BZJZ_Material/文档润色流和知识库构建流/claude-scholar/skills/ml-paper-writing/references/literature-research/arxiv-search-guide.md

arXiv Literature Search Guide

Overview

This guide provides workflows for discovering and evaluating recent ML research papers on arXiv. Use this when conducting literature reviews, finding related work, or staying updated on recent publications.

---

Search Strategies

1. Keyword-Based Search

arXiv Search URL Pattern:

https://arxiv.org/search/?searchtype=all&query=KEYWORDS&abstracts=show&order=-announced_date_first

Common ML Search Keywords:

• General ML: machine learning, deep learning, neural networks
• Specific Areas: reinforcement learning, transformer, attention mechanism, graph neural networks
• Applications: computer vision, natural language processing, reinforcement learning
• Methods: self-supervised learning, contrastive learning, foundation models

Tips:

• Combine keywords with + for AND operation
• Use | for OR operation
• Put phrases in quotes for exact matches: "attention is all you need"

2. Category-Based Search

Recommended arXiv Categories for ML:

• cs.LG (Machine Learning)
• cs.AI (Artificial Intelligence)
• cs.CV (Computer Vision and Pattern Recognition)
• cs.CL (Computation and Language)
• cs.NE (Neural and Evolutionary Computing)
• stat.ML (Machine Learning - Statistics)

Category Filter URL:

https://arxiv.org/search/?cat:cs.LG+OR+cat:cs.AI+AND+all:transformer&abstracts=show&order=-announced_date_first

3. Time-Based Filtering

Recent Papers (Last 3 Months):

• Use order=-announced_date_first for newest first
• Manually filter by submission date
• Check paper metadata for submission date

---

Using Chrome MCP for arXiv Search

When available, use Chrome MCP tools for automated arXiv searching:

1. Navigate to arXiv search using Chrome MCP navigation

2. Extract paper information from search results:

   • Paper title
   • Authors
   • arXiv ID
   • Abstract preview
   • Publication date

3. Navigate to individual papers for detailed review

---

Paper Quality Evaluation

Evaluate papers using the 5-dimension criteria below:

Dimension	Weight	Key Points
Innovation	30%	Novelty of contribution
Method Completeness	25%	Clarity and reproducibility
Experimental Thoroughness	25%	Validation depth
Writing Quality	10%	Clarity of expression
Relevance & Impact	10%	Domain importance

Scoring Guidelines (1-5 scale)

Innovation (30%):

• 5: Breakthrough contribution, major impact
• 4: Significant improvement, new insights
• 3: Methodological innovation
• 2: Incremental improvement
• 1: Minor improvements

Method Completeness (25%):

• 5: Complete and rigorous, easily reproducible
• 4: Very detailed, mostly reproducible
• 3: Core method clear, basically reproducible
• 2: Lacks key details
• 1: Unclear description

Experimental Thoroughness (25%):

• 5: Comprehensive multi-dataset, ablation studies
• 4: Multiple datasets, reasonable ablations
• 3: Main experiments complete
• 2: Limited experiments
• 1: Minimal validation

Writing Quality (10%):

• 5: Excellent clarity and rigor
• 4: Clear and well-structured
• 3: Understandable
• 2: Some ambiguity
• 1: Confusing

Relevance & Impact (10%):

• 5: Solves important problem, wide impact
• 4: Important domain problem
• 3: Meaningful contribution
• 2: Niche problem
• 1: Limited impact

Selection Process

1. Screen by title/abstract for relevance
2. Navigate to full paper for detailed review
3. Score each dimension (1-5)
4. Calculate weighted total
5. Rank and select top papers

---

Extracting Paper Metadata

From arXiv Abstract Page (https://arxiv.org/abs/ARXIV_ID):

• Title (from <h1> tag)
• Authors (from .authors element)
• Abstract (from blockquote.abstract)
• Submission date (from .dateline)
• arXiv ID (from URL)
• Categories (from .subjects)
• Comments (if present)
• Code repository (check abstract for GitHub links)

---

Integration with Citation Workflow

After finding relevant papers:

1. Verify citations using Semantic Scholar API (see ../citation-workflow.md)
2. Fetch BibTeX programmatically via DOI
3. Store in bibliography with verification status

---

Common Use Cases

Finding Related Work

When writing a paper, use arXiv search to:

1. Find recent papers on your topic
2. Identify state-of-the-art methods
3. Discover competing approaches
4. Find baseline comparisons

Staying Updated

Set up regular searches for:

• Your specific research area
• Competing labs/researchers
• New methods in your domain
• Conference proceedings (preprints)

Literature Reviews

For comprehensive reviews:

1. Start with broad keyword searches
2. Filter by recent publications (last 1-3 years)
3. Use citation chaining (forward and backward)
4. Evaluate and select high-quality papers
5. Organize by theme/contribution

---

Tips for Effective Searching

1. Use specific keywords rather than broad terms
2. Combine techniques (keywords + categories + time filters)
3. Check code availability (many arXiv papers link to GitHub)
4. Look for citations to understand impact
5. Read abstracts carefully before full papers
6. Use paper metrics (citation count, code stars) as indicators

---

External Resources

• arXiv: https://arxiv.org/
• Semantic Scholar: https://www.semanticscholar.org/
• Papers With Code: https://paperswithcode.com/
• Connected Papers: https://www.connectedpapers.com/
• arXiv API: http://export.arxiv.org/api_help/