Paper Interpretation

Goal

Read the whole paper, including text, tables, figures, captions, appendices, and experiment details, then write a Chinese Markdown article that helps non-specialists understand the paper while preserving its academic and industrial value.

Always save the final Markdown under the current project's markdown/ directory. Create the directory if missing. Use a descriptive filename derived from the paper title, for example markdown/论文标题-通俗解读.md.

Intake

Accept either:

• A local PDF path.
• A PDF URL.

If the user gives a URL, download the PDF or use scripts/extract_paper_assets.py to download and extract it. If the PDF is scanned or extraction is incomplete, use OCR or available PDF/image tooling before writing. Do not rely only on abstract/introduction unless the PDF cannot be fully processed; state any limitation explicitly.

Extraction Workflow

Use the helper script when useful:

python3 /Users/digoal/.codex/skills/paper-interpretation/scripts/extract_paper_assets.py INPUT_PDF_OR_URL --out .paper-work

The script creates:

• paper.pdf for the normalized PDF.
• paper_text.md with page-level text.
• paper_tables.md with table candidates when pdfplumber is available.
• figures/ with extracted embedded images when PyMuPDF is available.
• manifest.json with extraction status and warnings.

Read the extracted files and inspect important figures/tables directly when they carry evidence, architecture, or results. If a library is unavailable, use other local tools if present; otherwise continue with the available extraction and mention the gap only if it affects confidence.

Reading Checklist

Before writing, identify:

• Paper title, authors, venue/date when available.
• Research problem and practical motivation.
• Prior methods or baselines the paper contrasts against.
• Proposed method, system architecture, algorithm, data, or theory.
• Main experiments, metrics, datasets, tables, and figures.
• Claimed academic contributions and industrial implications.
• Assumptions, limitations, failure cases, and future work.

Article Structure

Write in Chinese. Make the result accessible, but do not dilute technical substance.

1. 论文定位

Start by answering "这是什么、跟我有什么关系、值不值得读":

• Use 1-3 sentences to explain the real-world problem.
• State academic value: what gap it fills or what new capability/evidence it adds.
• State industrial value: where it can be used or what decision it can improve.
• Give one intuitive analogy tailored to the paper. Avoid reusing examples blindly; for a multi-agent finance paper, an analogy like "给 AI 配了一个完整的投资团队" is appropriate.

2. 前置知识地图

Build a prerequisite knowledge map:

核心概念（必须懂） -> 支撑概念（有助于理解） -> 扩展概念（感兴趣再看）

For each important concept, explain with:

• A chart or structured list.
• An analogy.
• A concrete example.

Use Mermaid when it improves clarity:

flowchart LR
  A["核心概念"] --> B["支撑概念"]
  B --> C["扩展概念"]

3. 论文精读

Use the 5W1H frame and map it to the paper:

Prefer comparison over isolated explanation:

• Compare with prior methods or baselines.
• Compare with a reader's intuitive expectation.
• Explain what changes in the workflow, cost, accuracy, robustness, or scalability.

4. 术语解释

Include a glossary for important terms. Use this template:

**Term（中文名）**
- 是什么：...
- 为什么重要：...
- 现实类比：...

Explain why the term is named that way when useful. Keep terms short and selective; prioritize terms required to understand the paper.

5. 批判性评估：论文强在哪里，边界在哪里

Evaluate, do not merely praise:

• Which assumptions must hold in real deployments?
• Are datasets, baselines, metrics, ablations, or statistical claims convincing?
• Where might the method fail?
• What cost, latency, safety, privacy, reproducibility, or integration risks exist?
• What future improvements are realistic?

Separate paper claims from your inference. Use wording like "论文证明了..." for supported claims and "可推断..." for reasoned extrapolation.

Visual Requirements

Use Markdown-supported visuals at key points:

• Mermaid for process, architecture, concept maps, causal chains, and experiment flow.
• Markdown tables for comparisons, result interpretation, limitations, and term dictionaries.
• Simple SVG only when a static diagram is clearer than Mermaid.
• Text diagrams when they are more readable than formal charts.

Do not add decorative diagrams. Every visual must help explain a concept, method, result, or critique.

Source Discipline

• Cite page numbers, section names, table numbers, or figure numbers whenever possible.
• Do not invent results, claims, datasets, or author intent.
• If extraction is incomplete, clearly label uncertain parts.
• Preserve important formulas or algorithms, but explain them in plain language.
• If the paper has supplementary material or appendices inside the PDF, include relevant evidence from them.

Output Checklist

The final Markdown must include:

• Title.
• Paper metadata.
• 1-3 sentence overview.
• Academic value and industrial value.
• Knowledge map.
• Problem-solution-validation deep read.
• Key figure/table interpretations.
• Term glossary.
• Critical evaluation and future directions.
• References/source notes pointing back to paper sections, pages, figures, or tables.

After saving, report the absolute output path and any extraction limitations.