vr-literature-review

name	vr-literature-review
description	Deep literature review with multi-stage search, citation graph traversal, and full-text analysis. Uses Semantic Scholar API, arXiv HTML reading, and snowball search to produce comprehensive research landscape analysis.
user-invocable	true
argument-hint	["research-topic or brainstorm-file"]

Literature Review (Deep Search)

Systematically investigate, read, and organize prior work on a research topic using multi-stage search and full-text analysis.

Input

$ARGUMENTS - Research topic, keywords, or brainstorm file path (e.g., "LLM reasoning evaluation" or "experiments/brainstorm_20260305_llm_reasoning.md")

Process

Phase 1-4: Deep Paper Search

1. If a brainstorm file is provided, read it to extract the research topic and hypotheses
2. Read .claude/skills/_shared/paper-search-protocol.md and execute a Level 3 search focused on the research topic
3. Use the search results to build the paper reading list for subsequent phases

Phase 5: Structured Paper Analysis

For each paper, produce a detailed analysis:

5a. Basic Information

• Title, Authors, Venue, Year
• Citation count (from Semantic Scholar)
• Reading depth achieved (full-text / abstract-only)

5b. Contribution Analysis

• Problem Statement: What problem does this paper address?
• Main Contribution: 1-2 sentence summary of the key contribution
• Novelty: What is genuinely new vs. incremental improvement?

5c. Method Deep Dive (full-text papers only)

• Architecture/Algorithm: Core technical approach with key equations or design choices
• Training/Optimization: Training procedure, loss functions, hyperparameters
• Key Design Decisions: Why specific choices were made (and alternatives considered)

5d. Experimental Analysis

• Datasets: Which benchmarks/datasets were used
• Baselines: What methods were compared against
• Key Results: Main performance numbers (include specific metrics)
• Ablation Findings: Which components contribute most to performance

5e. Limitations and Open Questions

• Stated Limitations: What the authors acknowledge
• Unstated Limitations: Issues identified through our analysis
• Reviewer Concerns: If available from OpenReview

5f. Relevance Assessment

• Direct Relevance: (high/medium/low) to our research
• What We Can Reuse: methods, datasets, evaluation protocols, baselines
• How We Differ: our approach vs. theirs
• Key Takeaway: one sentence on what to learn from this paper

Phase 6: Research Landscape Synthesis

6a. Taxonomy

Classify all papers into a multi-dimensional taxonomy:

• By approach: (e.g., supervised, self-supervised, RL-based, prompting-based)
• By scale: (e.g., small model, large model, API-based)
• By task focus: (e.g., reasoning, generation, evaluation) Create a comparison table with key dimensions

6b. Evolution Timeline

• Trace the chronological development of the field
• Identify paradigm shifts and their catalysts
• Note which papers triggered new research directions

6c. Performance Landscape

• Create a benchmark comparison table across papers (where applicable)
• Identify state-of-the-art on each benchmark
• Note performance trends over time

Phase 7: Gap Analysis and Research Positioning

7a. Research Gaps

• Unsolved Problems: Problems that remain open despite multiple attempts
• Methodological Gaps: Approaches not yet explored or combined
• Scale Gaps: Model/data scales not yet tested
• Evaluation Gaps: Missing benchmarks, metrics, or evaluation dimensions
• Theoretical Gaps: Missing formal analysis or understanding of why methods work

7b. Opportunity Analysis

• Which gaps are most impactful to fill?
• Which gaps are most tractable given available resources?
• What is the expected novelty level? (incremental / significant / paradigm-shift)

7c. Our Research Positioning

• How does our proposed idea fill identified gaps?
• What is our unique angle compared to the closest existing work?
• What are the risks of our approach (based on lessons from related work)?

Phase 8: BibTeX and Reference Management

Generate accurate BibTeX entries following the BibTeX Generation section in .claude/skills/_shared/paper-search-protocol.md.

