Academic Research

This skill provides comprehensive guidance for academic paper search, literature reviews, and research synthesis using the exa-search scripts and arxiv-mcp-server.

When to Use This Skill

• Searching for academic papers on a topic
• Conducting literature reviews
• Finding papers by specific authors
• Discovering recent research in a field
• Downloading and analyzing arXiv papers
• Synthesizing findings across multiple papers
• Tracking citation networks and influential papers
• Researching state-of-the-art methods in AI/ML

Available Tools

Exa Search (Web Search with Academic Filtering)

Scripts: ~/.claude/skills/exa-search/scripts/ (exa_search.py, exa_contents.py, exa_research.py, exa_agent.py)

Key Flags for Academic Search:

• --category publication - Filter results to papers, preprints, and proceedings
• --domains arxiv.org - Restrict to arXiv (repeat for more venues; cannot combine with --category publication)
• --after / --before - Filter by publication date

The hosted Exa MCP server is an alternative; its filters live only in the opt-in web_search_advanced_exa tool (see references/exa-academic-search.md).

ArXiv MCP Server (Paper Search, Download, Analysis)

Tools: search_papers, download_paper, list_papers, read_paper

Capabilities:

• Search arXiv by keyword, author, or category
• Download papers locally (~/.arxiv-papers)
• Read paper content directly
• Deep paper analysis with built-in prompts

Core Workflows

Workflow 1: Quick Paper Discovery

Use case: Find papers on a specific topic quickly

Step 1: Use Exa with the publication category
python3 ~/.claude/skills/exa-search/scripts/exa_search.py \
  "a survey of transformer attention mechanisms" \
  --category publication --highlights --no-text -n 10

Step 2: Review titles and abstracts
Step 3: Note arXiv IDs for deeper analysis

Workflow 2: ArXiv-Focused Search

Use case: Search specifically within arXiv

Step 1: Use arxiv MCP search_papers
search_papers({
  query: "large language models reasoning",
  max_results: 20,
  sort_by: "relevance"
})

Step 2: Download papers
download_paper({ arxiv_id: "2301.00234" })

Step 3: Read and analyze
read_paper({ arxiv_id: "2301.00234" })

Workflow 3: Comprehensive Literature Review

Step 1: Broad discovery with Exa (--category publication)
Step 2: Identify key papers and authors
Step 3: Deep dive with arXiv MCP (download + read_paper)
Step 4: Synthesize findings by methodology/approach

Workflow 4: Recent Developments Tracking

Step 1: Time-filtered Exa search
python3 ~/.claude/skills/exa-search/scripts/exa_search.py \
  "multimodal large language models" \
  --category publication --after 2026-01-01 --highlights --no-text

Step 2: Sort arXiv by submitted_date
search_papers({ query: "multimodal LLM", sort_by: "submitted_date" })

ArXiv Categories Reference

Category	Description
cs.AI	Artificial Intelligence
cs.CL	Computation and Language (NLP)
cs.CV	Computer Vision
cs.LG	Machine Learning
cs.NE	Neural and Evolutionary Computing
stat.ML	Statistics - Machine Learning
cs.RO	Robotics

Academic Domain Filtering

For Exa searches, restrict to academic sources (repeated flags accumulate; omit --category publication, which rejects --domains):

--domains arxiv.org --domains aclanthology.org --domains openreview.net \
--domains proceedings.mlr.press --domains papers.nips.cc --domains openaccess.thecvf.com

Tool Selection Guide

Task	Primary Tool	Alternative
Broad topic search	Exa (--category publication)	arXiv search_papers
ArXiv-specific	arXiv search_papers	Exa with --domains arxiv.org
Download paper	arXiv download_paper	-
Full paper content	arXiv read_paper	-
Code implementations	Exa with --domains github.com	-
Very recent papers	arXiv (submitted_date)	Exa with --after
Multi-hop review, paper lists	exa_agent.py run	-
Bot-protected sites	Obscura --stealth	Scrapling (Turnstile)
Batch stealth scrape	Obscura scrape	-

