Scientific Paper Writing

Stage 1: High-Recall Retrieval

• 20-30 keyword queries via search.py -o "site:arxiv.org ..."
• Each taxonomy cell: 3+ query variants (core terms, synonyms, method names)
• Snowball: seed paper citation networks
• Target: 200-500 raw candidates

Stage 2: LQS Multi-Dimensional Scoring

Thresholds: LQS≥7.0 must-cite, 5.0-7.0 conditional, <5.0 drop

Stage 3: Citation Depth Classification

• A-level (1-3 paragraphs): section protagonist, 3-5 per chapter
• B-level (2-5 sentences): important insight, 5-10 per chapter
• C-level (1 sentence): supporting evidence
• D-level: dropped, not cited

Stage 4: Venue Upgrade

• Cross-check DBLP + OpenReview for acceptance status
• arXiv with "Accepted at X" → @inproceedings
• Target: arXiv-only ratio ≤ 60%

Verification

• Every 20 citations: title match, author, year, venue check
• Target: verification rate ≥80%, hallucinated = 0
• Year distribution: within-1yr ≥40%, accepted ≥30%

Chapter Architecture (Survey Standard)

• §1 Introduction: Hook → Gap → Contributions → Roadmap
• §2 Background: formal definitions, taxonomy overview
• §3-6 Core: one method family per chapter, with critical assessment
• §7 Benchmarks + Experiments
• §8 Future: specific open problems (Barrier + Attack vector)
• §9 Conclusion: numbered key findings (not repeat of abstract)

Paragraph Logic Patterns

Taxonomy Design

• Multi-axis matrix (not flat list)
• MECE: mutually exclusive, collectively exhaustive
• Must have empty cells → gap analysis material
• Spanning methods show taxonomy tension (good)

Formal Claims

• Default: Conjecture + Remark (not Theorem)
• Hedge ladder: demonstrates > suggests > may > hypothesize
• Rule: claim strength ≤ evidence strength

Related Work Differentiation

• Mandatory comparison table with existing surveys
• "We're more recent" is NOT sufficient differentiation
• Need structural novelty: new taxonomy, new angle, new experiment

Stage 1: Design (Most Important)

• Must answer: "which paper claim does this support?"
• Experiment spec: hypothesis, independent/dependent vars, control vars, expected results
• Statistical plan decided BEFORE running (no HARKing)
• Principles: falsifiable, minimal first, pre-registered, has control

Stage 2: Execute

• API: 3-5 frontier models × 2-3 conditions × 15-25 tasks × 3 trials
• GPU: via cluster job submission + auto-monitoring loop

Stage 3: Iterate

• Ceiling effect → increase difficulty
• Floor effect → decrease difficulty or check for bugs
• Not significant → increase trials or change hypothesis
• Surprise finding → design follow-up
• Max 5 iterations, then accept best result

Stage 4: Report (Data Only)

• Output: results.json (schema: config + results + statistics + findings)
• Output: experiment_summary.md (purpose, results, limitations)
• Does NOT produce LaTeX tables or figures — that's the Figures skill's job

Table Types

Table Rules

• No vertical lines — booktabs three-line style only
• Alternating row color: \rowcolor{gray!6}
• Bold best results in each column
• All experimental data: mean ± std
• Caption must contain key finding, not just description

Figure Types & Tools

• Data-driven (curves, bars, heatmaps): matplotlib → PDF
• Architecture/flow diagrams: TikZ or SVG→PDF
• Simple schematics: PIL → PNG (acceptable per reviewer feedback)
• Priority: TikZ > matplotlib PDF > SVG→PDF > PIL PNG

Quality Checklist

• Vector format (PDF) preferred, PNG ≥ 300 DPI
• Font size ≥ 10pt after scaling
• Academic palette: blue #2196F3, red #F44336, green #4CAF50, orange #FF9800
• All axes labeled, all lines have legend
• Light grid (alpha=0.3) for readability
• Self-contained: understandable without reading main text

Quantity Targets

• Full survey (50+ pages): ≥10 tables, ≥6 figures
• Short survey (30 pages): ≥5 tables, ≥3 figures