Survey Pipeline

• AUTO_PROCEED = true — When true, auto-continue through all gates without waiting for confirmation. Set to false to pause at each gate.
• USE_EXISTING_ARXIV_JSON = true — When true, check for existing arxiv_results.json before fresh search. If found, use batch-triage to process all papers.
• MAX_PAPERS = 20 — Maximum number of papers to analyze
• PAPER_DOWNLOAD = true — Download PDFs for deeper analysis
• MIN_PAPERS_FOR_TAXONOMY = 5 — Minimum papers needed to build taxonomy

Stage 0: Brainstorm & Topic Refinement (/survey-brainstorm)

Command: /survey-brainstorm "$ARGUMENTS"

When to run: If the user's idea is fuzzy (e.g., "I want to write a survey about AI systems") rather than a specific subfield, invoke this stage FIRST to jointly refine the scope.

What happens:

1. Explores the broad field via arXiv search and web search
2. Presents sub-directions to the user
3. Asks clarifying questions (problem scope, target entities/tasks, method families, target venue)
4. Outputs SURVEY_SCOPE.md with refined topic and parameters

Output: SURVEY_SCOPE.md — refined survey specification

🚦 Gate 0 — Scope confirmation: After Stage 0, present the refined scope. If AUTO_PROCEED=true, auto-continue. Otherwise ask for confirmation:

📋 Refined Survey Scope:

Topic: Robust graph learning under distribution shift
Keywords: graph neural network, robustness, distribution shift, benchmark
Problem scope: method + evaluation protocol
Venue: TPAMI

[Auto] Proceeding to literature search... (AUTO_PROCEED=true)

Pipeline Stages

Stage 1: Research Literature (/research-lit OR batch-triage)

Logic (runs automatically when skill is invoked):

1. Check for existing arxiv_results.json:

   • Check ./arxiv_results.json
   • Check ./tpami_tem/arxiv_results.json
   • Check ../tpami_tem/arxiv_results.json (parent dir)

2. If arxiv_results.json exists AND USE_EXISTING_ARXIV_JSON=true:

   • Run batch-triage via CLI:

     python3 tools/surveymind_run.py --stage batch-triage --arxiv-json <path>
     

   • This processes ALL papers in arxiv_results.json (up to 170 papers)
   • Output: corpus_report.json, corpus_report.md with tier classification

3. If arxiv_results.json NOT found:

   • Run fresh search: /research-lit "$ARGUMENTS — arxiv download: true"
   • This searches arXiv API and creates paper_list.json

What happens (batch-triage mode):

1. Reads arxiv_results.json (all papers)
2. Calls arXiv API for metadata/enrichment for each paper
3. Applies 12-field classification
4. Outputs tiered corpus report

Output (batch-triage mode):

• corpus_report.json — Machine-readable corpus with tier 1-4 classification
• corpus_report.md — Human-readable tier summary

Output (fresh search mode):

• paper_list.json — Machine-readable paper list with paper_id, title, authors, year, venue, arXiv ID, and pdf_path
• Saved PDFs in papers/ or literature/

🚦 Gate 1 — Confirmation: After Stage 1, present paper count. If AUTO_PROCEED=true, auto-continue. Otherwise ask for confirmation:

Batch-triage mode:

📚 Using existing arxiv_results.json: N papers
├── Tier 1 (core): X papers
├── Tier 2 (high relevance): X papers
├── Tier 3 (related): X papers
└── Tier 4 (peripheral): X papers

[Auto] Proceeding to paper analysis... (AUTO_PROCEED=true)

Fresh search mode:

📚 Found N papers on "$ARGUMENTS"

Top papers:
1. [Title] - [Authors] ([Year])
2. [Title] - [Authors] ([Year])
...

[Auto] Proceeding to paper analysis... (AUTO_PROCEED=true)

When auto-proceeding, also append to findings.md:

- [Gate 1] research-lit complete: N papers for "$ARGUMENTS" (source: batch-triage/fresh-search)

Stage 2: Paper Analysis (/paper-analysis)

Command: /paper-analysis "$ARGUMENTS"

What happens:

1. Reads paper_list.json from Stage 1 output to get paper_id and pdf_path
2. Loads papers from the paths specified in paper_list.json
3. Applies 8-dimensional classification to each paper
4. Generates structured analysis files

Data Flow:

Stage 1 output (paper_list.json) ──▶ Stage 2 input
         │
         └── paper_id, pdf_path, title, authors, venue

Output:

• Directory: surveys/survey_<topic_slug>/gate2_paper_analysis/
• Files: {paper_id}_analysis.md for each paper

🚦 Gate 2 — Review: After Stage 2, report analysis completion. If AUTO_PROCEED=true, auto-continue:

✅ Paper analysis complete: N papers analyzed
📁 Results saved to: gate2_paper_analysis/

Classification summary:
- Model family A: N
- Model family B: N
- Method family A: N
- Method family B: N
...

