literature-search

name	literature-search
description	Academic literature search best practices and strategies.
metadata	{"nanobot": {"always": true}}

Literature Search Skill

This skill guides you on how to effectively search for academic papers.

Core Principles

1. Use English keywords — Academic databases are primarily English-based. Even for Chinese queries, translate to English for searching.
2. Start broad, then narrow — Begin with 2-3 core keywords, then add constraints (year, method, dataset).
3. Exploit multiple sources — arXiv, Semantic Scholar, Google Scholar, PubMed each have strengths.

Keyword Construction

Basic Rules

• Use specific technical terms over general words: "transformer efficiency" beats "fast neural network"
• Include method + task + dataset when possible: "LoRA fine-tuning LLaMA CommonsenseQA"
• Use abbreviations AND full names in OR queries: "(LLM OR large language model)"

Advanced arXiv Query Syntax

cat:cs.CL AND ("chain of thought" OR CoT) AND NOT survey

• cat:cs.CL — Computational Linguistics category
• AND / OR / AND NOT — Boolean operators
• "exact phrase" — Phrase search
• au:Smith_J — Author search

Example Queries by Intent

User Intent	Search Query
"LLM推理加速"	cat:cs.CL AND ("inference acceleration" OR "speculative decoding" OR "KV cache" OR "quantization") AND "large language model"
"Transformer效率"	cat:cs.LG AND ("transformer efficiency" OR "linear attention" OR "flash attention" OR "sparse attention")
"多模态融合"	cat:cs.CV AND ("multimodal fusion" OR "vision language model" OR "cross-modal")
"图神经网络综述"	cat:cs.LG AND ("graph neural network" OR GNN) AND (survey OR review)

Search Strategy Workflow

Step 1: Quick Scan (5 minutes)

• Search with 2-3 core keywords, max_results=10
• Read titles and abstracts only
• Goal: understand the terminology landscape

Step 2: Targeted Search (10 minutes)

• Refine keywords based on Step 1 findings
• Add year constraint if needed: submittedDate:[20240101 TO 20241231]
• Search with max_results=20
• Identify 3-5 most relevant papers

Step 3: Snowballing

• For each key paper, use get_related_papers or citation_graph
• Check references (what they built upon) and citations (who built upon them)
• This often finds the most impactful papers missed by keyword search

Source-Specific Tips

arXiv

• Strengths: Latest preprints, CS/Physics/Math heavy
• Best for: Cutting-edge methods, benchmark papers
• Tip: Use export.arxiv.org API for structured results; check announce_type for updates

Semantic Scholar

• Strengths: Citation counts, influential citations, author disambiguation
• Best for: Finding highly-cited foundational papers
• Tip: Filter by "highly influential" citations rather than raw count

Google Scholar

• Strengths: Broadest coverage, includes books/theses
• Best for: Interdisciplinary topics, finding older foundational work
• Tip: Use "Cited by" and "Related articles" features

Common Pitfalls

• ❌ Using too many AND terms → zero results
• ❌ Searching only in Chinese → misses 90% of literature
• ❌ Ignoring publication year → wasting time on outdated methods
• ❌ Not checking citations → missing follow-up improvements