2nd/10_Wiki/Topics/Other/Related-Work.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-related-work	Related Work	10_Wiki/Topics	verified	self	Related Work Section Prior Art Literature Review	none	A	0.9	applied	research academic-writing papers		2026-05-10	pending	language framework english-korean research-writing

Related Work

매 한 줄

"매 academic paper 의 mandatory section 의 prior literature 의 position 의 own contribution". ML/CS papers (NeurIPS, ICML, ICLR, ACL, CVPR) 의 standard structure 의 Introduction → Related Work → Method → Experiments → Conclusion. 매 reviewer 의 first read — 매 novelty claim 의 here 의 stand or fall.

매 핵심

매 Purpose

1. Position — 매 own work 의 landscape 의 place.
2. Differentiate — 매 prior 의 limitation 의 explicit, 매 own gap 의 fill.
3. Credit — 매 intellectual lineage 의 acknowledge.
4. Scope — 매 reviewer 의 not-cited prior work 의 reject 의 prevent.

매 Structure (typical)

• Thematic grouping (preferred 2026) — 매 theme 의 paragraph 의 each.
• Chronological — only for survey papers.
• Per-paper — verbose, avoid.

매 Typical Categories (ML paper)

1. Foundation / closest direct prior.
2. Methodology family (e.g., diffusion vs flow-matching).
3. Application domain.
4. Concurrent work (last 6 months).

매 응용

1. Conference paper — 매 0.5-1 page Related Work section.
2. Thesis — 매 standalone chapter (10-30 pages).
3. Grant proposal — 매 "Innovation" section 의 backbone.
4. Patent — 매 "Background of the Invention".

💻 패턴

LaTeX section template

\section{Related Work}

\paragraph{Foundation models for X.}
\citet{vaswani2017} introduced the Transformer, which subsequent work
\citep{devlin2019,brown2020,touvron2023llama} scaled to billions of parameters.
Unlike these, our method targets edge inference (\textsection\ref{sec:method}).

\paragraph{Efficient inference.}
Quantization \citep{dettmers2022int8,frantar2023gptq} and speculative decoding
\citep{leviathan2023speculative,chen2023accelerating} reduce latency, but
neither addresses our setting of dynamic batch size.

\paragraph{Concurrent work.}
\citet{smith2026concurrent} appeared on arXiv in March 2026; we differ in
that we additionally support streaming output.

BibTeX management (.bib)

@inproceedings{vaswani2017,
  title={Attention is all you need},
  author={Vaswani, Ashish and others},
  booktitle={NeurIPS},
  year={2017}
}

@article{brown2020,
  title={Language Models are Few-Shot Learners},
  author={Brown, Tom and others},
  journal={NeurIPS},
  year={2020}
}

Paper graph extraction (Python semanticscholar)

from semanticscholar import SemanticScholar
sch = SemanticScholar()

paper = sch.get_paper("10.48550/arXiv.1706.03762")  # Attention is all you need
for ref in paper.references[:10]:
    print(ref.title, "→", ref.year)

# Forward citations
cites = sch.get_paper_citations(paper.paperId, limit=50)

Related work table (Markdown)

| Method | Modality | Latency | Param-free | Ours |
|---|---|---|---|---|
| GPTQ | LLM | medium | no | -- |
| AWQ | LLM | low | no | -- |
| FlashAttn | LLM | low | yes | similar |
| **Ours** | **LLM+Vision** | **lowest** | **yes** | -- |

LLM-assisted citation finder (Anthropic SDK)

import anthropic
client = anthropic.Anthropic()

resp = client.messages.create(
    model="claude-opus-4-7",
    max_tokens=600,
    tools=[{"type": "web_search_20250305", "name": "web_search"}],
    messages=[{
        "role": "user",
        "content": "Find 5 ICLR/NeurIPS 2024-2026 papers on speculative decoding with multi-token prediction. Return BibTeX."
    }]
)
print(resp.content[0].text)

매 결정 기준

상황	Approach
Conference paper	Thematic, 0.5-1 page
Survey paper	Chronological + thematic
Thesis	Dedicated chapter
Industry blog	"Inspired by" + light citation
Patent	"Background" with prior art table

기본값: thematic grouping, 매 group 당 3-5 cites, concurrent work 의 explicit paragraph.

🔗 Graph

• 부모: Academic-Writing · Research-Methodology
• 변형: Literature-Review

🤖 LLM 활용

언제: paper search 의 expand, group prior work 의 thematically, 매 differentiation paragraph 의 draft. 언제 X: 매 hallucinated citation 의 risk — 매 always verify 의 DOI / arXiv ID.

❌ 안티패턴

• Citation dump (no commentary): "[Smith 2020, Jones 2021, Lee 2022] also did X." — 매 reader 의 differentiation 의 unclear.
• "To the best of our knowledge": 매 cliché — 매 specific 의 prefer.
• Concurrent work 의 ignore: 매 reviewer 의 catch — proactive 의 cite.
• Hallucinated citations: 매 LLM-generated 매 always 의 verify.
• Self-citation 의 over-rely: 매 inflate own lineage.

🧪 검증 / 중복

• Verified (NeurIPS/ICML author guidelines, Goodson "How to Write Related Work" 2024).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — thematic structure + LaTeX template + comparison table