SuperAPI AI Directory
Partner中文AI newsFavorites
Skillssurvey-paper

survey-paper

Writing & research

View repository

Updated 10-15 20:09

About this skill

Survey Paper Skill

Read an arxiv paper (or similar external article) and convert it into an Obsidian survey note.

Workflow

[입력: arxiv ID/URL, 또는 arxiv 페이지 덤프, 또는 블로그/에세이 등 외부 글]
      ↓
[1] 대상 목록 추출 (단일 건이면 바로 진행)
      ↓
[1.5] 링크가 사용자가 설명한 맥락(기관·주제 등)과 실제 내용이 일치하는지 확인.
    불일치하면 그대로 진행하지 않고 사용자에게 먼저 알린다
      ↓
[2] 각 건에 대해:
    ├── 원문 다운로드 및 분석
    ├── Paper Digest 생성 (한 문단 통찰 요약) + 한눈에 요약(프레이밍 한 줄 + 3포인트) 초안
    ├── Iterative Writing으로 세부 정리
    ├── 플로우 다이어그램 생성
    └── Obsidian 페이지 저장 → Survey 폴더 (Digest 바로 아래에 한눈에 요약 블록 포함)
      ↓
[3] 읽기 가이드 생성 (5분 사전 가이드)
      ↓
[4] 품질 체크리스트 검증
      ↓
[5] 미충족 시 수정 후 재검증
      ↓
[6] 세줄 다이제스트 메시지 생성 (채팅 응답에 항상 포함, 노트에 이미 쓴 한눈에 요약을 채팅 형식으로 재사용)

Multi-target processing

When dumping an arxiv page, extract the list:

  • Search results page: extract titles and arxiv IDs
  • Author page: extract the author's list of papers
  • Repeat the above workflow for each item
  • Confirm the list to be processed with the user, then proceed

If the user provides a description of the item along with the link (institution name, topic, research direction, etc.), after reading the body text, be sure to cross-check whether that description matches the actual content. If there are typos or mismatches (specifically, if you determine that the papers are from different institutions or topics), do not proceed silently as-is; first inform the user of the discrepancy and confirm how to proceed (use AskUserQuestion if needed). At this point, the context the user provides may be an instruction to "use it only as a format reference," so judge by distinguishing between the two cases.

Phase 1: Source analysis

For arxiv papers, download with the arxiv MCP tool and then read the full text:

  • mcp__arxiv-mcp-server__download_paper: download the paper
  • mcp__arxiv-mcp-server__read_paper: read as markdown
  • Alternatively, the alphaXiv MCP tool (mcp__alphaXiv__get_paper_content, if needed fullText: true) can also be used. If the output is too large and causes an error, read the automatically saved file in chunks using offset/limit
  • For non-arxiv materials (blog posts, etc.), obtain the source text via WebFetch or similar

Key extraction items:

  1. Research motivation: Why was this research done? What were the existing limitations?
  2. Key insight: What others missed, the unique perspective of this piece
  3. Method: How the insight connects to the method
  4. Results: Concrete comparison with alternatives (if the piece presents only theory/design without experiments, state that fact itself explicitly)
  5. Significance: New possibilities this research opens up

⚠️ Principle of Complete Section-by-Section Coverage: Every section and subsection of the source text must be summarized without omission. If the source consists of Sections 1–7, the survey note must also include the corresponding content in full. No section is omitted, including Introduction, Related Work, Method, Experiments, Results, Discussion, Conclusion, Limitations, etc. This is so that later, by looking at the note alone, one can fully grasp the entire flow and detailed content of the source. Faithfully reflect the key content, figures, tables, and main insights of each section, and if there are subsections, preserve that structure in the summary as well.

If the paper is not an arXiv paper but a blog post/essay, etc., adapt the above template to the actual structure of that source (the source's own subheadings). Omit fields for which there is no arXiv metadata, or state them explicitly, e.g. "Blog essay (not an arXiv paper)."

