Gap Analysis

Strategy 2: Benchmark Gap Analysis

• Search Papers With Code for SOTA results (use web_fetch on paperswithcode.com)
• Find tasks/datasets where performance plateaus or no solution exists
• Identify performance gaps between methods
• Look for datasets/tasks that exist but have few papers (low competition)

Strategy 3: Methodological Monoculture Detection

• Identify if most papers use the same dataset, benchmark, or architectural pattern
• Ask: what happens if we change the dataset? the modality? the evaluation metric?
• Cross-domain transfer: find methods successful in domain A (e.g., NLP) not yet applied to domain B (e.g., genomics)
• Check feasibility of transfer via paper-search

Strategy 4: Temporal Trend Analysis

• Compare paper counts by subtopic over years using paper-search with --year-from
• Find "cooling" areas (fewer recent papers) vs "heating" areas (accelerating interest)
• Identify emerging trends that haven't yet been well-studied
• Note: a "cooling" area may indicate saturation OR an opportunity if other fields have advanced

Strategy 5: Method Combination

• Identify papers using method X and papers using method Y
• Check if anyone combined X+Y — use paper-search
• Assess feasibility, prerequisites, and likely impact

Strategy 6: Novelty Validation

• When user proposes a research direction, verify it is genuinely novel
• Search variations: different keywords, synonyms, related formulations
• Check arXiv recent submissions (last 6–12 months) for near-duplicates
• Report confidence in novelty: High / Medium / Low with supporting evidence

Output Format

For each gap found:

### Gap: [Title]
- Type: empirical | theoretical | applied | methodological
- Score: novelty(1-5) × feasibility(1-5) × impact(1-5) = [composite]
- Evidence: [paper excerpts or findings pointing to this gap]
- Strategy: [which strategy found it]
- Suggested approach: [how to tackle it]
- Required resources: [compute, data, skills needed]
- Related papers: [arXiv IDs]
- Novelty confidence: High / Medium / Low

Organize output as a gap map with sections:

1. Empirical gaps — unstudied conditions, datasets, populations
2. Theoretical gaps — missing formal explanations or proofs
3. Applied gaps — real-world applications not yet explored
4. Methodological blind spots — everyone uses the same approach

Save Results

• Write to gaps_{topic}_{date}.md
• Update MEMORY.md with top 3–5 gaps and their novelty scores

Integration Tips

• Combine with contradiction-detection skill — contradictions that remain unresolved ARE gaps
• Combine with evidence-grading skill — weak evidence areas signal gaps worth filling
• Run survey first to build the corpus, then run this skill on the results