Paper2Web: Interactive Scientific Publishing

• Paper2Web is a framework that converts static PDFs into dynamic, interactive web artifacts using automated pipelines and multimodal parsing.
• It leverages structured extraction, metadata embedding, and machine-actionable semantics to enable fine-grained navigation and enhanced accessibility.
• Benchmark results demonstrate high connectivity, completeness, and interactivity, driven by iterative human-agent refinements and robust evaluation protocols.

Paper2Web refers broadly to the suite of computational methods, pipelines, and frameworks designed to automatically or semi-automatically convert scientific papers—primarily static PDF documents—into web-native, interactive, and semantically rich digital artifacts. These systems target robust dissemination, fine-grained navigation, interactivity, and enhanced accessibility for scientific content, advancing beyond static HTML rendering to support layout fidelity, executable mechanisms, and knowledge-level benchmarks. Applications range from homepage generation and cross-lingual information retrieval to dynamic web applications that mirror the algorithmic content of research.

1. Conceptual Overview and Scope

Paper2Web initiatives aim to automate the transformation of academic papers into web-based representations featuring interactivity, structured content, and machine-actionable semantics. Distinguished from mere PDF-to-HTML pipelines, modern Paper2Web systems extract document structure (sections, figures, equations), mechanisms (algorithms, simulations), and interaction logic to produce deployable web applications or enriched homepages. Key frameworks explicitly integrate multimodal parsing, agent-driven refinement, and evaluation suites that measure properties such as connectivity, completeness, interactivity, and knowledge retention (Chen et al., 17 Oct 2025, Dai et al., 24 Mar 2026).

Notable domains include:

• Static enrichment: Semantic publishing, metadata embedding, and enhanced navigation (e.g., citation parsing, math rendering) (Lord et al., 2012).
• Homepage/webpage synthesis: Automated generation of project or paper homepages with interactive assets (Chen et al., 17 Oct 2025).
• Mechanism-aware applications: Conversion of dynamic scientific content into interactive web systems, e.g., for algorithm visualization (Dai et al., 24 Mar 2026, Dai et al., 30 May 2026).
• Human-in-the-loop and autonomous agent frameworks: Systems using LLMs, VLMs, and tool APIs to iteratively refine generated artifacts (Chen et al., 17 Oct 2025).

2. Benchmark Datasets and Evaluation Protocols

Recent Paper2Web research emphasizes rigorous, multi-faceted benchmarks:

• Structured Dataset: The Paper2Web dataset comprises 10,716 AI papers with verified project homepages and 85,843 without, each record triple consisting of a paper (PDF/Markdown), metadata, and homepage assets including HTML and screenshots (Chen et al., 17 Oct 2025).
• Annotation Pipeline: Datasets are curated via LLM-driven relevance checks and human adjudication to filter true project homepages.
• Interactivity Benchmarks: I-WebGenBench and PaperVoyager benchmarks include pairs of research papers with expert-built interactive systems as ground truth. These benchmarks cover domains such as algorithms, ML, systems, and more (Dai et al., 30 May 2026, Dai et al., 24 Mar 2026).

Evaluation metrics developed for this task include:

• Rule-based metrics:
  • Connectivity ($S_{\mathrm{conn}}$): Fraction of valid external/internal links.
  • Completeness ($S_{\mathrm{cmp}}$): How faithfully core paper sections and text-image balance are preserved, with penalization for excessive verbosity.
• Holistic scoring: Multimodal LLM (MLLM)-as-a-Judge for interactivity, aesthetics, and informativeness, further validated by domain-expert annotators (Chen et al., 17 Oct 2025).
• Knowledge retention: PaperQuiz—LLM-generated MCQs on paper content, answered by AI models using screenshots of generated pages.
• Interactivity-specific metrics: Build Success Rate (BSR), Interaction Rate (IR), and VLM-judge scores for dynamic behavior verification (Dai et al., 30 May 2026).

Metric	Description	Source
Connectivity	Valid link ratio (internal/external)	(Chen et al., 17 Oct 2025)
Completeness	Section coverage, img–txt balance, verbosity	(Chen et al., 17 Oct 2025)
PaperQuiz QA	Knowledge retention on rendered artifact	(Chen et al., 17 Oct 2025)
Interaction Rate	Events causing DOM change/response	(Dai et al., 30 May 2026)
VLM-Judge Score	Visual, interaction, and topical correctness	(Dai et al., 30 May 2026)

3. Methodologies and Architectures

Paper2Web methods span traditional, hybrid, and agent-based paradigms:

3.1. Structured Autonomous Pipelines

PWAgent: Converts PDF to structured Markdown and then decomposes into JSON assets (text, image, link), ingested by a Model Context Protocol (MCP) server. The agent iteratively refines the generated webpage, applying tools for emphasis, balance, and aesthetics. Iterative correction continues until no further improvements are suggested, or a maximum refinement depth is reached. Core tools apply rule-based CSS transformations and layout adjustments for presentation and clarity (Chen et al., 17 Oct 2025).

