim is also backed by a forced-choice human preference study (Sec. 5.3) that does not rely on the VLM rubrics. Concerns raised by the manuscript itself—VLM rubric content in Appendix H appearing interchanged, no human calibration, and the abstract's '<$0.1' vs Sec. 6's '$0.06-$0.20'—are validity/consistency risks, not reductions to inputs. No load-bearing self-citation is present; the VLM-as-Judge and compression-aware formulations cite external prior work. Therefore no defined circularity pattern is exhibited.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No physical entities, forces, or fitted scientific constants are introduced. The design-level thresholds listed above are hand-set choices that shape the benchmark results. The empirical claims depend on domain assumptions about the validity of LLM-based evaluation and the representativeness of PageBench.

free parameters (4)
  • Maximum page sections (Planner constraint) = 5
    The Section Generation prompt in Appendix H instructs 'The number of sections you design must not exceed five.' This hand-set cap shapes content completeness and is not varied or ablated.
  • Table count cap (Content Checker and Revise) = keep 2; total tables <= 3
    The Full Content Review prompt restricts tables to 3 and asks to keep the 2 most important; this constraint affects Visual Content Accuracy and Layout metrics and is not ablated as a design choice.
  • QA pairs per paper for Compression-Aware Accuracy = 100
    Appendix B.3 generates 100 questions per source paper. Metric stability with respect to this number is not reported.
  • Compression weighting in S_final = A * ln(C) = 1 (multiplicative, log-weighted)
    The product of accuracy and log compression ratio is a hand-chosen scoring rule; alternative weightings would change system rankings.
assumptions (6)
  • domain assumption VLM-as-Judge scores are a valid proxy for human visual-quality judgments
    All Visual Quality metrics use VLM-as-Judge (Sec. 4.2.2, Appendix H) with no calibration against human ratings or inter-rater agreement.
  • domain assumption The QA pipeline faithfully measures factual fidelity
    Compression-Aware Info Acc (Sec. 4.2.1, App. B.3) uses GPT-o3-generated questions and six LLM answerers, but the correctness adjudication protocol is not fully specified.
  • domain assumption PageBench's 100-page test set is representative of real project pages
    The test set is clustered and sampled from 1,500 pages from NeurIPS/ICML/ICLR 2023-2025; no external validation or comparison to pages outside these venues is provided.
  • domain assumption 20 undergraduate/graduate students are representative expert evaluators
    The user study (Sec. 5.3, App. E) relies on 20 students with no power analysis or stated domain expertise.
  • domain assumption Cross-pairing papers with templates from different project pages prevents information leakage
    Sec. 5.1 claims this decoupling prevents copying, but the template contents are not shown to be stripped of the source paper's text or figures.
  • ad hoc to paper Checker agents materially reduce hallucination and layout errors
    The ablation (App. A) supports this within the paper's own metric suite, but checkers are not evaluated for precision/recall against human-identified errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Human-Agent Collaborative Paper-to-Page Crafting." pith.science (2026). https://pith.science/paper/VKXHZBJ5

@misc{pith2026251019600,
  author       = {Pith},
  title        = {Pith review of: Human-Agent Collaborative Paper-to-Page Crafting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VKXHZBJ5}},
  note         = {Machine review of arXiv:2510.19600}
}
abstract

In the quest for scientific progress, communicating research is as vital as the discovery itself. Yet, researchers are often sidetracked by the manual, repetitive chore of building project webpages to make their dense papers accessible. While automation has tackled static slides and posters, the dynamic, interactive nature of webpages has remained an unaddressed challenge. To bridge this gap, we reframe the problem, arguing that the solution lies not in a single command, but in a collaborative, hierarchical process. We introduce $\textbf{AutoPage}$, a novel multi-agent system that embodies this philosophy. AutoPage deconstructs paper-to-page creation into a coarse-to-fine pipeline from narrative planning to multimodal content generation and interactive rendering. To combat AI hallucination, dedicated "Checker" agents verify each step against the source paper, while optional human checkpoints ensure the final product aligns perfectly with the author's vision, transforming the system from a mere tool into a powerful collaborative assistant. To rigorously validate our approach, we also construct $\textbf{PageBench}$, the first benchmark for this new task. Experiments show AutoPage not only generates high-quality, visually appealing pages but does so with remarkable efficiency in under 15 minutes for less than \$0.1. Code and dataset will be released at $\href{https://mqleet.github.io/AutoPage_ProjectPage/}{Webpage}$.

Figures

Figures reproduced from arXiv: 2510.19600 by the authors.