[Auto] Proceeding to taxonomy building... (AUTO_PROCEED=true)

When auto-proceeding, also append to findings.md:

- [Gate 2] paper-analysis complete: N papers classified across X categories

Stage 3: Taxonomy Building (/taxonomy-build)

Command: /taxonomy-build "$ARGUMENTS"

What happens:

1. Reads all gate2_paper_analysis/*.md files
2. Builds hierarchical taxonomy
3. Analyzes coverage and interconnections

Output:

• taxonomy.md — Hierarchical classification structure

🚦 Gate 3 — Review: After Stage 3, present taxonomy summary. If AUTO_PROCEED=true, auto-continue:

📊 Taxonomy built successfully

Hierarchy:
├── Representation Enhancement (N papers)
│   ├── Learnable Scaling (N)
│   └── Rotation Transform (N)
├── Sparsity Exploitation (N papers)
│   └── ...
...

Coverage:
- Method Categories: N
- Submethods: N
- Most common: [Method] (N papers)

[Auto] Proceeding to gap identification... (AUTO_PROCEED=true)

When auto-proceeding, also append to findings.md:

- [Gate 3] taxonomy built: X categories, Y submethods identified

Stage 4: Gap Identification (/gap-identify)

Command: /gap-identify "$ARGUMENTS"

What happens:

1. Reads taxonomy.md
2. Identifies 5 types of research gaps
3. Prioritizes gaps by severity and confidence

Output:

• gap_analysis.md — Research gaps and opportunities

🚦 Gate 4 — Review: After Stage 4, present gap summary. If AUTO_PROCEED=true, auto-continue:

🔍 Gap analysis complete

Gap Summary:
| Gap Type | Count | Top Priority |
|----------|-------|--------------|
| Unexplored Combinations | N | Yes |
| Benchmark Gaps | N | |
...

Top Research Opportunities:
1. [Opportunity 1] (Impact: High, Difficulty: Med)
2. [Opportunity 2] (Impact: Med, Difficulty: High)

[Auto] Proceeding to survey writing... (AUTO_PROCEED=true)

When auto-proceeding, also append to findings.md:

- [Gate 4] gap analysis complete: Z gaps identified, top opportunity: [description]

Stage 5: Survey Writing (/survey-write)

Command: /survey-write "$ARGUMENTS"

What happens:

1. Reads taxonomy.md and gap_analysis.md
2. Synthesizes into standard academic survey structure
3. Generates comprehensive survey document

Output:

• SURVEY_DRAFT.md — Complete survey document

Final Report

After Stage 5, present completion summary:

🎉 Survey pipeline complete!

📄 Output files:
├── SURVEY_SCOPE.md (refined survey specification)
├── paper_analysis_results/ (N analysis files)
├── taxonomy.md
├── gap_analysis.md
└── SURVEY_DRAFT.md

Survey Statistics:
- Papers analyzed: N
- Method categories: N
- Research gaps identified: N
- Survey sections: 6

Next steps:
- Review SURVEY_DRAFT.md
- Add missing references
- Refine gap prioritization
- Customize for target venue

Key Rules

• Fuzzy topic detection: If $ARGUMENTS is too broad or vague (e.g., "survey on AI", "写综述", "optimization"), invoke /survey-brainstorm FIRST to refine the scope before proceeding to Stage 1.
• Sequential execution: Each stage depends on the previous
• Auto-proceed by default: When AUTO_PROCEED=true, automatically continue through gates (default behavior)
• Graceful degradation: If a stage finds few papers, continue with warning
• Evidence binding: All claims must be traceable to original papers
• Machine-readable outputs: All intermediate files must be parseable by downstream stages
• Findings tracking: At each gate, append a one-line summary to findings.md for session recovery

Error Handling

• No papers found: Report error, suggest broadening topic
• Insufficient papers for taxonomy: Warn but continue (MIN_PAPERS_FOR_TAXONOMY = 5)
• Stage failure: Report error with specific issue, offer to retry

Integration with Existing Skills

The pipeline chains these skills: 0. /survey-brainstorm — Topic refinement & scope definition (Stage 0, pre-survey)

1. /research-lit — Paper discovery
2. /paper-analysis — Paper analysis
3. /taxonomy-build — Taxonomy construction
4. /gap-identify — Gap identification (NEW)
5. /survey-write — Survey generation (NEW)

Typical Use Cases

/survey-pipeline "graph neural network robustness"
/survey-pipeline "multimodal reasoning evaluation"
/survey-pipeline "privacy-preserving federated optimization"

Files Generated

paper_analysis_results/
├── 2401.12345_analysis.md
├── 2402.23456_analysis.md
└── ...
taxonomy.md
gap_analysis.md
SURVEY_DRAFT.md

This skill orchestrates the complete survey construction workflow.