How to Read a Research Paper — The Three-Pass Method That Actually Works.

You open a paper. It's 14 pages. Dense equations. Seven figures you don't understand. References to four other papers you've never read. You start from the abstract, read linearly, get lost by page three, and close the tab. Sound familiar?

Most people read papers wrong. They treat a research paper like a novel — start at the beginning, push through to the end, and hope that understanding arrives somewhere along the way. It doesn't. Research papers aren't written to be read linearly. They're written to be navigated.

In 2007, S. Keshav at the University of Waterloo published a short, brilliant guide called "How to Read a Paper". It introduced the three-pass method — a systematic approach that researchers worldwide now treat as gospel. This post distills that method, extends it with practical advice from Simon Peyton Jones and Timothy Roscoe, and adds the workflow I actually use when reading papers for this blog.

Part 1 of 2. This post covers reading. Part 2 covers writing.

Why you need a method at all.

Researchers spend hundreds of hours per year reading papers. To keep up with a field, to do literature surveys, to review conference submissions. Without a method, you either spend too long on papers that don't matter, or too little time on papers that do.

A method gives you three things:

An exit ramp. Not every paper deserves your full attention. A method lets you decide at each stage whether to go deeper or stop.
Layered understanding. Each pass builds on the last. You don't need to understand everything on the first read — you just need to know what the paper is about.
Efficient allocation. You spend 5 minutes on most papers, 30 minutes on relevant ones, and 4–5 hours only on papers that truly matter to your work.

The three-pass method.

The core idea: you read the paper three times, each with increasing depth. Each pass has a specific goal, a specific scope, and a specific exit condition.

Pass 1 — The bird's-eye view (5–10 minutes).

The first pass gives you a general idea of the paper. You're not trying to understand the details. You're trying to answer one question: should I keep reading?

Here's what you read:

The title, abstract, and introduction. Carefully.
The section headings. Read every one, but nothing underneath.
The conclusion. In full.
The references. Scan them. Mentally tick off any you've already read.

Skip everything else — all the body text, proofs, equations, figures, and experiments. You'll get to them later if the paper warrants it.

After this pass, you should be able to answer five questions — Keshav calls these the "five Cs":

Category. What type of paper is this? An empirical evaluation? A new system? A theoretical proof? A survey?
Context. Which other papers is it related to? What theoretical bases does it use?
Correctness. Do the assumptions appear valid? Does the logic feel sound from what you've seen?
Contributions. What are the main claims? What does this paper add to the field?
Clarity. Is this well-written? Is the structure logical?

Using these answers, you make a decision: stop, file for later, or continue to Pass 2. Most papers stop here. That's fine. That's the point. You've invested 5 minutes and you know whether this paper matters to you.

Fig 1 — The three-pass pipeline. Each pass has an exit ramp. Most papers never make it past Pass 1.

Pass 2 — Grasp the content (30–60 minutes).

Now you read the paper with more care — but still not at the deepest level. You're reading for content, not detail.

Rules for Pass 2:

Read the figures, diagrams, and tables carefully. These are the paper's backbone. Pay special attention to graphs — are the axes labeled properly? Are error bars shown? Are the results statistically significant? Sloppy figures often signal sloppy work.
Mark unread references. As you encounter citations, mark the ones that seem important. These form your reading list for the related work.
Annotate. Write notes in the margins (physical or digital). Mark things you don't understand. Circle key terms. Underline claims that surprise you.
Ignore proofs. Skip dense mathematical proofs on this pass. Note that they exist, note what they claim to prove, but don't work through them.

After Pass 2, you should be able to summarize the paper's main thrust to someone else, with supporting evidence. You know the paper's contribution. You understand the approach. You can identify the strengths and weaknesses.

For most purposes — keeping up with the field, writing a related-work section, deciding whether to build on this paper — Pass 2 is enough. You stop here for maybe 80% of papers you choose to read past Pass 1.

Sometimes, though, you finish Pass 2 and you still don't understand the paper. This happens. It might be a new area for you, the math might be unfamiliar, or the paper might just be badly written. You have three options: (a) set it aside and hope you never need to understand it, (b) come back later after reading background material, or (c) continue to Pass 3.

Pass 3 — Virtual re-creation (4–5 hours).

This is the deep dive. The goal of Pass 3 is to virtually re-create the paper — to mentally re-implement the work, making the same assumptions and decisions as the authors. You're not just reading anymore. You're rebuilding.

Here's the key technique: at every step, you ask yourself how would I do this? Then you compare your approach with the authors'. Every difference — every place where their choice surprises you — is a learning opportunity. It reveals either a clever insight you missed or a questionable decision you can challenge.

During Pass 3, you also:

Challenge every assumption. Why did they use this dataset and not another? Why this baseline? Why this metric?
Verify the proofs. Now is the time. Work through the math line by line.
Identify implicit assumptions. What are the authors taking for granted that they never state explicitly?
Think about future work. Not the vague "future work" section at the end — your own ideas for what could come next.
Note missing citations. Are there relevant papers the authors should have cited but didn't?

After Pass 3, you should be able to reconstruct the paper from memory. You know the paper's structure, its strengths, its hidden assumptions, and its weak points. You could explain it at a whiteboard without notes. You could write a review.

This level of understanding is necessary for only a handful of papers — the ones directly related to your own research, the seminal works in your field, or papers you're reviewing for a conference.

Doing a literature survey.

