PaperlayerView exact revision on arXiv

Source-linked AI summary

Paper2Web: Let's Make Your Paper Alive!

Yuhang Chen, Tianpeng Lv, Siyi Zhang, Yixiang Yin, Yao Wan, Philip S. Yu, Dongping Chen

arXiv:2510.15842v1cs.CLcs.CV

TL;DR

Academic papers are mainly disseminated as static PDFs, while existing webpage-generation approaches struggle with layout, interactivity, multimedia integration, and comprehensive evaluation. Paper2Web introduces a paired paper–homepage dataset and multidimensional benchmark, while PWAGENT uses MCP-based iterative refinement to generate interactive academic webpages. PWAGENT improves benchmark quality substantially over end-to-end baselines while maintaining low generation cost.

  • Problem

    Existing academic dissemination and webpage-generation methods provide limited interactivity, multimedia integration, layout quality, and comprehensive evaluation for paper-based websites.

  • Method

    Paper2Web builds a paper–homepage dataset and evaluation suite, while PWAGENT transforms papers into structured interactive webpages through MCP-based iterative refinement.

  • Results

    PWAGENT improves Connectivity and Completeness by roughly 12% on average, gains 28% over the arXiv HTML baseline, and improves MLLM-as-a-Judge scores by 18%.

  • Takeaways & Limitations

    Paper2Web establishes a benchmark and PWAGENT offers a practical, scalable approach for evaluating and generating interactive academic homepages.

  • Takeaways & Limitations

    The work is an initial step that defines scope and standards, while the contribution of multimedia elements to effective academic communication remains an open challenge.

Abstract

from arXiv · show

Academic project websites can more effectively disseminate research when they clearly present core content and enable intuitive navigation and interaction. However, current approaches such as direct Large Language Model (LLM) generation, templates, or direct HTML conversion struggle to produce layout-aware, interactive sites, and a comprehensive evaluation suite for this task has been lacking. In this paper, we introduce Paper2Web, a benchmark dataset and multi-dimensional evaluation framework for assessing academic webpage generation. It incorporates rule-based metrics like Connectivity, Completeness and human-verified LLM-as-a-Judge (covering interactivity, aesthetics, and informativeness), and PaperQuiz, which measures paper-level knowledge retention. We further present PWAgent, an autonomous pipeline that converts scientific papers into interactive and multimedia-rich academic homepages. The agent iteratively refines both content and layout through MCP tools that enhance emphasis, balance, and presentation quality. Our experiments show that PWAgent consistently outperforms end-to-end baselines like template-based webpages and arXiv/alphaXiv versions by a large margin while maintaining low cost, achieving the Pareto-front in academic webpage generation.

1 INTRODUCTION

Paper2Web addresses the difficulty of turning static academic papers into accessible, interactive webpages while preserving core content. It introduces a benchmark, dataset, and PWAGENT pipeline that jointly improve evaluation and webpage generation.

  • 1 INTRODUCTION: Existing PDF, HTML-conversion, template-based, and direct LLM approaches often lose multimedia or textual detail and produce disordered, weakly interactive layouts.Reported failure modes include rigid figure grids, detached captions, missing responsiveness, limited author control, and difficulty handling long contexts.
  • 1 INTRODUCTION: PAPER2WEB provides a dataset linking academic papers with verified project homepages and supports quantitative analysis of web-based scholarly dissemination.The dataset construction combines conference crawling, metadata extraction, relevance assessment, and human resolution of ambiguous cases.
  • 1 INTRODUCTION: PWAGENT iteratively refines content and layout by organizing paper assets, allocating spatial budgets, inspecting rendered views, and issuing targeted MCP edits.The process links segmented screenshots to HTML fragments for localized correction of imbalance, hierarchy, and multimedia placement.
  • 1 INTRODUCTION: PWAGENT improves Connectivity and Completeness by roughly 12% on average, gains 28% over the arXiv HTML baseline, and improves MLLM-as-a-Judge scores by 18%.It also triples the average score of the strongest end-to-end baseline while remaining competitive with template-assisted variants.
  • 1 INTRODUCTION: The benchmark combines Connectivity and Completeness with human-assisted MLLM-as-a-Judge assessments and PaperQuiz-based knowledge-transfer evaluation.These dimensions address webpage quality, interactivity, informativeness, aesthetics, and retention of paper-level knowledge.
  • 1 INTRODUCTION: PWAGENT is an MCP-based agent that transforms academic papers into structured, interactive webpages through end-to-end processing.Its design targets the integration of textual content, multimedia, and navigable presentation.