Phase 2: Generate Paper Digest + at-a-glance summary

Paper Digest Structure (one paragraph, 4 elements):

  1. Context: What is the problem?
  2. Insight: What did others miss?
  3. Solution: Insight → natural connection to method
  4. Evidence: Concrete comparison with alternatives

Core Rules:

  • Assume the reader is smart but unfamiliar with the field
  • On first appearance of technical terms, briefly explain in parentheses Add (e.g., "ReAct (reasoning-action iteration framework)")
  • Use concrete examples/analogies (e.g., "trash can role" >> "focus on specific tokens")
  • Show causal relationships explicitly
  • Compare/contrast with alternatives ("X failed while Y succeeded")
  • Boldto emphasize 2-4 key concepts
  • Cite sources for numbers (예: "93% (Table 1)")
  • If the original text is a pure theoretical/design specification without experimental/benchmark results, explicitly state that fact in the Digest (e.g., "This report does not present experimental results").

Also draft the at-a-glance summary (one-line framing + 3 points) at this stage: While writing the Paper Digest, also compress the core of this text into a one-line framing and 3 points. This goes right below the Digest in the note in Phase 5, and is reused as-is in Phase 7 for the three-line digest for chat. Follow the "One-line Framing Writing Principles" / "3 Points Writing Principles" from Phase 7 (same content, only the format differs for note/chat).

Phase 3: Refine details through Iterative Writing

Build Fact Base (first):

  • Extract only verifiable facts from the original text
  • Cite sources for all numbers
  • Do not use unverified claims

Apply 19 Principles (14 Original + 5 Naturalness):

Category Principle
Factuality Every claim must be backed by the Fact Base, numbers must have sources (e.g., "93% (Table 1)")
Structure Topic-first, state section purpose, conclude with significance
Comprehensibility Define first (include basic terms in the key terms summary), compare-contrast, beginner-friendly, logical flow
Format Be concise, use concrete numbers, prose only (no bullets), maintain level of abstraction
Naturalness No translationese (NL1), no excessive rhetoric (NL2), one idea per sentence (NL3), no word repetition (NL4), connector variety (NL5)

Term Definition Priority Principle:

  • On first appearance of technical terms in the Digest, briefly explain in parentheses
  • Include the paper's basic terms in the Key Terms Summary section (e.g., ReAct, trajectory)
  • Define from basic concepts so that beginners in the field can understand

Section Structure (reflect the actual section structure of the original text as-is):

The survey note must follow the section numbers and titles of the original text exactly. Instead of four fixed sections, the actual structure of the original text (e.g., 1. Introduction, 2. Related Work, 3. Method, 4. Experiments, 5. Results, 6. Discussion, 7. Conclusion) is carried over as-is into h2/h3 headings. Within each section, faithfully reflect the key content, figures, tables, and main insights, and if there are subsections, maintain that structure as well. Through this, it must be possible to reconstruct the entire original text from the note alone.

However, the following elements must be included:

  1. Why do this research? (key question, existing limitations, key insight) — extracted from the Introduction
  2. Method — include a flow diagram
  3. Findings — include a quantitative results table (if it is a theoretical paper with no experiments, omit this item or state "no experiments")
  4. Theoretical significance — extracted from the Conclusion/Discussion

Phase 4: Generate flow diagram

Required diagrams:

  1. Main framework: the core flow of the overall methodology (mermaid or ASCII)
  2. Comparison diagram: existing method vs proposed method (if needed)

Mermaid format example:

graph TB
    A[입력] --> B[처리1]
    B --> C{조건}
    C -->|Yes| D[결과1]
    C -->|No| E[결과2]

ASCII format example:

┌─────────────┐     ┌─────────────┐
│   Step 1    │ ──→ │   Step 2    │
└─────────────┘     └─────────────┘

Visualize the core flow of the methodology so it can be grasped at a glance.