Figure 1
Figure 1. Overview of our work. (a) End-to-end LLMs directly convert papers into project pages, resulting in unreasonable layouts and lacking human feedback. (b) Our proposed AutoPage integrates human-agent col￾laboration into automated page generation with higher content and visual quality. The figure also illustrates the process for constructing the PageBench benchmark. showcase demos. However, this is a manual, repet￾itive… view at source ↗
Figure 2
Figure 2. Overview of AutoPage. AutoPage conducts a multi-agent pipeline for transforming papers into interactive webpages: (1) Narrative Planning and Structuring parses PDFs into Markdown and generates section-level outlines; (2) Multimodal Content Generation produces coherent text–visual sections; (3) Interactive Page Rendering matches templates, compiles full HTML pages, and performs final layout checks. Throughout all pha… view at source ↗
Figure 3
Figure 3. Human preference study. The bar chart shows that AutoPage attains the highest user preference score, surpassing all baselines with more informative, coherent, and visually engaging webpages. user experience. This conclusion is decisively sup￾ported by our user study (Sec. 5.3), as shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison illustrating AutoPage’s superior generation quality over baselines. The figure highlights four common scenarios where AutoPage demonstrates superior performance: (a) Formula Presentation; (b) Image Layout; (c) Table Presentation; (d) Content Plan…
Figure 5
Figure 5. Figure 5: Comprehensive Evaluation of Model Per￾formance with and without AutoPage. The radar plot shows that integrating AutoPage consistently boosts both content and visual quality, thereby demonstrating its strong advantage in generating more accurate, coher￾ent, and visually…
Figure 6
Figure 6. Figure 6: Visual comparison of baseline and AutoPage. The webpage generated by the baseline (left) exhibits rendering failures, including an inability to display the formula and a distorted layout for images and tables. Conversely, the page generated by AutoPage (right) renders …
Figure 7
Figure 7. Figure 7: Additional Visual comparison of baseline and AutoPage. The baseline-generated page (left) suffers from improper image scaling, leading to a catastrophic visual presentation. In contrast, AutoPage (right) styles the tables to match the page’s theme and displays the imag…
Figure 8
Figure 8. Figure 8: Additional Visual comparison of baseline and AutoPage. The webpage generated by the baseline method (left) fails to render the contents within their respective sections. Conversely, the page generated by AutoPage (right) successfully displays both the textual and visua…
Figure 9
Figure 9. Figure 9: Correcting Page Layout with Human Feedback. The baseline generation (left) results in a poor layout with an oversized image. By integrating human feedback, AutoPage (right) produces a corrected layout with a properly-sized image. Prefect !! Image too large! Before Huma…
Figure 10
Figure 10. Figure 10: Impact of Human Feedback on Visual Layout. The initial page generated without human feedback (left) suffers from a flawed layout, featuring a disproportionately large image. In contrast, after incorporating human feedback, AutoPage (right) corrects the layout and rend…
Figure 11
Figure 11. Figure 11: Impact of Human Feedback on Vertical Layout Spacing. The initial page generated by AutoPage without human feedback (left) exhibits excessive vertical whitespace between content modules, resulting in a sparse and poorly structured layout. In contrast, after incorporati…
Figure 12
Figure 12. Figure 12: Human-in-the-Loop Correction for Page Assets. This figure demonstrates how human feedback is used to refine UI components. The initial output (left) features an incorrect or broken logo image in the header. After processing the feedback, AutoPage generates a corrected…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. OmniPresent: Generating Coherent Presentation Suites from Scientific Papers

    cs.SE 2026-07 conditional novelty 6.5 of 10

    A multi-agent HTML pipeline with shared knowledge and cross-artifact verify-and-repair generates coherent poster/slides/video/page suites from papers and beats specialized baselines on OmniPreBench.

  2. DECKBench: Benchmarking Multi-Agent Frameworks for Academic Slide Generation and Editing

    cs.AI 2026-02 conditional novelty 6.0 of 10

    A benchmark for paper-to-slide generation and multi-turn editing, built from 294 pairs and simulated users, with an editing-evaluation design that is partly circular.