3.2. Interactive System Synthesis

PaperVoyager: Operates via a modular three-stage pipeline (Dai et al., 24 Mar 2026, Dai et al., 30 May 2026):

• Paper Understanding: Multimodal layout parsing, extraction of figures, captions, and mechanisms, followed by chain-of-thought prompts for interactable module identification.
• System Modeling: Mechanisms are formalized as $(S, U, T)$, where $S$ is the set of states, $U$ the set of user controls, and $T: S \times U \to S$ the transition function, mirrored as event handlers in React/TypeScript.
• Webpage Synthesis: Each module is generated as a React component, validated via VLM-based evaluation of rendered screenshots, merged into a deployable single-page app.

Block-level candidate generation, rendering, and filtering improve coverage and correctness of interactions, as measured by module coverage and exploration success rates (Dai et al., 24 Mar 2026).

3.3. Semantic and Crowdsourced Enrichment

OntologyNavigator: Uses a graph-based ontology (extended ACM CCS) for IT corpus navigation. The directed graph $O$ contains hierarchical, semantic, and descriptor edges; supports cross-lingual queries via MT and crowdsourced translation; and links ontology nodes directly to meta-queries for bibliographic search. Visualization leverages hyperbolic (“Topic Map”) projections, enabling focus+context navigation and annotation (Kembellec et al., 2011).

3.4. Vision and Layout Extraction Approaches

Sketch2Code: Employs classical and deep segmentation for paper-to-website conversion from drawings and wireframes. DeepLab v3+ models, combined with object detection and OCR, extract structural information from sketches or rendered templates, facilitating domain adaptation and rapid bootstrapping of web code via a JSON DSL (Robinson, 2019).

Semantic Publishing Workflows: Three-tier architectures combine authoring environments (LaTeX, Word), CMS middleware (e.g., WordPress plugins for citation and math injection), and client-side JavaScript for rendering and linked-data export, incrementally enhancing existing publication workflows with minimal author overhead (Lord et al., 2012).

4. Empirical Results and Comparative Analyses

Experiments across benchmarks reveal the following:

• PWAgent achieves state-of-the-art performance on connectivity and completeness, with average holistic scores (MLLM and human) of $S_{\mathrm{conn}}=3.10$ and $S_{\mathrm{cmp}}=3.56$ across held-out samples, and a per-page cost 82% lower than GPT-4o (Chen et al., 17 Oct 2025).
• PaperVoyager delivers an Interactive Exploration Success Rate (SR) up to 84.4% (k=full, VLM screenshots), surpassing proprietary and open-source LLM-based baselines (Qwen-Max 80.2%, GPT5.2 68.1%) (Dai et al., 24 Mar 2026). Ablations confirm that modular block generation and mechanism modeling are critical for high interaction coverage.
• Sketch2Code deep segmentation yields F1 scores of 0.65–0.80 for UI element detection, compared with 0.28–0.57 for classical CV (Robinson, 2019).
• Semantic publishing with linked-data plugins demonstrates tangible benefits for both human and machine consumers, including structured citation extraction and live math rendering, without disrupting author workflows (Lord et al., 2012).

5. Limitations and Open Challenges

Despite substantial progress, several limitations persist:

• Dynamic behavior modeling: LLM-generated web applications frequently achieve high visual fidelity, yet true event-driven interactivity (measured by IR/interaction recall) lags, with failures in properly wiring state transitions and event propagation (Dai et al., 30 May 2026).
• Scalability: Token and sequence constraints restrict end-to-end generation for complex, multi-page, or collaborative web systems.
• Dataset bias and generalization: Many pipelines are tuned on AI/ML papers, Bootstrap-style layouts, or a narrow set of hand-drawn templates, limiting domain generality (Robinson, 2019, Chen et al., 17 Oct 2025).
• Semantic standardization: Incomplete adoption of universal citation, math, or data markup standards inhibits advanced automation and seamless cross-platform compatibility (Lord et al., 2012).
• Human validation: Most high-performance systems depend on a layer of human adjudication, especially for ambiguous cases in homepage classification or translation correctness (Chen et al., 17 Oct 2025, Kembellec et al., 2011).

A plausible implication is that hybrid approaches—combining LLMs, VLMs, modular prompt engineering, symbolic reasoning, and iterative tool use—are currently required for robust end-to-end performance.

6. Future Directions

Emergent research focuses on:

• Joint VLM+LLM fine-tuning for highly structured, mechanism-aware synthesis of interactive systems (Dai et al., 24 Mar 2026, Dai et al., 30 May 2026).
• Continual learning via active user feedback, cross-domain sketch collection, and RL-based refinement of UI generation policies (Robinson, 2019).
• Generalization to new genres: Expansion of benchmarks to cover scientific domains beyond AI and ML, and to accommodate chemical kinetics, multidimensional simulations, or multi-user collaborative platforms.
• Semantic enrichment: Deeper integration of RDFa/JSON-LD, automated discovery of domain-specific predicates, and workflow interoperability with scholarly communication tools (Lord et al., 2012).
• Multi-agent collaboration: Orchestrating sets of specialist agents—text interpreters, diagram analysts, presentation stylists—for end-to-end, high-fidelity Paper2Web rendering (Chen et al., 17 Oct 2025).

Progress along these axes is expected to narrow the gap between static scholarly documents and rich, executable, web-native scientific narratives.