cross-paper-synthesis

name	cross-paper-synthesis
description	Synthesize findings across multiple papers into a coherent narrative, structured comparison table, or temporal evolution. Use after collecting papers via survey or paper-search. Goes beyond summarizing individual papers to produce insights that only emerge when reading across the corpus as a whole.
always	false

Cross-Paper Synthesis

When to Use

• User asks "what's the big picture across all these papers?"
• User wants a related work section that actually synthesizes, not just lists papers
• User asks "how has the field's understanding of X evolved over time?"
• User wants to compare methodological approaches across papers
• User needs a structured comparison table of models/methods/results

Synthesis Modes

Choose the mode that best fits the user's goal:

Mode 1: Narrative Synthesis

Best for: related work sections, overview documents, explaining field evolution

• Identify the "through line" — the core insight that unifies disparate work
• Structure as: what we knew → what changed → what it means
• Write prose with inline citations (Author et al., Year)
• Highlight key turning points (papers that shifted the field)

Mode 2: Structured Comparison Table

Best for: method comparison, benchmark results, dataset overview

• Define dimensions to compare (e.g., model, dataset, metric, result, year, venue)
• Build Markdown table with one row per paper
• Flag missing data with —
• Add a "Key Insight" column summarizing each paper's main contribution

| Paper | Method | Dataset | Metric | Score | Key Insight |
|-------|--------|---------|--------|-------|-------------|
| Author et al. (Year) | ... | ... | ... | ... | ... |

Mode 3: Temporal Evolution

Best for: showing how a field progressed year by year

• Group papers by year or phase (e.g., "Pre-2020 foundations", "2020–2022 scaling era", "2023+ efficiency focus")
• Identify what changed between phases
• Show which papers started each phase

Mode 4: Cross-Domain Synthesis

Best for: connecting insights from different research areas

• Identify conceptual parallels between fields
• Note which techniques transferred successfully
• Flag potential transfers not yet attempted (this surfaces gaps)

Workflow

Step 1: Collect Papers

Use paper-search or read from MEMORY.md to get a paper set (10–50 papers is ideal). For papers needing deep understanding, use paper-fetch or paper-read-pdf for full text.

Step 2: Extract Per-Paper Information

For each paper, extract:

• Core claim/contribution
• Method summary
• Datasets used + metrics reported
• Stated limitations
• Year and venue

Step 3: Synthesize

Apply chosen mode. Look for:

• Consensus points — what most papers agree on
• Contested claims — where papers disagree (hand off to contradiction-detection)
• Evolution — how approaches changed over time
• Patterns — recurring themes, methods, or failure modes

Step 4: Write Output

• Narrative: prose with citations, ~500–1000 words
• Table: Markdown table, one row per paper
• Timeline: year-by-year with key papers anchoring each phase

Step 5: Save

• Save to synthesis_{topic}_{date}.md
• Update MEMORY.md with key synthesis insights

Notes

• Best results with 10–50 papers; >100 may require narrowing the topic first
• Always review AI-generated synthesis before including in manuscripts — factual errors possible
• Combine with contradiction-detection to surface and address conflicting findings
• Combine with evidence-grading to distinguish what's established vs. speculative