Reference graph

Works this paper leans on

58 extracted references · 3 linked inside Pith · cited by 2 Pith papers

  1. [1]

    If a figure does not appear in the paper content, then original_section should be set to null

    Determine which section of the article each image and table in the figures belongs to, and then add a field called original_section to every figure in the original figures, filling it with the determined section. If a figure does not appear in the paper content, then original_section should be set to null. Your output should be json format

  2. [2]

    Key of these tags istag

    Extract figure and table tags from figure or table captions. Key of these tags istag

  3. [3]

    sur- prise

    Judgelm: Fine-tuned large language models are scalable judges.arXiv preprint arXiv:2310.17631. Zeyu Zhu, Kevin Qinghong Lin, and Mike Zheng Shou. 2025. Paper2video: Automatic video gen- eration from scientific papers.arXiv preprint arXiv:2510.05096. 10 A Ablation on Verifiers We evaluate the performance of AutoPage’s self- correction mechanism by conducti...

    arXiv 2025

  4. [4]

    The teaser figure that appears early in the paper must be included in the content. 21

  5. [5]

    Don’t leave any important figure in the research paper

  6. [6]

    Remove the extracted tag from caption of each figure. Output Format ```json <Your output> ``` /da◎abasePrompt: Section Generation System Prompt You are an expert content planner specializing in creating engaging project pages for research papers. Your role is to analyze research content and plan an effective structure that communicates the research clearl...

  7. [7]

    Research Paper in markdown format

  8. [8]

    List of images extracted from the paper

Show all 58 references
  1. [9]

    assets/paper-picture-8.png

    List of tables extracted from the paper. Your goal is to create a section plan that organizes the research into an effective project page structure. Template Description Please analyze the paper content and identify the key sections that should be included in the project page....

  2. [10]

    The figure_index MUST be an integer starting from 1, and no other text should be used in the figure_index position

  3. [11]

    Each figure should be used at most once, with precise and accurate placement

  4. [12]

    Prioritize pictures and tables based on their relevance and importance to the content

  5. [15]

    If a chapter has multiple tables,only the one most relevantto the chapter should be included

  6. [16]

    Your output must be in JSON format, and the section names in your output must exactly match those in theproject_page_content

  7. [17]

    Please ensure that the images you insert are closely related to the context and align well with the content of the section. Output Format ```json <Your output> ``` /da◎abasePrompt: HTML Generation System Prompt You are an expert web developer specializing in creating professio...

  8. [18]

    The HTML files you generate should follow the format and style of the reference HTML template as closely as possible, but you can replace the content in the reference HTML template

  9. [19]

    The content of your project page should be filled completely with the project page content, without any omissions

  10. [20]

    All content sections in Project Page Content should be properly formatted in your html file

  11. [21]

    Images and tables in Project Page Content should be integrated into your html using the correct image path or table path

  12. [22]

    You should make sure that paths of css files included in the html file should not be changed

  13. [23]

    https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js

    If there is a formula in the generated content, please add the relevant js code such as: <script> window.MathJax = { tex: { inlineMath: [['$','$'], ['\\(','\\)']], displayMath: [['$$','$$'], ['\\[','\\]']] } 22 }; </script> <script src="https://cdn.jsdelivr.net/npm/mathjax@3/e...

  14. [24]

    Figure 1., Table 2.)

    For images and tables, only the caption needs to be retained and the tag before the caption needs to be deleted (e.g. Figure 1., Table 2.)

  15. [25]

    If the table is large, place it in a larger container

    Do not place the table in a small container. If the table is large, place it in a larger container. 9.The image (not table) should be placed in the container to limit its size

  16. [26]

    Formulas should not be placed in separate paragraphs

  17. [27]

    section-text

    Be careful not to let the image overflow the screen, the screen width is generally 1280. You can add a container that is the same width as the screen to limit. Layout Specification • For example, if there are two images in two columns whose aspect ratios are 1.2 and 2 respecti...

  18. [28]

    4.Include at least one image in visualization section if it is included in generated sections

    Include at least one table in experiment and ablation section if these two are included in generated sections. 4.Include at least one image in visualization section if it is included in generated sections. Remember that you should just restrict number of tables under 4, rather...

  19. [29]

    • For each section, you should check whether the text content and captions of figures and tables it includes is tightly related

    Figure Placement and Usage • Verify whether figures are inserted in the correct sections according to their meaning in the paper. • For each section, you should check whether the text content and captions of figures and tables it includes is tightly related. • Check if two fig...

  20. [30]

    • If the correlation between them is weak, please suggest to remove the figure or move it to other section

    Relation between figures and text • Check if the core idea that a figure shows is mentioned in text of its section. • If the correlation between them is weak, please suggest to remove the figure or move it to other section

  21. [31]

    If it is, youshould choose 2 most important table to remain

    Number of tables • You should tell whether the number of tables is more than 2. If it is, youshould choose 2 most important table to remain. • Do not restrict the number of figures. Below is the figures with captions, paths, width and height in the paper: <figures> {{figures}}...

