2nd/10_Wiki/Topics/AI_and_ML/Research.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-research	Research	10_Wiki/Topics	verified	self	Research-Methodology Literature-Review Deep-Research	none	A	0.9	applied	research methodology literature ai-aided		2026-05-10	pending	language framework python anthropic-sdk

Research

매 한 줄

"매 모든 답은 누군가 이미 reformulated". Research는 매 question → literature → synthesis → novel contribution의 매 disciplined loop — 2026 의 매 AI-aided synthesis (Claude Opus 4.7 deep research, GPT-5 with browsing, Elicit, Consensus, undermind.ai) 가 매 weeks of work 를 매 hours로 단축.

매 핵심

매 Phases

1. Question framing — vague curiosity → specific testable question (PICO, FINER criteria).
2. Literature scoping — keywords, citation graph (forward/backward), Connected Papers / Litmaps.
3. Reading & extraction — structured notes (Zettelkasten, claim-evidence-source).
4. Synthesis — themes, gaps, contradictions.
5. Hypothesis / contribution — what novel claim this work adds.
6. Validation — experiment / proof / case study.
7. Communication — paper, blog, talk.

매 Modern toolchain (2026)

• Search: Semantic Scholar API, Google Scholar, OpenAlex.
• Discovery: Connected Papers, Litmaps, Inciteful (citation graph viz).
• AI synthesis: Claude Opus 4.7 deep-research mode, GPT-5 deep research, Elicit (extracts data per paper), Consensus (claim-level), undermind.ai (deep retrieval).
• Notes: Obsidian + Zotero integration; Logseq; Reflect.
• Reproducibility: Quarto, Jupyter Book, Code Ocean.

매 AI-aided literature review pattern

1. Seed papers (3–5 known relevant) → Connected Papers graph.
2. Snowball (citations both ways) → ~100 candidates.
3. LLM screen abstracts: relevance score 0–10.
4. Top 30 → full-text PDF → AI structured extraction (claim, method, evidence, limitations).
5. AI cluster into themes; human reviews + writes synthesis.

매 안전장치 (필수)

• 매 hallucination 의 적: AI 의 매 fake citation 매 흔함 → DOI 의 매 verify 의 must.
• 매 echo chamber: AI synthesis 의 매 popular sources 매 over-weight → manually 의 매 deliberate diverse sampling.
• 매 confirmation bias: AI 의 매 user의 매 hypothesis 매 align — 매 explicit "steelman opposite" prompt.

매 응용

1. PhD literature review.
2. Industry tech radar / market research.
3. Due diligence (M&A, investment).
4. Pre-implementation prior-art search (patents, OSS).

💻 패턴

Claude deep-research synthesis (verify-first)

from anthropic import Anthropic
import httpx

client = Anthropic()

def synthesize(question: str, papers: list[dict]) -> str:
    """papers: [{title, abstract, doi, year}]"""
    corpus = "\n\n".join(
        f"[{i}] {p['title']} ({p['year']}, doi:{p['doi']})\n{p['abstract']}"
        for i, p in enumerate(papers)
    )
    msg = client.messages.create(
        model="claude-opus-4-7", max_tokens=4096,
        system=("Synthesize evidence. Cite EVERY claim with [index]. "
                "If evidence is weak/contradictory, say so explicitly. "
                "Never fabricate citations."),
        messages=[{"role": "user", "content": f"Q: {question}\n\nPapers:\n{corpus}"}],
    )
    return msg.content[0].text

def verify_dois(text: str, papers: list[dict]) -> list[str]:
    """Hallucination check — every cited DOI must exist in our set."""
    import re
    cited = re.findall(r"doi:(10\.\d+/\S+)", text)
    valid = {p["doi"] for p in papers}
    return [d for d in cited if d not in valid]  # offenders

