literature-research

Literature Research SOP

Layer Rules

• Layer: sop — wraps MCP tools directly
• Called by: Any tactic or strategy requiring deep paper reading and rigorous analysis
• Calls: alphaxiv MCP tools, semantic-scholar MCP tools (never calls other SOPs)

Purpose

Deep reading. Raw full text, targeted PDF queries. For rigorous analysis, experiment design, and paper writing. This is the highest-depth skill — you read the actual paper content, not summaries.

Use this when you need to:

• Understand exact methodology details (equations, algorithms, architectures)
• Extract specific experimental setup (hyperparameters, datasets, baselines)
• Compare approaches at a technical level
• Design experiments based on prior work
• Write a paper that cites specific claims with precision

This skill reads RAW FULL TEXT. AI summaries are not acceptable at this depth.

Tools

HARD-GATE

AI summaries (fullText: false) are NOT acceptable for this skill.

PROHIBITED:

• Using get_paper_content with fullText: false (AI summaries)
• Basing analysis on abstracts or discover_papers snippets
• Claiming to understand methodology without reading the full methods section
• Citing specific numbers (accuracy, parameters) without reading the results section

REQUIRED:

• Call get_paper_content(fullText: true) for every key paper (minimum 3)
• Use answer_pdf_queries for targeted extraction of specific details
• Read actual equations, tables, and experimental details from full text
• Every claim must be traceable to specific content in the full paper

Workflow

Step 1: Search

Primary (arXiv):

alphaxiv.discover_papers(
  keywords: ["keyword1", "keyword2", "keyword3"],
  question: "Detailed description of papers needed for deep analysis",
  difficulty: 7
)

Supplementary (non-arXiv):

ss.relevanceSearch(
  query: "search terms",
  limit: 20,
  year: "2022-2024"
)

Use higher difficulty (7-10) for research-depth searches — you need comprehensive coverage.

Step 2: Enrich Metadata

ss.paperBatch(
  paper_ids: ["ARXIV:2301.xxxxx", "ARXIV:2302.xxxxx", ...]
)

Step 3: Select Key Papers

Choose 3-10 papers for deep reading based on:

• Direct relevance to your specific research question
• Methodological significance (introduces the technique you're studying)
• Recency (most recent results and baselines)
• Citation impact (highly-cited = foundational)

Fewer papers, read deeply > many papers, read shallowly.

Step 4: Read Raw Full Text

For each selected paper:

alphaxiv.get_paper_content(
  url: "https://arxiv.org/abs/XXXX.XXXXX",
  fullText: true
)

fullText: true returns the raw extracted text — complete paper content including:

• Full methodology sections
• All equations and algorithms
• Complete experimental setup
• Full results tables
• Appendices and supplementary details

Step 5: Targeted PDF Queries

For specific details that need precise extraction:

alphaxiv.answer_pdf_queries(
  url: "https://arxiv.org/pdf/XXXX.XXXXX",
  queries: [
    "What is the exact model architecture?",
    "What hyperparameters were used for training?",
    "What datasets were used for evaluation?",
    "What are the ablation study results?"
  ]
)

Notes:

• Accepts any PDF URL (not just arXiv)
• Multiple queries on the same paper are nearly free (cached)
• Returns filtered page content as XML with page numbers
• Use for: equations, hyperparameters, ablation results, specific claims

Step 6: Citation Graph Expansion

Find important related work:

ss.citations(paper_id: "ARXIV:XXXX.XXXXX", limit: 50)
ss.references(paper_id: "ARXIV:XXXX.XXXXX", limit: 50)

For promising papers from the graph, repeat Steps 3-5.

Tool-Specific Notes

alphaxiv.get_paper_content (fullText: true)

• Returns raw extracted text — slower but complete
• Includes all sections, equations (as LaTeX), tables, figure captions
• May be large (10-30 pages of text) — read carefully, don't skim
• Only works for arXiv papers

alphaxiv.answer_pdf_queries

• Accepts ANY PDF URL (not limited to arXiv)
• Returns XML with <page num="N"> tags showing relevant content
• Multiple queries in one call = efficient (paper is cached after first query)
• Best for: extracting specific facts, numbers, equations, or claims
• Use AFTER reading full text to drill into specific details

ss.citations / ss.references

• Include citation context (the sentence where the paper is cited)
• Include intent (background, methodology, result comparison)
• Include isInfluential flag (significant vs. passing citation)
• Use these to find papers that extend or challenge the work you're reading

Examples

Deep analysis for experiment design: "LoRA variants for LLM fine-tuning"

# Step 1: Search
alphaxiv.discover_papers(
  keywords: ["LoRA", "parameter-efficient", "fine-tuning", "PEFT"],
  question: "Papers proposing variants or improvements to LoRA for LLM fine-tuning",
  difficulty: 7
)

# Step 2: Enrich
ss.paperBatch(paper_ids: ["ARXIV:2106.09685", "ARXIV:2305.14314", ...])

# Step 3: Select top 5 most relevant

# Step 4: Read full text
alphaxiv.get_paper_content(url: "https://arxiv.org/abs/2106.09685", fullText: true)  # Original LoRA
alphaxiv.get_paper_content(url: "https://arxiv.org/abs/2305.14314", fullText: true)  # QLoRA
# ... repeat for all 5

# Step 5: Extract specific details
alphaxiv.answer_pdf_queries(
  url: "https://arxiv.org/pdf/2106.09685",
  queries: [
    "What is the rank r used in experiments?",
    "What is the training compute compared to full fine-tuning?",
    "Which layers have LoRA applied?"
  ]
)

Methodology comparison: "diffusion model sampling strategies"

# Step 1: Search
alphaxiv.discover_papers(
  keywords: ["diffusion", "sampling", "DDPM", "DDIM", "DPM-Solver"],
  question: "Papers proposing fast sampling methods for diffusion models",
  difficulty: 8
)

# Step 4: Read full text of key papers
alphaxiv.get_paper_content(url: "https://arxiv.org/abs/2010.02502", fullText: true)  # DDPM
alphaxiv.get_paper_content(url: "https://arxiv.org/abs/2010.02502", fullText: true)  # DDIM
alphaxiv.get_paper_content(url: "https://arxiv.org/abs/2211.01095", fullText: true)  # DPM-Solver++

# Step 5: Compare specific details
alphaxiv.answer_pdf_queries(
  url: "https://arxiv.org/pdf/2211.01095",
  queries: [
    "What is the FID score with 10 sampling steps?",
    "How does it compare to DDIM at the same step count?",
    "What is the computational overhead of the solver?"
  ]
)

# Step 6: Find newer work
ss.citations(paper_id: "ARXIV:2211.01095", limit: 30)