How to read a research paper efficiently: the three pass method for PhD students

A PhD involves reading a very large number of papers. Reading each one line by line, from the abstract to the last reference, is how many first year students fall behind. Efficient readers read in layers, and they decide early how deep each paper deserves to go.

Why reading linearly fails

Papers are not written to be read in order. The introduction sells, the method section assumes context you may not have yet, and the key result often sits in a single figure. Reading linearly spends the same effort on every paper, when most of them only deserve a few minutes.

The three pass method

This approach was described by S. Keshav in a short, widely shared article, How to Read a Paper. Each pass has a clear goal and a time budget.

Pass 1. The bird's eye view (about five to ten minutes)

Read the title, abstract, and introduction. Skim the section headings. Read the conclusion. Glance at the references and note the ones you already know. At the end, you should be able to answer five questions:

• Category. What kind of paper is it: a new method, an analysis, a review, a system?
• Context. Which other work does it build on?
• Correctness. Do the assumptions look reasonable?
• Contributions. What does it claim to add?
• Clarity. Is it well written?

Most papers stop here, and that is fine.

Pass 2. The content (up to about an hour)

Read with more care, but skip proofs and fine details. Study every figure and table: axes, units, error bars, and what is being compared. Mark references you should read next. At the end, you should be able to summarize the main argument, with evidence, to a colleague.

Pass 3. The reconstruction (several hours)

Reserve this for papers central to your own work. Try to recreate the work in your head: make the same assumptions, then challenge each one. Ask what you would have done differently, and what the authors did not test. This is where your own research ideas come from.

Read like a reviewer

From reviewing papers, these are the questions I ask of every result. They are useful long before you review for a journal.

• Do the claims match the evidence? Is "outperforms" backed by a meaningful margin, or by a small difference on one dataset?
• Are the baselines fair? Same data, same splits, similar tuning effort.
• Is uncertainty reported? Variation across random seeds or folds, confidence intervals, or statistical tests. A single number hides a lot. For uncertainty inside the model itself, see explainable AI and uncertainty in medical imaging.
• Are failure cases shown? A method that never fails in the paper has probably not been tested hard enough.
• Could you reproduce it? Data, code, and enough detail about preprocessing and hyperparameters.

A note template that turns reading into writing

For every paper that reaches pass 2, keep a short structured note. These notes become your literature review almost directly.

Build a reading rhythm

Skim widely and read deeply rarely. A steady weekly routine, with a fixed slot for first passes and a smaller slot for deep reads, beats occasional reading marathons. Store every note in your reference manager so that nothing you read is lost.