literature-survey

Literature Survey Skill

Overview

The literature-survey skill performs a comprehensive topic-driven literature survey for AI/ML research. It systematically searches multiple academic databases (arXiv, Semantic Scholar, DBLP), identifies research gaps, performs cross-domain exploration to discover transferable methods, and proposes 2-3 innovation directions with detailed feasibility assessments.

Arguments

Parse $ARGUMENTS as follows:

• Topic (required): The first argument is the research topic string to survey. Example: "vision transformer", "graph neural networks", "federated learning privacy".

• --date-range (optional): Time window for paper search. Default: 1y (1 year). Supported values: 1y, 2y, 3y. Controls the lookback period from today.

• --max-papers (optional): Maximum number of papers to retrieve per search query. Default: 50. Useful for scoping large topics. Values: 10-200.

• --venues (optional): Comma-separated list of conference/journal abbreviations to filter results. Example: --venues NeurIPS,ICML,ICCV. If omitted, all venues are included.

Setup

This skill requires one-time dependency installation. Run:

pip install -r BASE_DIR/scripts/requirements.txt

Replace BASE_DIR with the base directory of the academic-research-plugin project (shown at the top of this skill's loaded context).

Required dependencies typically include:

• requests — for HTTP queries to research databases
• arxiv — Python client for arXiv API
• bibtexparser — for parsing and generating BibTeX
• pandas — for data aggregation and analysis

Workflow

Follow these 7 steps to complete a comprehensive literature survey:

Step 1: Decompose Topic into Search Queries

Break the user's topic into 3-5 focused search queries to capture different aspects and terminology variations:

• Use synonyms and alternative phrasing
• Include sub-problems and related terms
• Target specific methodologies, applications, and theoretical angles

Example: For "vision transformer", generate:

• vision transformer
• ViT image classification
• self-attention computer vision
• visual attention mechanism
• transformer architecture image tasks

Document all queries for reproducibility.

Step 2: Execute Parallel Paper Searches

For each decomposed query, execute a paper search:

python BASE_DIR/scripts/paper_search.py --query "<query>" --max-results 20 --output json

• Each query searches up to 20 papers (scales with --max-papers flag from arguments)
• Output format: JSON with fields: title, authors, year, venue, abstract, url, citations
• Parse JSON results and aggregate papers across all queries
• Remove duplicates (same title or arXiv ID)
• Record the total number of papers found

Step 3: Identify Seminal and Milestone Papers

Perform targeted searches for foundational works in the field:

• Search with a broader date range (e.g., last 5-10 years) to find seminal papers
• Manually identify must-read papers known in the field
• Example: For transformers, include "Attention Is All You Need" (Vaswani et al., 2017)
• Cross-reference high-citation papers (those cited 1000+ times)
• Mark seminal papers in the output with a "Seminal" label

Step 4: Classify Papers into Themes and Identify Research Gaps

Organize all papers into 3-6 coherent sub-themes based on their core contributions:

Sub-themes should cluster papers by:

• Research methodology or technique
• Application domain
• Problem formulation
• Theoretical framework

For each theme, write a 2-3 sentence summary of the papers it contains.

Research gaps are identified by examining what is NOT covered:

• Missing combinations of techniques
• Unexplored application domains
• Unresolved theoretical questions
• Contradictions between papers
• Opportunities where existing methods haven't been tried

Document at least 3-5 distinct gaps with concrete descriptions.

Step 5: Cross-Domain Exploration

Systematically search adjacent fields for transferable methodologies:

Approach:

1. Identify the core methodology or insight from papers in your main topic
2. Reframe it for adjacent domains (e.g., NLP→CV, CV→Audio, Supervised→Unsupervised)
3. Execute new searches in those domains with adapted keywords

Example: Vision Transformers (ViT)

• Core insight: Self-attention architecture from NLP (Transformers) applied to vision patches
• Cross-domain queries:
  • "attention mechanism audio signal processing"
  • "transformer architecture time series forecasting"
  • "self-attention graph neural networks"

Document findings that show successful methodology transfer and what made the transfer effective.

Step 6: Propose Innovation Directions

Generate 2-3 concrete innovation proposals, each with:

• Description: 1-2 sentences explaining the core idea. Combine insights from gaps + cross-domain findings.

