Research Paper Study

Complete workflow for studying ML and research papers with summarization, interactive discussion, note compilation, and optional Obsidian integration.

Quick Start

Fetch from arXiv

bash

python scripts/fetch_arxiv.py "https://arxiv.org/abs/1706.03762"
# Or just the ID:
python scripts/fetch_arxiv.py "1706.03762"

Extract from Local PDF

bash

python scripts/extract_pdf_text.py paper.pdf extracted-text.txt

Extract Images from PDF

bash

python scripts/extract_pdf_images.py paper.pdf output-dir/ --prefix arxiv-id --min-size-kb 20

Compile Study Notes

bash

python scripts/compile_study_notes.py arxiv-metadata.json --lang ko --output study-notes.md

Complete Workflow

Phase 1: Acquire Paper

From arXiv:

•Run scripts/fetch_arxiv.py with URL or ID
•Downloads PDF and extracts metadata (title, authors, abstract, year)
•Outputs: {arxiv-id}.pdf and {arxiv-id}-metadata.json

From Local PDF:

•If user provides PDF path, use scripts/extract_pdf_text.py
•Ask user for metadata (title, authors, year, link) or extract from filename
•Create metadata JSON manually for consistency

Phase 2: Summarize Paper

Use sub-agent for deep reading:

•
Spawn isolated session with higher context model (recommended: sonnet or gpt)
•
Provide extracted text + metadata
•
Check USER.md for user role/preferences to determine technical depth

•

Generate technical summary with following structure:

code

Summarize this research paper with technical depth. Include:
- Problem statement with complexity/efficiency analysis
- Key contributions with architectural details
- Methodology: equations, hyperparameters, design choices
- Results: concrete metrics, ablation studies, comparisons
- Computational complexity analysis where relevant
- Limitations and future work

[Language: English/Korean based on user preference]

Do NOT include phrases like "ML Engineer perspective" or "for technical audience" 
in the summary title or section headers. Just write technically by default.

•
Return structured summary to main session

Summary structure:

•Problem definition (include bottlenecks, complexity issues)
•Key contributions (specific innovations with details)
•Methodology (equations, architectures, hyperparameters)
•Experimental results (tables, ablations, comparisons)
•Computational analysis (complexity, efficiency)
•Limitations and future directions

Phase 3: Interactive Discussion

•
Present FULL summary to user:
- •Show the complete summary generated by sub-agent
- •Do NOT summarize or condense it further
- •Do NOT show only "highlights" or "key points"
- •The user wants to read the entire technical analysis
•
User reads and asks questions
•
Answer questions conversationally
•
Track Q&A pairs for compilation:
- •Store each question and answer
- •Maintain order of discussion

Example questions to anticipate:

•"What's novel about this approach?"
•"How does this compare to [other method]?"
•"What are the practical applications?"
•"What experiments did they run?"

Phase 4: Extract Important Figures

If figures are needed:

•Run scripts/extract_pdf_images.py
•Filter by size (default --min-size-kb 20)
•Show extracted images with page numbers.
•Ask user which figures to keep (e.g., “1, 3, 5”).

Naming convention:

•{arxiv-id}-fig{N}.{ext}

Phase 5: Compile Study Notes

When user requests compilation (e.g., "save this" or "create study notes"):

•
Use appropriate template:
- •English: references/hugo-post-template-en.md
- •Korean: references/hugo-post-template-ko.md
•
Fill template with:
- •Metadata (from Phase 1)
- •Study date: Use date +"%Y-%m-%d" to get current date (NOT hardcoded)
- •Summary (from Phase 2)
- •Q&A discussion (from Phase 3)
- •User insights (ask if not provided)
- •Image embeds (from Phase 4, if applicable)
•
Image placement:
- •Do NOT place all figures at the end in a separate section
- •Embed figures inline at contextually appropriate locations
- •
  Examples:
  - •Architecture diagrams → within "Methodology" section when describing the architecture
  - •Attention mechanism diagrams → when explaining attention details
  - •Result graphs → within "Results" section
- •Use Obsidian wiki-link format: ![[attachments/papers/image.png]]
- •Add descriptive captions under each figure
•
Reference links:
- •For cited papers in "References" section, add hyperlinks to arXiv/DOI
- •Format: [Paper Title](https://arxiv.org/abs/XXXX.XXXXX)
•
Section naming:
- •Use "인사이트" NOT "개인적 인사이트" for Korean
- •Use "Insights" NOT "Personal Insights" for English
•
Generate output filename: YYYY-MM-DD-paper-title-slug.md

Phase 6: Generate Hugo Blog Post

After compiling the study notes, generate a Hugo blog post.