• Separate into:
  • references_core.bib - Papers we will definitely cite
  • references_extended.bib - Additional papers for broader context

Output

Save results to the experiment directory.

File Structure

experiments/<experiment-dir>/
├── literature_review.md            # Comprehensive literature review report
├── paper_analyses/                 # Individual detailed paper analyses
│   ├── paper_01_[short_name].md
│   ├── paper_02_[short_name].md
│   └── ...
├── references_core.bib             # Core BibTeX entries (will definitely cite)
├── references_extended.bib         # Extended BibTeX entries (broader context)
└── search_log.md                   # Search queries and discovery process log

literature_review.md Format

# Literature Review: [Research Topic]
Date: [YYYY-MM-DD]
Papers Analyzed: [N] (Full-text: [M], Abstract-only: [K])

## 1. Search Strategy
- Primary keywords: [...]
- Search queries used: [...]
- Sources: Web Search, Semantic Scholar API, Citation Graph
- Snowball rounds: 2 (backward + forward)

## 2. Paper Summary Table

| # | Paper | Venue | Year | Citations | Relevance | Reading Depth |
|---|-------|-------|------|-----------|-----------|---------------|
| 1 | ...   | ...   | ...  | ...       | High      | Full-text     |

## 3. Research Taxonomy

### 3a. Approach Classification
| Category | Papers | Core Idea | Strengths | Weaknesses |
|----------|--------|-----------|-----------|------------|
| ...      | ...    | ...       | ...       | ...        |

### 3b. Evolution Timeline
[Chronological narrative of field development]

### 3c. Performance Landscape
| Method | Benchmark A | Benchmark B | ... |
|--------|-------------|-------------|-----|
| ...    | ...         | ...         | ... |

## 4. Detailed Paper Analyses
[Summary of each paper with cross-references to paper_analyses/ for full details]

## 5. Research Gap Analysis
- **Unsolved Problems**: [...]
- **Methodological Gaps**: [...]
- **Scale Gaps**: [...]
- **Evaluation Gaps**: [...]
- **Theoretical Gaps**: [...]

## 6. Opportunity Assessment
| Gap | Impact | Tractability | Our Fit |
|-----|--------|--------------|---------|
| ... | ...    | ...          | ...     |

## 7. Research Positioning
[How our work fits in the landscape and what makes it unique]

## 8. Key Takeaways for Our Research
- [Actionable insight 1]
- [Actionable insight 2]
- ...

## Next Steps
Based on the literature analysis, run `/vr-experiment-plan [idea]` to create an experiment plan.

search_log.md Format

# Search Log
## Query 1: [query text]
- Source: [WebSearch / Semantic Scholar API]
- Results: [N papers found]
- Relevant: [papers selected]

## Snowball Round 1: Backward from [paper names]
- New papers discovered: [N]
- Added to reading list: [paper names]
...

Notes

• Maximize full-text reading: Always attempt ar5iv HTML first for arXiv papers. Full-text analysis is far superior to abstract-only.
• Use Semantic Scholar API extensively: It provides structured data (citation counts, references, venues) that web search alone cannot.
• Snowball search is critical: The most relevant papers are often found through citation chains, not keyword search. Two rounds (backward + forward) catch both foundational work and cutting-edge follow-ups.
• Record search provenance: Log which search channel discovered each paper in search_log.md. This helps evaluate search strategy effectiveness.
• Handle rate limits gracefully: Semantic Scholar API allows ~100 requests per 5 minutes without an API key. Space requests if needed.
• Ask the user: Before deep reading, present the candidate paper list and ask if they know additional relevant papers or want to adjust the scope.
• Parallel execution: Use parallel tool calls for independent searches (Channel A, B, C can run simultaneously).
• If an experiment directory does not yet exist, create one following the naming convention: experiment_YYYYMMDD_<descriptive-name>
• If no brainstorm file is available, proceed with the investigation using just the topic