• Feasibility Assessment:

  • Data availability: Is sufficient labeled/unlabeled data accessible?
  • Compute requirements: GPU/TPU needs, training time estimates
  • Expected novelty: How different from existing work? Publication venue fit?
  • Timeline: Realistic implementation duration (weeks/months)

• Potential Weaknesses: Critical evaluation of limitations, edge cases, or reasons the idea might not work

• Landing Plan: Concrete first steps

  • What to implement first (MVP)
  • Expected intermediate milestones
  • Success metrics
  • Fallback strategy if initial approach fails

Step 7: Collect BibTeX References

For all papers mentioned in the report, fetch complete BibTeX entries:

python BASE_DIR/scripts/bibtex_utils.py fetch --title "<paper_title>"

• Aggregate all BibTeX entries into a single references.bib file
• Use standard citation keys: FirstAuthorYear format (e.g., vaswani2017)
• Verify all references are parseable and contain required fields (author, title, year, venue)

Output Format

Generate a comprehensive report file at ./output/literature-survey/YYYY-MM-DD-HHMMSS/survey_report.md using this exact structure:

# Literature Survey: <Topic>

**Date:** YYYY-MM-DD | **Papers Found:** N | **Date Range:** [e.g., "Last 1 year"] | **Search Queries:** K

## Paper Summary Table

| # | Title | Authors | Year | Venue | Citations | Notes |
|---|-------|---------|------|-------|-----------|-------|
| 1 | [Title with link to PDF/arXiv] | First Author et al. | YYYY | Conference/Journal | NNNN | Seminal / Key contribution |
| 2 | ... | ... | ... | ... | ... | ... |

---

## Theme Clusters

### Theme 1: <Name>

**Summary**: [2-3 sentences describing papers in this theme]

**Key Papers**:
- Paper A (Year)
- Paper B (Year)
- Paper C (Year)

**Contribution**: [What this theme contributes to the field]

### Theme 2: <Name>

[Same structure as Theme 1]

[Additional themes as needed...]

---

## Research Gaps

1. **Gap Name 1**: [Concrete description of missing research area or unresolved question]

2. **Gap Name 2**: [Another gap]

3. **Gap Name 3**: [Another gap]

[Additional gaps as identified...]

---

## Cross-Domain Findings

This section documents successful methodology transfer opportunities:

- **Finding 1**: [Methodology X from domain Y successfully applied to domain Z because of reason A. Example: Author Year]

- **Finding 2**: [Another cross-domain insight]

- **Finding 3**: [Another cross-domain insight]

---

## Innovation Proposals

### Proposal 1: <Innovation Title>

**Description**: [1-2 sentences of core idea]

**Feasibility**:
- Data: [Availability and requirements]
- Compute: [Estimated GPU/compute needs]
- Novelty: [How different from existing work, target venues]
- Timeline: [Realistic implementation duration]

**Potential Weaknesses**: [Critical evaluation of limitations and risks]

**Landing Plan**:
1. [First concrete step - MVP scope]
2. [Second step - intermediate milestone]
3. [Third step - validation/refinement]
- Success Metrics: [How to measure success]
- Fallback Strategy: [What to try if approach fails]

### Proposal 2: <Innovation Title>

[Same structure as Proposal 1]

### Proposal 3: <Innovation Title>

[Same structure as Proposal 1]

---

## References

See `references.bib` in this directory for complete BibTeX entries.

---

**Generated**: YYYY-MM-DD HH:MM:SS UTC | **Tool**: literature-survey skill v1.0

Output Directory

All outputs are saved to:

./output/literature-survey/YYYY-MM-DD-HHMMSS/

The directory contains:

• survey_report.md — Main report with all findings and proposals
• references.bib — Complete BibTeX file with all cited papers
• search_log.json — Metadata on all executed searches (queries, result counts, dates)
• papers_raw.json — Full JSON dump of all retrieved papers for reference

Tips for Best Results

1. Topic Decomposition: Spend time on Step 1. Better queries lead to more relevant papers.

2. Date Range Selection: Use --date-range 3y for emerging fields (last 3 years of rapid innovation). Use 1y for stable fields with good coverage.

3. Venue Filtering: For rigorous surveys, use --venues NeurIPS,ICML,ICCV,ICLR to focus on top-tier venues.

4. Seminal Papers: Always manually verify that truly foundational papers are included, even if they're older.

5. Gap Identification: Gaps are most valuable when they're specific and actionable (i.e., suggest a concrete research direction rather than vague limitations).

6. Innovation Proposals: The best proposals combine insights from research gaps + cross-domain findings. Avoid purely speculative ideas.

7. Feasibility Assessment: Be honest about compute and data requirements. This makes proposals more credible and actionable.