Consult references/hugo-integration.md for detailed guidelines.

•
Determine blog root:
- •Default: /Users/gil-yoonhee/VSCodeProject/moripiri.github.io
- •Or check current working directory
•
Create slug from paper title:
- •Lowercase, hyphens for spaces, ASCII only
- •Example: "Attention Is All You Need" → attention-is-all-you-need
- •Max ~60 chars
•
Create page bundle directory:
bash
```
mkdir -p content/posts/<slug>/images
```
•
Copy extracted images:
- •Source: paper-images-<arxiv-id>/ or workspace
- •Destination: content/posts/<slug>/images/
- •
  Rename descriptively if needed:
  - •2512.24601-fig1.png → figure1-architecture.png
  - •2512.24601-fig3.png → figure2-main-results.png

•

Generate frontmatter:

yaml

---
title: "<Paper Title> 요약"  # or "Summary" for English
date: <YYYY-MM-DD>
draft: true
tags: ["Paper Review", "<topic-tag>"]
ShowToc: true
description: "<one-line summary>"
---

•

Add AI notice (MUST be at top of content):

markdown

> **🤖 AI Summary Notice**
> 이 글은 AI(Claude)가 논문을 읽고 작성한 요약입니다. 부정확한 내용이 있을 수 있으니, 정확한 정보는 원문을 참고해주세요.

<!--more-->

**저자:** [Author names]  
**발행년도:** [Publication year]  
**링크:** [arXiv/DOI URL]

•
Convert markdown format and paste content:
- •Convert Obsidian wiki-links → Hugo image syntax
- •![[attachments/papers/image.png]] → <p align="center"><img src="images/image.png" alt="description"></p>
- •IMPORTANT: Keep ALL other content AS-IS from research-paper-study summary
- •DO NOT simplify or restructure the technical content
- •Preserve: complexity analysis, equations, tables, ablations, detailed insights
•
Write to file:
- •Save as content/posts/<slug>/index.md

•

Start Hugo server (optional):

bash

cd /Users/gil-yoonhee/VSCodeProject/moripiri.github.io && hugo server -D

•
Confirm to user:
- •Post location: content/posts/<slug>/index.md
- •Preview URL: http://localhost:1313/posts/<slug>/
- •Status: draft (set draft: false when ready)
- •Images copied: list filenames

Language Support

Default: English

Korean mode: Triggered when:

•User explicitly requests Korean (e.g., "한국어로", "write in Korean")
•USER.md indicates Korean preference
•Use --lang ko flag for scripts

What gets translated:

•Summary content
•Q&A responses
•Template headers and labels
•Output markdown file

What stays in English:

•Paper metadata (title, authors, as published)
•Technical terms (when appropriate)
•arXiv links and identifiers

Dependencies

Python packages required:

•requests (arXiv API)
•pdfplumber (PDF text extraction)
•PyMuPDF (PDF image extraction)

Install if missing:

bash

pip install requests pdfplumber PyMuPDF

Output Example

See references/hugo-post-template-en.md and references/hugo-post-template-ko.md for full templates.

Hugo blog posts are generated as page bundles:

code

content/posts/<slug>/
├── index.md (frontmatter + content)
└── images/
    ├── figure1-architecture.png
    └── figure2-main-results.png

English output:

markdown

# Attention Is All You Need

**Authors:** Vaswani et al.
**Year:** 2017
**Link:** https://arxiv.org/abs/1706.03762
**Studied:** 2024-02-02

## Summary
[Summary content]

## Discussion
### Q: What's the key innovation?
**A:** The Transformer architecture eliminates recurrence...

Korean output:

markdown

# Attention Is All You Need

**저자:** Vaswani et al.
**발행년도:** 2017
**링크:** https://arxiv.org/abs/1706.03762
**학습일:** 2024-02-02

## 요약
[요약 내용]

## 토론 및 질의응답
### Q: 핵심 혁신은 무엇인가요?
**A:** Transformer 아키텍처는 순환 구조를 제거하고...

Advanced Usage

Batch processing:

•Process multiple papers in sequence
•Generate comparison notes across papers

Custom templates:

•User can provide custom template in workspace
•Override default templates if found

Integration with other tools:

•Use hugo server -D for live preview while editing
•Use git to commit and push blog posts
•Generate multiple posts in batch for paper comparison series

Troubleshooting

PDF extraction issues:

•Some PDFs are image-based (scanned) - requires OCR
•Try alternative: Use arXiv source LaTeX if available

arXiv rate limits:

•Be respectful with API calls
•Add delays if fetching multiple papers

Large papers:

•May hit context limits during summarization
•Use chunking strategy or higher context model
•Focus on abstract + key sections only