2 PAPER2WEB: A NEW TASK AND DATASET

PAPER2WEB constructs a large corpus of academic papers and human-created project homepages to study webpage-based scholarly dissemination. Its collection pipeline combines automated extraction, website crawling, LLM filtering, human review, and feature analysis across conferences and categories.

  • 2 PAPER2WEB: A NEW TASK AND DATASET: The collection pipeline harvests project links from papers and code repositories, crawls corresponding webpages, and covers multiple conferences and categories.Figure 4 presents this data-collection pipeline for constructing paired paper–homepage records.
  • 2.1 DATA COLLECTION: An LLM assigns each paper to one of 13 topical categories using its introduction, enabling standardized cross-paper analysis.Metadata extraction also records title, authors, venue, year, and Semantic Scholar citation counts.
  • 2.1 DATA COLLECTION: Project Website Search maps each paper to at most one canonical homepage by scanning paper and repository links, crawling target HTML, applying LLM analysis, and using human review for ambiguous cases.Papers lacking relevant links are defined as having no project homepage.
  • 2 PAPER2WEB: A NEW TASK AND DATASET: The dataset contains 10,716 papers with human-created project homepages and 85,843 without, grouped into 13 conference-based categories.The corpus spans major AI conferences from 2020–2025 and includes metadata such as authors, venues, years, and citation counts.
  • 2.2 DATA CHARACTERISTICS: Manual auditing of 2,000 webpages defines static, multimedia, and interactive sites and characterizes their feature distribution.Many pages remain static, multimedia use is notable, and interactive capabilities remain comparatively rare and unevenly implemented.

3 EVALUATION METRICS

PAPER2WEB evaluates academic webpages through complementary structural, holistic, and knowledge-transfer measures. The suite combines connectivity and completeness analysis, human-verified MLLM judging, and screenshot-based PaperQuiz assessment.

  • 3 EVALUATION METRICS: PAPER2WEB evaluates webpages with Connectivity & Completeness, human-verified MLLM-as-a-Judge, and PaperQuiz.The three modules assess structure and links, holistic quality, and knowledge transfer from webpage screenshots.
  • 3.1 CONNECTIVITY & COMPLETENESS: Connectivity measures link quality and navigation, while Completeness measures preservation of the source paper’s core sections.HTML analysis is supplemented by URL parsing, human evaluation, image–text balance, and information-efficiency priors.
  • 3.1 CONNECTIVITY & COMPLETENESS: The Image–Text Balance Prior penalizes deviation from an ideal 1:1 visual-to-text ratio using a scaled deviation term D.The score is defined through ζ = 5/(1 + γ·D) and Simg-txt = 5 − ζ.
  • 3.1 CONNECTIVITY & COMPLETENESS: The Information Efficiency Prior compares generated length L with median human-designed length W and penalizes ratios r = L/W above one.Its decay is controlled by β, with β=0.6 given as an example.
  • 3.2 HOLISTIC EVALUATION WITH HUMAN-VERIFIED MLLM-AS-A-JUDGE: Human-verified MLLM judging scores webpages from 1 to 5 on interactivity, aesthetics, and informativeness.The dimensions cover responsiveness and usability, layout and visual appeal, and content clarity and coherence.
  • 3.3 PAPERQUIZ: PaperQuiz uses 50 source-paper questions—25 Verbatim and 25 Interpretive—to measure knowledge transfer from webpage screenshots while penalizing verbosity.MLLMs answer using only rendered webpage content, and the penalty discounts high scores caused by excessive text transfer.

