Arxiv2beamer

Execution Positioning

Act as an academic presentation and LaTeX Beamer execution agent. Based on the auto-detected Beamer main file and the paper source, produce high-quality slides for a doctoral group meeting. Prioritize compilability, visual clarity, and verifiability. Edit files directly instead of giving vague suggestions. Faithfully preserve the original paper's claims and wording intent. Do not hallucinate methods, experiments, results, or numbers that are not present in the source. If information is missing, use conservative wording instead of fabrication. For compilation stability, always remove citation/bibliography commands.

Input Scope

• Main file: auto-detected Beamer .tex file in the current target directory (template + preamble included).
• Paper source (one of the following):
  • Local path: LaTeX paper source files in the same directory (abstract, body text, equations, figure/table references).
  • Arxiv link: call pdf-vision MCP download_arxiv_source to download and extract source files, then parse content from the extracted source.

Dependency Check (Mandatory)

• This skill heavily depends on pdf-vision MCP (download_arxiv_source, get_pdf_metadata, inspect_pdf_visually).
• If pdf-vision MCP is unavailable, immediately warn the user: pdf-vision MCP is unavailable; this skill cannot run, then stop execution.

Beamer Main File Discovery Rule

This rule is only for discovering the Beamer slide entry file, not the Arxiv paper source entry. Scan the target directory for .tex files containing \documentclass{beamer} (or equivalent Beamer class declaration) and select that file as the Beamer main file.

Task Goals

• Keep the preamble unchanged.
• Rewrite or fill content from \section{Introduction} onward (\begin{frame} blocks after that section).
• Produce technically deep doctoral-meeting slides that compile successfully and have solid visual layout quality.

Paper Outline Extraction (Mandatory Before Generation)

Before generating or editing any slide content, read ALL paper .tex source files and reconstruct the paper's original section outline as faithfully as possible.

• Parse the paper structure down to subsection / subsubsection granularity whenever available, e.g., 2.1 xxx, 2.2 xxx, 2.3.1 xxx.
• If explicit subsection headings are absent, summarize the paragraph-level logical units under the nearest section heading.
• Produce a compact but complete outline of the paper's organization before planning slides.
• Use this extracted outline as the primary basis for slide planning.
• Follow the paper's original organization, hierarchy, and narrative order as much as possible; reorder only when strictly necessary for presentation readability.

The extracted outline should cover:

• Title and abstract intent
• Section / subsection / subsubsection titles
• Paragraph-level logical blocks when headings are missing
• A tentative mapping from each major paper block to one or more slide groups

Do not start slide generation until this outline is complete.

Mandatory Execution Loop (Generate -> Audit -> Fix)

Follow this loop strictly. Do not skip visual auditing.

1. Generate and compile

• Update LaTeX content according to this skill.
• Compile with xelatex -interaction=nonstopmode only.
• Do not switch to pdflatex, lualatex, or any other compiler.

2. PDF initialization

• Call get_pdf_metadata to get Total_Pages.

3. Batch visual audit

• Build page batches with a sliding window. Each batch must contain fewer than 4 pages (e.g., 1-3, 4-6, 7-9).
• Before judging any slide image for clipping, occlusion, or truncation, identify the actual source image file(s) used on that slide from the Beamer source.
• Open each source image file first with view_image (or equivalent local-image inspection) so you can see the uncropped original image before looking at the rendered slide page.
• Then call inspect_pdf_visually for the corresponding PDF batch and compare the rendered slide image against the original image.
• When comparing original vs rendered image, explicitly check:
• whether any semantically important region from the original image is missing on the slide;
• whether labels, legends, axes, arrows, callouts, captions inside the bitmap, or borders were cut off;
• whether the image is partially hidden by text/blocks/other figures;
• whether LaTeX options such as trim, clip, or aggressive resizing introduced accidental loss of content.
• Only decide that a slide image is "complete" after this original-vs-rendered comparison. Do not infer completeness from the rendered slide alone.
• Check each page for:
• Overflow: text/equations/tables clipped or out of bounds.
• Overlap: figure/text/column overlap, including image regions hidden by other slide elements.
• Crowding: too much content on one slide (split if needed).
• Whitespace: excessive empty space that needs re-layout.
• Image sizing: key figures (e.g., method overview) too small; usually give a dedicated slide.
• Orientation: landscape figures forced into narrow columns and becoming unreadable.
• If subagents are available and parallel work is allowed in the current session, distribute independent page batches to subagents for visual inspection in parallel. Each subagent must inspect only its assigned batch, and must follow the same original-image-first comparison rule above.
• Parallel audit is valid only for a single immutable PDF snapshot. If any batch reports an issue and you edit/recompile the slides, discard older batch results tied to the previous PDF and launch a fresh audit round on the new PDF. If subagents are unavailable, inspect the batches serially.