Phase 5: Generate Obsidian page

Filename rules: {핵심키워드}_{연도}_{분류}.md

  • 예: ReTRAC_2026_DeepResearch.md

Save location: /Users/iamseungpil/LSP/study/Study/01_Papers/Survey/[Category]/

  • Existing categories: check the current folder list and choose the category that fits best (e.g., Memory, Meta-cognition, Skill_LM, Sequence_Models, Generative_Models, Scaling_Laws, Safety_Detox, Self_Improving_Agent, Competitive_Programming, AI_Safety, etc.)
  • If there is no suitable category, create a new one, and also create a landing file Self_Training.mdthat follows the style of the existing simple auto-query categories (e.g., <Category>/<Category>.md) (frontmatter title/tags/domain + auto Obsidian query block + in the "Connections" section a [[Survey Index]] link)
  • If a new category was created or it is an index row in a style that displays the number of papers per category, Survey Index.mdis also updated

Page template:

# {제목 한글 번역}

> **Digest**: {Paper Digest 한 문단}

> **한눈에**: {프레이밍 한 줄 — 이 논문/글이 더 큰 흐름·맥락 속에서 어떤 의미를 갖는지}
> 1. {핵심 포인트 1}
> 2. {핵심 포인트 2}
> 3. {핵심 포인트 3}

---

## 메타데이터

| 항목 | 내용 |
|------|------|
| **제목** | {원제} |
| **저자** | {저자 목록} |
| **소속** | {기관} |
| **연도** | {연도} |
| **발표** | arXiv:{ID} (비-arxiv면 "블로그 에세이 (arXiv 논문 아님)" 등으로 명시) |
| **링크** | [arXiv](URL), [GitHub](URL) |
| **키워드** | {핵심 키워드들} |

---

## 왜 이 연구를 하는가?

### 핵심 질문

{한 문장으로 핵심 질문}

### 기존 접근법의 한계

| 한계 | 설명 |
|------|------|
| ... | ... |

### 핵심 통찰

- {통찰 1}
- {통찰 2}

---

## 섹션별 상세 요약

⚠️ 아래는 예시 구조이다. 실제로는 **원문의 섹션 번호와 제목을 그대로** 사용한다.
예를 들어 논문이 "1. Introduction / 2. Related Work / 3. Method (3.1 Overview, 3.2 Module A, 3.3 Module B) / 4. Experiments / 5. Results / 6. Conclusion"으로 구성되어 있다면, 아래 헤딩도 동일하게 따라간다. 블로그 글처럼 공식 섹션 번호가 없는 글은, 원문이 실제로 쓰는 소제목을 그대로 헤딩으로 사용한다.

### {Section 1 제목} (예: Introduction)

{핵심 내용 요약}

### {Section 2 제목} (예: Related Work)

{핵심 내용 요약}

### {Section 3 제목} (예: Method)

#### {3.1 하위 섹션 제목}

{내용}

#### {3.2 하위 섹션 제목}

{내용}

### {Section 4 제목} (예: Experiments)

{벤치마크, 셋업, 베이스라인 등 — 실험이 없는 글이면 이 섹션은 생략}

### {Section 5 제목} (예: Results)

| 모델 | 메트릭1 | 메트릭2 |
|------|---------|---------|
| ... | ... | ... |

{정량적 결과에서 도출된 핵심 발견}

### {Section N 제목} (예: Conclusion)

{결론 및 향후 연구}

---

## 방법 프레임워크 다이어그램

{mermaid 또는 ASCII 다이어그램 — Method 섹션의 핵심 흐름 시각화}

---

## 이론적 의의

### {의의 1 제목}

{설명}

### {의의 2 제목}

{설명}

---

## 관련 연구

- [[관련논문1]] — 연결점 설명
- [[관련논문2]] — 연결점 설명

---

## 핵심 용어 정리

| 용어 | 정의 |
|------|------|
| ... | ... |

---

## 태그
#paper #{연도} #{분류1} #{분류2} #{핵심키워드}

The "at a glance" block of the note must have the same content as the three-line digest for chat created in Phase 7 (the same one-line framing + the same 3 points). The only difference is the packaging: in the note, use a pure Obsidian blockquote (>), without bold titles or [-](url)-> chat-only link syntax. For chat, follow the Phase 7 format separately.