4 PWAGENT: A STRONG BASELINE

PWAGENT converts scientific papers into interactive homepages by structuring paper assets in an MCP repository and iteratively refining generated pages. Its workflow spans decomposition, MCP ingestion, and agent-driven visual optimization.

  • 4 PWAGENT: A STRONG BASELINE: PWAGENT transforms papers into interactive, multimedia-rich homepages through structured assets, MCP tools, and iterative optimization.The pipeline provides a foundational draft before refinement of layout, visual coherence, and semantic alignment.
  • 4.1 PAPER DECOMPOSITION: Paper decomposition converts PDFs to Markdown and uses an LLM to extract metadata, reconstruct tables, and model layout and reading order.The resulting machine-readable representation captures key contributions rather than merely summarizing the paper.
  • 4.1 PAPER DECOMPOSITION: The decomposition produces textual, visual, and link assets with synopses, captions, references, URLs, and semantic roles.These asset types preserve context and provide structured access to supplementary materials and related work.
  • 4.2 MCP INGESTION: MCP ingestion stores cross-referenced assets with stable IDs, relational metadata, layout budgets, and standardized tools for retrieval, composition, and editing.Visual elements are aligned with relevant text, links are typed by function, and spatial allocation balances visual density.
  • 4.3 AGENT-DRIVEN ITERATIVE REFINEMENT: The agent initializes a webpage from repository metadata and assets, then repeatedly detects and corrects layout and visual flaws.The workflow continues until corrective actions cease or a predefined iteration limit is reached.
  • 4.3 AGENT-DRIVEN ITERATIVE REFINEMENT: The Orchestrator combines tile-level and global reasoning to prevent overflow, occlusion, cross-section drift, and other visual artifacts.Adjacent tiles are merged after local refinement so inter-section dependencies can be optimized before a final global pass.

5 HOW PWAGENT MAKE PAPER ALIVE?

Experiments compare PWAGENT with end-to-end generators, templates, and existing HTML versions using structural, holistic, PaperQuiz, efficiency, and qualitative analyses. PWAGENT achieves the strongest overall quality while remaining substantially cheaper.

  • 5.1 EXPERIMENT SETUPS: The evaluation compares Oracle websites, end-to-end models, arXiv and alphaXiv HTML versions, and PWAGENT across structural and holistic metrics.Table 1 covers Completeness, Connectivity, and holistic MLLM evaluation, while Table 2 reports PaperQuiz raw and penalized scores.
  • 5.2 MAIN RESULTS: PWAGENT achieves 2% higher LLM-judged completeness than ground truth while balancing text, images, and links.The comparison also shows that arXiv-HTML’s high rule-based connectivity can inflate scores despite 64% lower human ratings.
  • 5.2 MAIN RESULTS: PWAGENT achieves the highest holistic scores, reaching 91% of ground-truth aesthetics, 94% informativeness, and 59% higher interactivity than alphaXiv.It outperforms all generation methods, while alphaXiv’s interactivity is 37% lower than PWAGENT’s.
  • 5.2 MAIN RESULTS: After verbosity penalties, PWAGENT attains the highest overall PaperQuiz score, with best or near-best performance across tasks and models.Its total information coverage rivals arXiv-HTML without relying on excessive text transfer.
  • 5.3 IN-DEPTH ANALYSIS: PWAGENT costs $0.025 per website, reducing cost by 82% versus GPT-4o and 54% versus Gemini while maintaining strong quality and usability.Template-aided open models cost about $0.069, or 2.8× more, without a clear advantage.
  • 5.3 IN-DEPTH ANALYSIS: Qualitatively, PWAGENT preserves paper structure while improving image–text balance, styling, interactivity, and aesthetics over the baselines.The human-designed version remains better, and the case study identifies verbosity, sparse visuals, monotony, and incomplete content among competing methods.