The three-pass method shines when you need to survey an unfamiliar field. Here's how:

Start with Google Scholar. Use well-chosen keywords. Find 3–5 recent, highly-cited papers. Do Pass 1 on each. Read their related-work sections. If you're lucky, one of them will be a survey paper — that's gold.
Find shared citations and repeated author names. These are the key papers and key researchers. Download the key papers, set them aside. Then go to the key researchers' websites and see where they've published recently. That tells you the top conferences in the field.
Go to those conferences' proceedings. Scan recent titles. Find papers related to yours. Combined with the key papers you already have, this is your first pass of the survey.
Do Pass 2 on the set. Now you have context, shared vocabulary, and a mental map of the field.

What makes a bad reading experience.

Keshav's method tells you how to read. But it helps to know what makes papers hard to read in the first place — so you can calibrate your expectations and not blame yourself.

Missing motivation. The authors jump into the solution without explaining why the problem matters. You don't know what question you're trying to answer.
Dense notation without introduction. Symbols appear without definition. Subscripts multiply. You need a notation index that doesn't exist.
Unclear contribution statement. You finish the introduction and you still don't know what's new.
Figures that confuse rather than clarify. Tiny fonts, missing labels, graphs with 12 overlapping lines.
Burying the lede. The important result is on page 8, paragraph 3. The first seven pages are context you already know.

If you run into these, it's not your fault. The paper is badly written. Adjust your expectations and your time budget accordingly.

Reading papers as a reviewer.

Timothy Roscoe's guide on reviewing for systems conferences adds a critical dimension: reading papers with the intent of evaluating them for publication. This is a different skill from reading for understanding.

A good review answers three questions:

What problem does the paper address? In one sentence, without jargon.
Is the problem significant? Does solving it advance the field? Would practitioners care?
Is the solution novel, correct, and well-evaluated? Has it been done before? Does the math check out? Are the experiments convincing?

Common reviewer mistakes:

Reviewing the paper you wish the authors had written instead of the paper they actually wrote.
Focusing on presentation over substance. Bad writing is fixable. Bad science is not.
One-line rejections. "Not novel enough" without explaining what prior work covers the same ground. Unhelpful to authors, lazy from the reviewer.
Forgetting that reviews are for the authors too. Even for rejected papers, the review should help the authors improve their work.

My actual workflow: how I read papers for this blog.

I read 2–4 papers per week for the Paper Juice series on this blog. Here's my real process, evolved over dozens of papers:

1. The triage (Pass 1 on steroids).

I keep a running list of papers from arXiv, Twitter/X, and conference proceedings. When I sit down to pick the next paper to cover, I do rapid Pass 1 on 5–10 candidates. I'm looking for:

A clear, surprising core idea. Something I can explain in one sentence that makes someone say "wait, really?"
Visual potential. Can I draw diagrams? Are there architectural innovations I can illustrate?
Practical relevance. Would a working engineer or ML practitioner benefit from understanding this?

2. The deep read (Pass 2 + selective Pass 3).

Once I've picked a paper, I do a full Pass 2 with aggressive annotation. I use a split-screen setup: paper on the left, notes on the right. I write down:

The one-sentence core idea
The problem they're solving and why it matters
The approach, decomposed into steps
Key equations (only the ones I'll need to explain)
Surprises — things that contradicted my expectations
Weaknesses — things the paper doesn't address or gets wrong

For the math-heavy sections, I do a selective Pass 3 — but only on the equations that are central to the paper's contribution. I work through them by hand. If I can't derive the result, I can't explain it.

3. The teach-back test.

Before I write a single word, I try to explain the paper to an imaginary reader who knows the basics of ML but hasn't read this specific paper. If I can't do it fluently, I haven't understood the paper well enough. I go back and re-read the parts I'm fuzzy on.

This is the most important step. Writing is just the teach-back test in slow motion.

The tools that make it easier.

A few tools I actually use:

Semantic Scholar — for finding related papers and citation graphs. Better than Google Scholar for exploring connections.
Connected Papers — generates a visual graph of related papers. Fantastic for discovering work you didn't know existed.
ar5iv — renders arXiv papers as HTML instead of PDF. Better readability, better search, works on mobile.
Zotero — reference manager. I tag papers by topic, reading status, and whether I've covered them on the blog.
Excalidraw — for sketching diagrams while reading. If I can draw the architecture, I understand it.

The meta-skill: knowing when to stop.

The most important skill in paper reading isn't comprehension. It's triage. The ability to quickly determine whether a paper deserves your time is worth more than the ability to deeply understand any single paper.

Most papers in your arXiv feed don't matter for your work. Many are incremental. Some are wrong. A few are transformative. The three-pass method is a filter that separates them efficiently.

Read widely at Pass 1. Read selectively at Pass 2. Read deeply at Pass 3. That's the whole system.

Cheat sheet: the three-pass method.

Pass	Time	What you read	What you skip	Goal
1	5–10 min	Title, abstract, intro, headings, conclusion, references	Everything else	Decide: read more or stop?
2	30–60 min	Full paper, figures, tables	Dense proofs, implementation details	Summarize to someone else
3	4–5 hours	Everything, line by line	Nothing	Reconstruct from memory

References.

S. Keshav, "How to Read a Paper," ACM SIGCOMM CCR, vol. 37, no. 3, July 2007. PDF
S. Peyton Jones, "Research Skills," Microsoft Research. Link
T. Roscoe, "Writing Reviews for Systems Conferences," ETH Zurich. PDF
H. Schulzrinne, "Writing Technical Articles," Columbia University. Link

Next up: Part 2 — How to Write a Research Paper.