Category classification criteria

Category Keywords
Memory memory, test-time learning, context, retrieval, RAG
Skill_LM reward, RL, RLHF, self-improvement, exploration, skill
Meta-cognition consciousness, self-awareness, metacognition, introspection
Sequence_Models architecture, recurrent, looped, state tracking, sequence modeling, latent space
(new category) If it does not fall under the above, create a new folder based on the topic of the article

Phase 6: Generate reading guide

Separately from the survey note, read the original text before Generate a pre-reading guide to skim in 5 minutes. The purpose of this guide is to help the reader grasp the structure and key points of the original text in advance, thereby increasing the efficiency of the actual reading. If survey notes are "organizing after reading," the reading guide is "navigation before reading."

Generation Principles:

  • A length that someone encountering the original text for the first time can read within 5 minutes
  • Easy tone: explain academic terms in everyday language
  • Acts as a map of the original text's structure: guide where is important and where it's okay to skim lightly
  • Encourage active reading with a worksheet that can be filled in while reading

File Naming Rules: {핵심키워드}_{연도}_읽기가이드.md

  • 예: ReTRAC_2026_읽기가이드.md

Save Location: Same folder as the survey notes

Reading Guide Template:

# 읽기 가이드: {대표 저자} ({연도})

> 이 가이드를 먼저 읽고 원문에 들어가세요. 5분이면 충분합니다.

---

## Obsidian 태그

아래 태그를 이 노트의 프로퍼티에 추가하세요:

| 태그 | 이유 |
|------|------|
| #{핵심주제} | 핵심 주제 |
| #{이론/방법명} | 다루는 이론 또는 방법 |
| #{저자핵심키워드} | 저자가 제안하는 핵심 개념 |
| #{연구유형} | 연구 유형 (실험/리뷰/이론 등) |
| #{분야} | 연구 분야 |

---

## 이 글이 뭘 하려는 건가?

{2~3문장으로 배경과 맥락을 설명. 이 분야를 모르는 사람도 "아, 그런 문제가 있구나"라고 느낄 수 있게.}

**한 줄 요약**: {글의 핵심 주장을 일상 언어로 한 문장}

---

## 논문 구조 미리보기

| 페이지/섹션 | 섹션 제목 | 역할 | 읽기 팁 |
|-------------|-----------|------|---------|
| {p.XX 또는 §1} | {섹션명} | {이 섹션이 하는 일} | {어떻게 읽으면 좋은지 조언} |
| ... | ... | ... | ... |

읽기 팁에는 다음을 포함한다:
- 가볍게/천천히/집중해서 등 속도 조절 안내
- 해당 섹션에서 특히 봐야 할 포인트 (예: "정의에 집중", "Table 3만 확인")
- ⭐ 표시로 가장 중요한 섹션 강조

---

## 미리 알아야 할 용어

| 용어 | 뜻 | 쉽게 말하면 |
|------|-----|------------|
| {영어 용어} ({한글}) | {학술적 정의} | {일상 비유나 쉬운 설명} |
| ... | ... | ... |

원문에서 사전 지식 없이 이해하기 어려운 용어만 선별한다.
"쉽게 말하면" 열에는 비유, 일상 상황, 또는 한 줄 비교를 사용한다.

---

## 워크시트: 읽으면서 채우기

원문을 읽으면서 아래 질문에 답을 적어주세요. 한두 줄이면 충분하고, 모르겠으면 `?`로 남겨두세요.