  22. [32]

    Do not suggest adding or deleting entire sections

  23. [33]

    The generated project page content should present the more important parts of the paper content in a concise manner, so your review should not require including too many unimportant details

  24. [34]

    Do not correlate the two sections together

    Remember that the original section of a image is not necessary to be same as the section it belongs to in the page. Do not correlate the two sections together

  25. [35]

    Do not give suggestions of including figure Captions, because they will be included during the generation of html, not full content

  26. [36]

    Do not give suggestion to change any text content in any section, you can just suggest to add or delete or move figures and tables

  27. [37]

    weakness

    Tables and Figures from Ablation section in the paper content should belong to Experiment section in the generated content if Ablation is not included in generated sections. Output Format You must return your review instrict JSON formatwith the following fields: Output Format ...

  28. [38]

    Carefully read theweakness,strength, andsuggestionfields in the review JSON

  29. [39]

    Improve the previously generated content by: • Fixing weaknesses • Preserving strengths • Applying suggestions directly and concretely

  30. [40]

    •Clear and fluent(scientifically precise, grammatically correct, and concise)

    Ensure the revised content is: •Accurate(aligned with the original intent of the paper and figures). •Clear and fluent(scientifically precise, grammatically correct, and concise). •Well-structured(logical flow, correct figure placement)

  31. [41]

    Please do not add or remove any sections

  32. [42]

    Do not change the name of any section in the page content

  33. [43]

    Do not include two identical figures in the page content

  34. [44]

    Output Format ```json <Your output> ``` /da◎abasePrompt: HTML Review System Prompt You are a professional reviewer specializing in images and figures on research project pages

    Do not change any text content. Output Format ```json <Your output> ``` /da◎abasePrompt: HTML Review System Prompt You are a professional reviewer specializing in images and figures on research project pages. Your task is to inspectall images individually and meticulously, con...

  35. [45]

    shrink to 90% of original width

    Width of images • Judge strictly whether each image is too wide, too narrow, or within the ideal visual range (approximately 70–90% of the main text block). • Mark as aweaknessif: –The image exceeds the main text block width. –The image is clearly smaller than 65% of the text ...

  36. [46]

    set both top and bottom margin to 24px

    Vertical spacing • Check if the spacing above and below each image is visually balanced. • Mark as aweaknessif one side is noticeably larger than the other. • Allowed adjustment:specify exact pixel values (e.g., “set both top and bottom margin to 24px”). • Spacing must be cons...

  37. [47]

    shrink to 80% of original width

    Text inside images • Judge whether text inside the image isdisproportionately larger than body text (~12–14px). •Font inside images cannot be changed directly. • Allowed adjustment:suggest resizing thewhole imageby a specific percentage of its original width (e.g., “shrink to ...

  38. [48]

    weaknesses

    Strictness Rules • Only consider the three aspects above: width, vertical spacing, internal text. • Skip tables when checking internal text size. • Provide precise, actionable suggestions using percentages for width adjustments and pixels for margins. • Be conservative: report...

  39. [49]

    Apply all the suggestions carefully to the existing HTML without omitting any improvements

  40. [50]

    Preserve the original formatting, style, and constraints from the existing HTML, unless explicitly adjusted by the suggestions

  41. [51]

    All content sections must remain properly formatted and intact; do not remove or lose any original content

  42. [52]

    Images and tables should retain correct paths and aspect ratios; apply size adjustments or centering only if suggested

  43. [53]

    Maintain responsive design, single-column full-width layout, and professional visual presentation

  44. [54]

    Ensure proper spacing, alignment, symmetry, and column height balance as specified in the previous layout rules

  45. [55]

    Comment <!– width = display_width, height = display_height –> before each image whose size is adjusted according to column width and aspect ratio. 28

  46. [56]

    Center images or tables horizontally within their column or content block where applicable

  47. [57]

    All other previous layout constraints and formatting rules must be respected

  48. [58]

    reason":

    Modify the image size according to the suggestions, making sure it is centered and there should be a certain amount of space between the images. Output Format ```html <html_content> ``` /da◎abasePrompt: Aesthetics Quality Judge System Prompt:You are an extremely strict visual ...

  49. [2023]

    arXiv preprint arXiv:2305.10142

    Improving language model negotiation with self-play and in-context learning from ai feedback. arXiv preprint arXiv:2305.10142. Jiaxin Ge, Zora Zhiruo Wang, Xuhui Zhou, Yi-Hao Peng, Sanjay Subramanian, Qinyue Tan, Maarten Sap, Alane Suhr, Daniel Fried, Graham Neubig, and 1 othe...

    2025 arXiv

  50. [2025]

    InThe Thirteenth Inter- national Conference on Learning Representations

    Cycleresearcher: Improving automated re- search via automated review. InThe Thirteenth Inter- national Conference on Learning Representations. Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, and 1 other...

    2024 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.