4. Self-correction

• In serial mode: if any issue is found, stop checking later batches immediately.
• In parallel mode: once any batch reports an issue, treat all remaining results from the same outdated PDF snapshot as non-authoritative after you edit/recompile.
• Fix the corresponding page in the Beamer main file and recompile.
• Re-check the affected batch on the new PDF until it passes; if the fix changes layout materially, re-run a fresh audit round for all later batches on the new PDF snapshot.
• Continue only after the currently authoritative batches fully pass.

5. Finalization

• Report completion only after all batches pass visual checks.

Content Structure Requirements

Preserve the original paper's section order, subsection hierarchy, and narrative logic as much as possible. Slide planning should be derived from the extracted paper outline first, rather than from an ad-hoc summary. Reorder content only when strictly necessary for presentation readability.

Ensure doctoral-level technical depth with this structure:

• Title Page: keep template mechanisms such as \title and \author, while updating concrete values from the paper.
• Motivation & Problem Definition:
  • Explain the concrete background and problem.
  • Identify limitations/gaps in prior work.
  • State the paper's core insight.
• Related Work:
  • Summarize this section by following the paper's original categorization, presentation order, and stated comparisons.
  • For each category of methods, provide a concise summary grounded in the paper's own Related Work discussion: the core idea of that category, the limitations noted by the paper, and how the paper positions itself relative to that category.
  • Do not introduce your own literature survey, reinterpretation, or extra comparisons beyond what is supported by the source text.
  • After summarizing one category, you may list a few representative method names or abbreviations exactly as they appear in the paper.
  • Do not add citations.
  • Do not invent method names, abbreviations, categories, limitations, or positioning claims not supported by the source.
• Methodology (highest priority):
  • Overview: show the full framework figure.
  • Module Breakdown: split core modules across multiple slides (e.g., Method I / Method II).
  • Technical Depth: reuse key equations from the paper; avoid over-simplification.
• Experiments:
  • Cover all key tables in the paper under the table rules below.
  • Setup: bullet points for datasets, metrics, training/evaluation settings.
  • Main Results: show tables/curves first, then corresponding visualizations from the paper.
  • Ablation: explain contribution/effectiveness of each module.
  • Place each experiment's table and related visualization in nearby frames, table first.
• Conclusion: summarize findings and "What we can learn".

Layout Constraints (Strict)

• Prefer concise bulletized slides; avoid long paragraphs (recommended 5-7 key lines per slide).
• Use exampleblock where helpful to emphasize definitions, assumptions, or key takeaways.
• Use columns appropriately:
• Left side 0.4-0.5\textwidth for text/equations.
• Right side 0.5-0.6\textwidth for figures.
• If a slide is crowded, split the frame rather than shrinking fonts excessively.

Figure and Table Policy

• Figures:
  • Parse real filenames from paper source \includegraphics{...} and reuse them.
  • Do not use example-image or placeholders.
• Tables:
  • Extract tabular from paper table / table* environments.
  • Preserve table semantics when moving to Beamer.
  • Always wrap with \resizebox{0.95\textwidth}{!}{...} to avoid overflow.
  • Put very large tables on dedicated slides.

\begin{table}
  \centering
  \resizebox{0.95\textwidth}{!}{
    \begin{tabular}{...}
      ...
    \end{tabular}
  }
\end{table}

Syntax and Compilation Safety

• Remove all citation commands: \cite{...}, \citep{...}, \citet{...}.
• Remove or comment out bibliography commands: \bibliographystyle{...}, \bibliography{...}.
• If text contains Figure \ref{...} / Table \ref{...}, keep readable wording but remove \ref to avoid ??.
• Strictly check environment closures, especially frame, columns, block, table, and figure.

Output Standard

• Directly edit the auto-detected Beamer main file; do not provide only snippets.
• The result must compile with xelatex -interaction=nonstopmode in the target directory.
• In the completion report, briefly include:
• which section range was rewritten,
• visual audit rounds and pass status,
• any remaining known risks (if any), with page-level specificity.