"Before you can add to a field, you need a map of it. Good researchers are great at finding what's already known — fast."
Level: Intermediate · Time: ~2 days · Prerequisites: Module 2
Learning Objectives
By the end of this module, you will be able to:
- Find relevant papers efficiently
- Build a "citation map" of a topic
- Tell foundational work from follow-ups
- Stay current without drowning in new papers
1. Where Papers Live
Most AI research is free and open. The key sources:
- arXiv — where papers are posted (often before peer review).
- Papers with Code — papers paired with runnable implementations.
- Google Scholar / Semantic Scholar — search and citation tracking.
- OpenReview — papers plus their public reviews.
A quick note on search technique: the exact keywords matter enormously. The same idea might be called "attention," "self-attention," or "transformer" depending on the year and author, so a good literature search means trying several phrasings and following the vocabulary the field actually uses. When one search term surfaces a strong paper, mine its wording for better terms to search next.
Concept: A paper on arXiv may have several versions (v1, v2…). Note which one you read — results and claims can change between versions. A v1 sometimes overstates a result that a later v3, after peer review, quietly walks back.
Explain like I'm new: Think of arXiv as a giant public bulletin board where scientists pin up work-in-progress, and journals/conferences as the slower "official" stamp of approval that comes later. In fast-moving AI, most people read the bulletin board first — but you always check which pinned version you're looking at.
2. Snowball Sampling
Start with one strong "seed" paper and expand outward:
- Backward (its references) → the foundations it builds on.
- Forward (papers that cite it) → what came after and improved it.
A few rounds of this "snowballing" reveals the key works in a subfield — usually a small cluster does most of the heavy lifting.
Here's the pattern in practice: you start with one 2024 paper on, say, image generation. Its references point you back to a handful of 2020–2021 foundations that keep reappearing. The papers citing it show you the 2025 follow-ups. After two rounds, the same 8–10 names show up again and again — those are the field's backbone. You've mapped a topic in an afternoon instead of reading blindly for weeks.
Explain like I'm new: It's like exploring a family tree. From one person you find parents (earlier work) and children (later work). After a couple of generations, you can see the whole family — the important papers everyone descends from.
3. Building a Citation Map
Tools like Connected Papers or Semantic Scholar draw a graph of how papers relate. The hubs — papers everyone cites — are usually the foundational ideas worth reading first. The visual layout does real work: papers that cite each other cluster together, so the picture literally shows you the field's neighborhoods before you've read a word.
| In your map, look for… | Because… |
|---|---|
| Highly-cited hubs | They're the field's foundations |
| Recent survey papers | They summarize a whole area for you |
| Clusters of papers | They mark distinct sub-approaches |
Try this: Pick any topic that interests you (say, "diffusion models"), find one seed paper, and list two papers it cites and two that cite it. You've just started a literature map.
Key idea: A recent, well-written survey is often the single best entry point to a new area — it has already done the mapping for you, grouped the sub-approaches, and cited the hubs. Reading one good survey can save weeks of blind searching. Look for a survey published in the last year or two, then snowball from its reference list.
4. Staying Current (Without Burning Out)
Hundreds of papers appear daily — you can't read them all, and you shouldn't try. Instead:
- Follow a few survey papers to get oriented.
- Skim titles/abstracts; deep-read only the ~20% that matter to you.
- Let citations surface what's important — good work gets cited.
Real-world use case: Before starting a project, a researcher spends two days doing exactly this — one survey, snowball to the hubs, skim 30 abstracts, deep-read 5. That small investment prevents the most painful mistake in research: spending a month "inventing" something that was already published (and improved on) three years ago.
Common mistake: Trying to read everything and feeling perpetually behind. Experts read selectively. A great survey plus a handful of hub papers teaches you more than 100 random skims.
✅ Checkpoint
- What's the difference between backward and forward citations?
- What is a "hub" paper, and why read it first?
- What's a healthier strategy than reading every new paper?
Answers: 1) Backward = papers it references (foundations); forward = papers that cite it (follow-ups). 2) A heavily-cited paper — usually a foundational idea, so it gives the most context. 3) Read selectively: use surveys and citation counts to focus on the ~20% that matter to you.
Key Takeaway: Map a field before adding to it. Papers live on arXiv, Papers with Code, and Semantic Scholar; grow your knowledge by snowballing backward and forward from a seed paper, and identify hub papers and surveys as your starting points. Stay current by reading selectively, letting citations point you to what matters.
Further Learning
- Connected Papers — visual citation maps
- arXiv Sanity — filter the firehose
Part of "Research & Advanced AI." Original content for this learning platform.