6 RELATED WORK

Related work spans automated HTML generation and derivative scholarly-content creation. Prior systems generate webpages, posters, slides, banners, videos, and other visual formats, increasingly using multi-agent workflows and MCP.

  • 6 RELATED WORK: HTML-generation research covers direct translation, divide-and-conquer, hierarchical generation, mobile interfaces, multi-page websites, and multi-agent development.Benchmarks include Design2Code, Websight, and WebCode2M, while iterative human feedback and model fine-tuning are also used.
  • 6 RELATED WORK: Automated scholarly-content methods include template- and rule-driven systems plus multi-agent pipelines for posters and presentation slides.Poster studies report high fidelity to human designs in layout, summarization, and visual aesthetics.
  • 6 RELATED WORK: Recent work broadens automated visual design to banners, videos, and MCP-enabled agents for complex unstructured academic information.Paper2Agent is cited as an example of MCP use in advanced academic-information workflows.

7 CONCLUSION AND DISCUSSION

PAPER2WEB defines a benchmark and task for generating project homepages from academic papers while identifying challenges in long-context and layout-sensitive generation. PWAGENT narrows the gap between machine- and human-designed webpages, but the work presents an initial framework rather than a definitive solution.

  • 7 CONCLUSION AND DISCUSSION: PWAGENT narrows the gap between machine- and human-designed webpages and offers a practical, scalable approach to web-based scholarly communication.
  • 7 CONCLUSION AND DISCUSSION: The work is an initial step that defines the scope and standards of exploratory web pages rather than presenting a definitive solution.
  • 7 CONCLUSION AND DISCUSSION: Evaluating how multimedia elements contribute to effective academic communication remains an open challenge for future agentic workflows and evaluation methods.

A.1 RULE-BASED METRIC FOR CONNECTIVITY

The connectivity metric counts external links and internal navigations in generated academic webpages. External links contribute only when they are reachable and contextually relevant.

  • A.1 RULE-BASED METRIC FOR CONNECTIVITY: Connectivity separates external links from internal anchor-based navigation within the same webpage.
  • A.1 RULE-BASED METRIC FOR CONNECTIVITY: The method records Sexternal and Sinternal after parsing HTML patterns for outside-domain URLs and local section anchors.
  • A.1 RULE-BASED METRIC FOR CONNECTIVITY: Only external URLs that are reachable and contextually relevant to the webpage content contribute to Sexternal.

A.2 RULE-BASED METRIC FOR COMPLETENESS

The benchmark evaluates webpage completeness through image–text balance and information efficiency, complemented by human ratings across interactivity, aesthetics, informativeness, completeness, and connectivity. These ratings inspect whether pages are usable, informative, visually coherent, complete, and well connected.

  • A.2 RULE-BASED METRIC FOR COMPLETENESS: The image–text prior weights image and text areas by container size, then aggregates them by page area to distinguish prominent images from small icons.This preserves the relationship between visual composition and the overall page layout.
  • A.2 RULE-BASED METRIC FOR COMPLETENESS: The information-efficiency prior compares generated text length L with median human-authored length W, while β controls penalties for verbosity when L > W.Smaller β values impose a stricter penalty on overly verbose text.
  • B HUMAN ANNOTATION AND VERIFICATION DETAILS: Six authors independently rate webpages from 1–5 on Interactivity, Aesthetic, Informative, Completeness, and Connectivity, using tutorials, cross-validation, and monitoring to support objectivity.
  • B.1 INTERACTIVITY: Interactivity covers basic interactions, interactive visualizations, live demos, and navigation aids, including controls that support discoverability, accessibility, and browsing.
  • B.2 AESTHETIC: Aesthetic quality evaluates visual hierarchy, typography, color and contrast, element quality, layout balance, engagement and style, and interface clarity.
  • B.3 INFORMATIVE: Informativeness emphasizes logical flow, depth of content, and scannability through structured sections, substantive explanations, and readable formatting.
  • B.4 COMPLETENESS: Completeness measures core research elements, rich media and artifacts, scholarly utility, and whether content and accompanying media are sufficiently developed and operationally intact.
  • B.5 CONNECTIVITY: Connectivity assesses functional links to research assets, scholarly context through related work, and smooth internal navigation between webpage sections.