Semantic Scholar fetch

def search_s2(query: str, limit: int = 50) -> list[dict]:
    r = httpx.get(
        "https://api.semanticscholar.org/graph/v1/paper/search",
        params={"query": query, "limit": limit,
                "fields": "title,abstract,year,citationCount,externalIds"},
    ).json()
    return [{"title": p["title"], "abstract": p.get("abstract") or "",
             "year": p.get("year"), "doi": p.get("externalIds", {}).get("DOI"),
             "cites": p["citationCount"]}
            for p in r["data"]]

Snowball expansion

def snowball(seed_ids: list[str], depth: int = 2) -> set[str]:
    frontier, seen = set(seed_ids), set(seed_ids)
    for _ in range(depth):
        next_frontier = set()
        for pid in frontier:
            r = httpx.get(f"https://api.semanticscholar.org/graph/v1/paper/{pid}/references",
                          params={"fields": "paperId", "limit": 100}).json()
            next_frontier.update(ref["citedPaper"]["paperId"]
                                 for ref in r.get("data", [])
                                 if ref["citedPaper"].get("paperId"))
        frontier = next_frontier - seen
        seen.update(frontier)
    return seen

Structured extraction prompt

EXTRACT_PROMPT = """Extract from this paper as JSON:
{
  "claim": "main thesis in one sentence",
  "method": "how they tested it",
  "evidence": "key result with numbers",
  "n": "sample size",
  "limitations": ["limit1", "limit2"],
  "novelty": "what this adds vs prior work"
}
If field unknown, use null. Don't invent."""

Steelman opposite (debias)

def steelman(claim: str) -> str:
    return client.messages.create(
        model="claude-opus-4-7", max_tokens=1024,
        messages=[{"role": "user", "content":
            f"Claim: {claim}\n\nWrite the strongest argument AGAINST this, "
            f"citing actual contrary evidence. Be a hostile reviewer."}],
    ).content[0].text

Zettelkasten note (atomic)

---
id: 2026-05-10-1432
tags: [retrieval, rag]
source: [[Lewis-2020-RAG]]
---
# Dense retrieval beats BM25 only when query-doc lexical overlap is low

In Lewis 2020 (Table 3), DPR > BM25 on NaturalQuestions (+6 EM)
but BM25 ≥ DPR on TriviaQA where queries copy doc tokens.

→ Hybrid search is robust: pick BM25 for lexical, dense for paraphrase.

Connects to: [[Hybrid Search]] · [[BM25]] · [[Dense-Retrieval]]

매 결정 기준

상황	Approach
매 빠른 scan (1h)	Elicit / Consensus / Claude deep-research
매 deep dive (1주)	Manual snowball + AI extraction
Systematic review (PRISMA)	PRISMA flow + Covidence + AI screening
매 cutting-edge (preprints)	arXiv-sanity + Twitter/Bluesky + Semantic Scholar alerts
매 industry / OSS	GitHub trending + State of X reports + AI synthesis

기본값: Connected Papers seed → S2 snowball → AI extract → manual synthesis with steelman.

🔗 Graph

• 부모: Scientific Method
• 변형: Literature-Review · Tech-Radar
• Adjacent: Hallucination · RAG

🤖 LLM 활용

언제: literature scan, abstract screening, structured extraction, synthesis draft, steelmanning. 언제 X: novelty claim 의 매 final assertion (LLM 의 매 ground truth 의 X), 매 quantitative meta-analysis (use proper stats software), 매 citation 의 verify 없이.

❌ 안티패턴

• Cite-without-verify: AI 의 매 만들어낸 fake DOI.
• Single-source synthesis: 매 한 paper 의 매 truth로 취급 — 매 replication 의 무시.
• Recency bias: 매 latest preprint 만 → 매 foundational work 의 무지.
• No gap analysis: literature dump 의 매 only — 매 "what's missing" 의 부재 → contribution 의 unclear.
• Hypothesis fishing: 매 data 부터 → 매 post-hoc theory (HARKing).

🧪 검증 / 중복

• Verified (PRISMA 2020 statement, Semantic Scholar API docs, Claude Opus 4.7 deep research, Elicit methodology).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — full rewrite covering methodology + AI-aided synthesis pipeline