Source Extraction Escalation

When a source isn't on ArXiv or Exa can't reach it, escalate through:

1. ArXiv MCP → paper is on arXiv (free, full text, best quality)
2. Exa contents → URL known, site allows crawling
3. Firecrawl → JS-heavy site, no anti-bot
4. Obscura --stealth → site fingerprints headless browsers (JSTOR, Scholar, Persée, PubMed, Academia.edu)
5. Scrapling → site uses Cloudflare Turnstile

Obscura Stealth Extraction (Tier 3)

For gated academic sources that block standard headless browsers via canvas/WebGL fingerprinting. Not for bypassing paywalls — for extracting publicly visible metadata, abstracts, and open-access content.

Verified sites (2026-04-24): Google Scholar, JSTOR, Persée, PubMed, Academia.edu, Perseus Digital Library.

# Quick metadata extraction (auto-detects site type from URL)
bash ~/.claude/skills/academic-research/scripts/academic_stealth_fetch.sh URL

# With explicit site type
bash ~/.claude/skills/academic-research/scripts/academic_stealth_fetch.sh URL scholar

# Direct Obscura usage
obscura fetch --stealth --quiet URL --eval "JS_EXPRESSION"

See references/obscura-academic-patterns.md for site-specific JS extraction patterns and gotchas.

Best Practices

1. Start broad with Exa's publication category, then narrow
2. Use date filtering for recent developments
3. Download key papers via arXiv MCP for persistent access
4. Cross-reference multiple search approaches
5. Use technical terms in queries for better results

Domain: Subquadratic Attention

Research domain for post-transformer attention mechanisms that break the O(n^2) barrier. Active area with rapid publication cadence (2024–2026).

Key Papers

Paper	Year	Key Contribution
FlashAttention-2 (Dao)	2023	IO-aware exact attention — foundation for all subsequent work
DuoAttention	2024	Split attention heads into retrieval (sparse) vs streaming (full)
Ring Attention	2024	Distributed sequence parallelism across devices
MoBA (Mixture of Block Attention)	2025	Block-sparse top-k gating with Triton kernel, 1M tokens
NSA (Native Sparse Attention, DeepSeek)	2025	Hardware-aligned sparse attention patterns
TokenSelect	2025	Dynamic per-layer token pruning

Pre-Built Search Queries

# Exa (publication category)
"subquadratic attention mechanism" --category publication --after 2024-01-01
"block sparse attention triton kernel" --category publication
"mixture of attention heads sparse" --category publication
"linear attention transformer approximation" --category publication --after 2024-06-01

# ArXiv (cs.LG + cs.CL)
search_papers({ query: "subquadratic attention sparse transformer", max_results: 20, sort_by: "submitted_date" })
search_papers({ query: "block sparse FlashAttention kernel", max_results: 10 })

Evaluation Criteria

When comparing subquadratic attention mechanisms, benchmark on:

Criterion	What to Measure
Quality	Perplexity degradation vs full attention at target sequence length
Speed	Wall-clock speedup on consumer GPUs (RTX 4090, M4 Max)
Memory	Reduction factor at 128K / 512K / 1M context
Compatibility	Drop-in replacement vs requires retraining
Sparsity	How much computation is actually skipped (e.g., 95% at 1M tokens)

Local Implementation Reference

Working MoBA implementation with Triton kernels: ~/Desktop/Aldea/01-Repos/perplexity-clone/model/moba_block_sparse.py

---

Reference Documentation

For detailed parameters and advanced usage:

• references/exa-academic-search.md - Exa script recipes and MCP tools for academic search
• references/arxiv-mcp-tools.md - ArXiv MCP server tool reference
• references/obscura-academic-patterns.md - Site-specific Obscura extraction patterns with JS expressions and gotchas