C.1 QA DATASET CURATION.

The QA dataset is built by converting each paper to Markdown and prompting an LLM to generate 50 questions: 25 verbatim and 25 interpretive. The curation process enforces aspect coverage, answerability, distractor quality, and exact formatting.

  • C.1 QA DATASET CURATION.: Each paper is converted to Markdown, then o3 generates 50 multiple-choice questions per paper.
  • C.1 QA DATASET CURATION.: The dataset divides questions evenly into 25 verbatim items answerable from the text and 25 interpretive items requiring higher-level comprehension.
  • C.1 QA DATASET CURATION.: Question generation covers research context, motivation, goals, contributions, methodology, and qualitative insights through predefined aspects.
  • C.1 QA DATASET CURATION.: The curation checks that every question has exactly four options labeled A–D and that the total count, distribution plan, and aspect classifications are correct.
  • C.1 QA DATASET CURATION.: Each question must be based on a located fact, classified under one aspect, answerable from the document, and paired with one correct answer plus three distractors.

C.2 EVALUATION WORKFLOW.

This section details the evaluation workflow, including PaperQuiz’s snapshot-based assessment and templates for parsing paper content, expanding content, and filtering visuals. It also illustrates the workflow through case studies of generated website variants and iterative, component-specific refinement.

  • C.2 EVALUATION WORKFLOW.: PaperQuiz queries six MLLM reader models on website snapshots, requiring answers to rely only on visible content and cite supporting webpage regions or return “NA”.The readers comprise three open-source models—LLaVA-OneVision-Qwen2-7B-ov-hf, DeepSeek-V3.2-Exp, and Qwen3-Coder-480B-A35B—and three closed-source models: o1, Gemini 2.5 Flash, and Grok Code Fast 1.
  • D.1 BASELINE TEMPLATE: The baseline generation template instructs an LLM to produce a complete academic website with modern aesthetics, responsive layouts, rich interactions, visualizations, navigation, external links, and implementation safeguards.It specifies features such as galleries, charts, scrolling indicators, theme toggles, responsive CSS, JavaScript initialization practices, and preservation of original content.
  • D.2 PARSING TEMPLATE: The parsing templates require comprehensive, technically detailed paper summaries and filtering of images and tables for website relevance, retaining every visual that enhances research understanding without artificial quantity limits.The summary prompt covers sections including introduction, related work, datasets, methodology, results, applications, and conclusions, while the filtering prompt outputs only relevant image and table arrays.
  • D.2 PARSING TEMPLATE: The parsing template retains all visuals relevant to paper sections or content, removes unsupported images or tables, and avoids artificially limiting their quantity.Relevance is determined by matching or relating captions to paper sections or content, with examples including methodology diagrams, result charts, and data summaries.
  • D.3 ORCHESTRATING TEMPLATE: The orchestrating template directs an MLLM to classify screenshot components and provide precise, component-specific recommendations for navigation, headers, content blocks, and cards.Protocols address alignment, readability, visual hierarchy, image sizing, adaptive layout, line-height, and contrast, and return whether a fix is needed, its category, and suggested fix.
  • E MORE EXAMPLES OF CASE STUDY: The case studies compare website variants generated by different methods for SMIRK [4], Interactive3D [5], Masked Audio Generation using a Single Non-Autoregressive Transformer [6], and MVDream [7].The examples cover four distinct scientific papers and present their alternative generated website variants.
Loading 2510.15842v1…