### Q1. {첫 번째 핵심 섹션 관련} ({페이지/섹션 범위})

{구체적 질문 — 해당 섹션에서 찾아야 할 핵심}

> 답:

### Q2. {두 번째 핵심 섹션 관련} ({페이지/섹션 범위})

{구체적 질문}

> 답:

{... 원문의 주요 섹션 수에 맞춰 Q3~Q6 정도까지 생성}

### Q{N-1}. 내 생각

읽으면서 가장 흥미롭거나 의문이 든 부분은?

> 답:

### Q{N}. 모르겠는 부분

이해가 안 되거나 더 설명이 필요한 부분은?

> 답:

Worksheet Question Writing Rules:

  • Each question must point to a specific section/page of the original text (make it clear where to find the answer)
  • Focus on questions asking "why?" and "how?" rather than simple fact-checking
  • Always include the last two questions ("My thoughts," "Parts I don't understand") to encourage metacognition
  • The number of questions is proportional to the length of the original text: for short texts (10p)은 Q1Q5, for long texts (20p+), about Q1–Q8

Phase 7: Generate a three-line digest message (always at the end, as a chat response)

After creating and delivering the survey notes and reading guide, every time also write a short "three-line digest" message that can be pasted directly into chat and include it in the response. This is not a file but the chat message itself, and is a separate output from the formal survey notes saved in Obsidian. The purpose is to create a short summary that can be shared as-is with colleagues or a study group.

The content of this message (one-line framing + 3 points) must be identical to what was drafted in Phase 2 and already saved as the note's "At a Glance" block in Phase 5. Phase 7 is the step of repackaging that content into a chat format (bold title, [-](url)-> links), not the step of creating new separate content.

Format:

**{논문/글 제목}**
[-]({링크})-> {한 줄 프레이밍 — 이 글이 왜 흥미로운지, 어떤 흐름/맥락 속에 있는지}

1. {핵심 기술적 포인트 1}
2. {핵심 기술적 포인트 2}
3. {핵심 기술적 포인트 3}

When bundling multiple items into a single message, line up the links side by side ([-](url1)[-](url2)) and write the titles side by side in the form "A & B."

One-line Framing Writing Principles:

  • It should not simply be a listing of facts like "this article came from institution X," but a single sentence that points out what meaning this result has within a larger flow, trend, or debate
  • If the user has a hypothesis or interest explicitly stated during the conversation (e.g., "LLMs are now moving toward neuralese," a critical awareness of a particular trend, etc.) and the actual content of the article supports that hypothesis, actively reflect that keyword/framing. However, if the article's content does not actually support that framing, do not use the user's hypothesis as-is; adjust to a framing that matches what the article actually says
  • Do not fabricate or exaggerate anything not in the source text — this message also follows the same factual standards as the survey notes.

Principles for Writing the 3 Points:

  • Choose 3 of the source text's key methods/results that are practically worth referencing or conceptually interesting (not a full summary, but excerpting and selecting)
  • Write each point as one sentence, in natural Korean without translationese.
  • When citing numbers, as with the survey notes, use only accurate figures that match the evidence (tables/sections).
  • If the source text is a theory/design paper without experiments, state that fact itself as one of the points (e.g., "However, this was all presented only as equations and proofs, with no empirical results").

When summarizing multiple items at once:

  • If multiple items can be grouped under a clear common theme or thread (e.g., when several items show different aspects of the same trend), ask the user whether to create a three-line summary for each item individually or to combine them into one integrated message (use AskUserQuestion).
  • If you decide to combine them into an integrated message, the one-line framing should point to the common trend/insight, and the 3 points should either be one drawn from each item (when there are 2 items) or 3 integrated points covering the whole — in either case, naturally include context so the reader can tell which item each point is about.
  • If the items differ clearly in topic, the default is an individual three-line summary per item.
  • In this case, the note's "At a Glance" block (Phase 5) is written individually for each item regardless of whether they are combined — since there is one note per item, it contains only the framing and 3 points corresponding to that note. Only the chat message can combine multiple items.

