Reading Research Papers — Building the Habit
Getting into the practice of reading papers consistently — finding them, organising them with Zotero, working through them with the three-pass method, and making it stick.
Why Bother
Most of what gets written about a field — blog posts, tutorials, YouTube explainers — is someone else’s digestion of someone else’s digestion of the actual work. Reading papers is reading the primary source. You stop playing a game of telephone and start reading what the people who built the thing actually found, what they were uncertain about, what they deliberately left out of the abstract.
The payoff isn’t immediate. The first paper in a new area takes hours and yields modest understanding. The tenth takes half the time and yields twice the insight. It compounds — and the compounding only starts once you build the habit.
Finding Papers
Google Scholar is the front door. Search by topic, author, or title. The “Cited by” number tells you roughly how significant a paper has been; the “Cited by” link shows you who built on it — which is often more useful than the paper itself. The “Related articles” link surfaces work in the same neighbourhood. Scholar also tracks your search history and can send email alerts when new papers match a saved query.
arXiv is where most ML, physics, mathematics, and computer science papers land before (and often instead of) journal publication. Papers go up within days of being written — the preprint is usually indistinguishable from the final version. Browse by category (cs.LG for machine learning, cs.AI, stat.ML, etc.) or search directly. The arXiv identifier (2301.07041, for example) is the standard citation handle in these communities.
Semantic Scholar is underrated. Better at surfacing influential papers you haven’t heard of, and its “TLDR” summaries are useful for the first-pass triage.
Connected Papers takes a seed paper and draws a graph of related work based on citation overlap. Useful when you’ve found one relevant paper and want to map the surrounding territory quickly.
Following people is often better than searching. When you find a paper you like, look at the authors’ Google Scholar or arXiv profiles. Follow the researchers who keep producing work you find interesting. Twitter/X still has an active ML research community that posts preprints the day they go up.
Organising With Zotero
Finding papers is easy. Having them scattered across browser tabs, PDFs on a desktop, and half-remembered titles is how you lose them.
Zotero is a free, open-source reference manager that solves this. The browser extension (“Zotero Connector”) adds a one-click save: when you’re on an arXiv page, a Google Scholar result, or a journal page, clicking the extension saves the full metadata — title, authors, abstract, year, DOI — and attaches the PDF automatically. No manual entry.
Inside Zotero:
- Collections are folders. Organise by topic, project, or reading status (To Read / Reading / Done).
- Tags cut across collections — tag a paper
foundationalorrevisitregardless of which collection it’s in. - Notes attach to individual papers. This is where the third pass lives — write your synthesis, questions, and connections here, not in a separate document that will drift away from the paper.
- Search is full-text across all saved PDFs. When you half-remember a paper, this finds it.
The Zotero iOS and Android apps sync your library so you can read on a tablet. The desktop app has a built-in PDF reader with annotation support — highlights and notes sync back to the item record.
One discipline worth keeping: when you save a paper, write one sentence in the abstract field about why you saved it. Six months later, “I saved this because it mapped how organisations actually measure ROI on agentic AI, not just hype” is more useful than a bare title.
The Three-Pass Method
S. Keshav’s framework for reading a paper efficiently. The insight is that a paper should be read multiple times, each time with a different question — not once, front to back, hoping understanding accumulates.
First pass — reconnaissance (5–10 minutes): Read the title, abstract, introduction, section headers, conclusion, and scan the references. Nothing else. The single question: does this paper deserve more of my time? What is the claim? What method? Is this the right paper for what I’m trying to understand, or am I looking for a different one? Most papers end here. That is the correct outcome.
Second pass — comprehension (up to an hour): Read the whole paper without getting stuck. When you hit a proof you can’t follow or a technique you don’t recognise, mark it and keep moving. Goal: be able to summarise the paper’s argument in a few sentences. Figures and charts deserve real attention — they usually contain more information than the text summarising them. A result that looks dramatic in the text but modest in the figure is worth noting.
Third pass — reconstruction (one to several hours): Try to re-derive the paper yourself. For each step in the argument, could you have made that move? Where did you need to look something up? Where did you accept a claim without knowing why it was true? The gaps in your reconstruction are exactly the gaps in your understanding. This is also where you interrogate the methodology — not just whether you can follow it, but whether it actually answers the question the authors claim it does.
Not every paper warrants the third pass. Save it for the ones you want to actually understand deeply, not just be aware of.
Making It a Habit
The failure mode is saving fifty papers in Zotero and reading zero. A library is not a reading practice.
A few things that help:
Fix a time, not a goal. Twenty minutes with one paper, three times a week, beats “I’ll read papers when I have time.” You will never have time.
Triage ruthlessly at the first pass. Most papers you encounter don’t need to be read — they need to be filed or discarded. A well-organised Zotero with fifty read papers is more valuable than a cluttered one with five hundred saved-but-unread.
Start with survey papers and review articles. Before diving into primary research, a good survey gives you the vocabulary, the major debates, and a map of who the significant contributors are. Reading primary papers before you have this map means reading without context.
Follow the citations. The references section of a paper you liked is a curated reading list from someone who has already done the work of identifying the relevant literature. The papers cited most frequently across several papers you’ve read are likely foundational — worth the third pass.
Write something after the second pass. Even three sentences: what the paper claims, how they tested it, what you think about it. Writing forces a clarity that passive reading doesn’t. These notes live in Zotero attached to the paper. Three months later, when you half-remember this paper, the note is what you’ll actually want.
The habit builds its own momentum once you’re past the initial friction. The first few papers in a new area feel like slogging through mud. By the tenth you have enough context that reading feels like conversation. That shift is worth the slog.