AI Skill Report Card
Mapping Literature Citation Networks
Quick Start12 / 15
Given a seed paper, expand outward through its citation graph:
- Pull the seed paper's References (what it builds on) and Citations (what builds on it)
- Rank related papers by a combination of: shared references/citations overlap, recency, and citation count
- Present results as a scannable list:
Author, Year — Title — Venue — #References — #Citations - Let the user drill into any result to re-center the map on that paper
Example query result:
Comanici, 2025 — Gemini 2.5: Pushing the Frontier with Advanced Reasoning...
arXiv.org | 0 References | 5k Citations
Recommendation▾
Clarify the actual data source/tool for pulling citation graphs (Semantic Scholar API, Google Scholar, etc.) — currently assumes access without specifying how to obtain it
Workflow13 / 15
Progress:
- Step 1: Identify seed paper(s) or topic keywords
- Step 2: Pull direct references (backward citations) and direct citations (forward citations)
- Step 3: Compute co-citation / bibliographic coupling to surface "More Like This" papers
- Step 4: Filter by date, keyword, journal, or venue as needed
- Step 5: Sort/present by relevance, recency, or citation count
- Step 6: Flag high-impact nodes (very high citation counts) as likely foundational works
- Step 7: Flag recent low-citation papers as emerging/frontier work worth watching
Step 1 — Seed selection Start from one anchor paper the user cares about, or a small cluster of 2-3 papers defining a topic.
Step 2 — Expand the graph
- Backward: references cited BY the seed (its intellectual ancestry)
- Forward: papers that cite the seed (its influence/descendants)
Step 3 — Relatedness scoring Two papers are "More Like This" if they:
- Share a high proportion of references (bibliographic coupling), or
- Are frequently cited together by later papers (co-citation) Prioritize these over simple keyword match.
Step 4 — Filtering Apply constraints the user specifies: date range, specific journal/venue, keyword in title/abstract, minimum citation count.
Step 5-7 — Synthesis Distinguish three paper types in the results:
- Foundational: old, extremely high citation count (e.g., Lin 2014 COCO — 65.9k citations)
- Active/central: moderate age, steadily accumulating citations, frequently in "More Like This" lists across multiple seeds
- Frontier: published in last 1-2 years, low citation count so far, but topically tight to the seed cluster
Recommendation▾
Add a concrete bad-output example (e.g., an overly keyword-matched or citation-count-only result) to contrast with good outcomes, since only positive examples are shown
Examples14 / 20
Example 1: Input: Seed paper = "Recipe1M: A Dataset for Learning Cross-Modal Embeddings for Cooking Recipes and Food Images" (Marín, 2021) Output:
- Foundational refs pulled in: Lin 2014 (COCO), Zhou 2018 (instructional videos)
- Related frontier work: Wang 2022 (recipe generation with structure prediction), Li 2025 (image-to-video transfer learning survey)
- Synthesis note: This paper sits in the cross-modal (vision+language) dataset lineage; cluster it with TextCaps, HowTo100M, and COCO as dataset-construction papers rather than modeling papers.
Example 2: Input: Topic keyword search = "crochet pattern generation" Output:
- Seitz, 2022 (Digital Crochet: visual language for pattern description) — foundational for this niche
- Li, 2025 (CrochetBench) — frontier, directly builds on Seitz 2022, benchmarks VLMs on the task
- Synthesis note: Small, emerging niche (~2 core papers); connect to broader multimodal reasoning surveys (Wang 2024) for methodological grounding since the niche itself is citation-sparse.
Recommendation▾
Quantify or give formulas for bibliographic coupling / co-citation scoring rather than just naming the concepts, so Claude has a reproducible method
Best Practices
- Always report both reference count and citation count — reference count signals scope/era, citation count signals impact
- When citation counts are very close in magnitude (e.g., two 2024 surveys with 224 vs 225 citations), treat them as duplicates/versions and merge rather than list separately
- Surface "more recently published" and "more citations" as two distinct sort modes — don't conflate impact with recency
- When a topic is thin (few directly related papers), explicitly widen to adjacent surveys to give the user grounding
- Note the venue/type (arXiv preprint vs peer-reviewed journal vs conference) since it affects reliability weighting
Common Pitfalls
- Don't treat raw citation count as a proxy for quality without checking age — a 2025 paper with 4 citations may be more relevant than a 2014 paper with 65k if the user wants current methods
- Don't ignore duplicate entries (same title/authors appearing twice with different citation snapshots) — dedupe before presenting
- Don't recommend only high-citation "foundational" papers when the user is looking for frontier/emerging work — always segment by recency
- Don't rely solely on keyword matching in titles; two papers can be highly related via citation network overlap despite dissimilar titles (e.g., cooking recipe generation and video instructional datasets)