Example (actual past case, per-item summary):

**NCP-ArchPreview (arXiv:2609.10715)**
[-](https://arxiv.org/abs/2609.10715)-> 사전학습 목적함수에 개념 단위 축을 하나 더 넣는 접근인데 규모가 꿤 크네요

1. 토큰 4개를 평균 풀링해 벡터양자화(코드북)로 이산화하고, 별도의 Concept Module이 다음 개념을 예측하도록 NTP와 함께 학습시켰다
2. 파라미터·연산량을 맞춘 baseline 대비로도 개선 폭이 유지돼, 단순 용량 증가가 아니라 개념 예측 구조 자체의 효과임을 어블레이션으로 확인했다
3. 사전학습 후 1,700만 파라미터짜리 VQ 모듈만 업데이트해도 LoRA·풀파인튜닝과 경쟁하는 경량 도메인 적응 인터페이스가 되고, 투기적 디코딩에 개념 표현을 얹으면 승인 길이도 늘어난다

Example (integrated message grouped under a common theme):

**NCP-ArchPreview & Recurrent Looped Transformer**
[-](https://arxiv.org/abs/2609.10715)[-](https://www.alphaxiv.org/abs/2609.recurrent-looped-transformer)-> 결은 다른 논문인데 같이 놓고 보니 LLM이 토큰(자연어) 대신 내부 잠재 공간에서 사고·기억하는 neuralese 쪽으로 가고 있다는 신호로 읽힙니다

1. NCP-ArchPreview는 토큰 4개를 묶어 만든 벡터양자화 코드북으로 "개념"이라는 별도의 이산 잠재 채널을 새로 만들어 토큰이 아닌 잠재 표현을 명시적 예측 목표로 승격시켰다
2. Recurrent Looped Transformer는 프롬프트·응답 구분 없이 모든 토큰에 걸쳐 연속적인 순환 은닉 상태를 그대로 이어붙여, 텍스트로 드러나지 않는 계산 경로가 시퀀스 전체에 걸쳐 쌓이게 설계했다
3. 목적은 다르지만(효율적 사전학습 vs. RL 리플레이 정합성), 두 논문 모두 사람이 읽을 수 있는 토큰 대신 압축되거나 연속적인 잠재 표현에 계산과 기억을 싣는 방향으로 아키텍처가 움직이고 있다는 공통점을 보여준다

Quality Checklist

Self-verification after generation (all must be ✅ to be complete):

  • Does the Digest include all of Context-Insight-Solution-Evidence?
  • Does the Digest provide parenthetical explanations when technical terms first appear?
  • Is the source cited for every number? (e.g., "93% (Table 1)")
  • Does the key terms list include basic terms (e.g., ReAct, trajectory)?
  • Does the flow diagram show the core of the methodology at a glance?
  • Is there a comparison with existing methods along with specific figures?
  • Are all sections and subsections of the source text summarized without omission? (Section-by-Section Coverage)
  • Can a beginner understand the why-what-how of the source text just by reading this note?
  • Is there no translationese ("~함으로써", "~에 있어서")? (NL1)
  • Is there no exaggerated rhetoric relative to the content ("innovative", "pervading")? (NL2)
  • Are there no over-compressed sentences listing 3 or more verbs? (NL3)
  • Are the same words not repeated in adjacent sentences? (NL4)
  • "However/Therefore/Also" is not mechanically repeated? (NL5)
  • Is the reading guide readable within 5 minutes?
  • Does the paper structure preview table include all major sections?
  • Do the worksheet questions specifically point to the key sections of the original text?
  • Is there an "at a glance" block (one-line framing + 3 points) directly below the Digest in the notes?
  • Was the three-line digest message always generated together?
  • Do the "at a glance" block in the notes and the three-line digest for chat (framing + 3 points) match without contradicting each other?
  • Are the one-line framing and the 3 points of the three-line digest accurately grounded in the actual content of the original text? (No exaggeration or fabrication)

⚠️ If the checklist is not satisfied: Fix the relevant part and re-verify