The first time you open a machine-learning paper, the instinct is to start at the first word and read to the last. It is almost the worst possible strategy. Papers are written to be archived and precise, not to be read linearly by a newcomer — the dense notation and the results buried mid-way will exhaust you before you reach the point. Reading papers well is a distinct skill, separate from understanding the underlying ideas, and it is learnable. The good news is that a simple, widely taught approach — reading in escalating passes rather than one linear sweep — makes even intimidating papers tractable. This is a practical guide to that skill, aimed at the AI literature but general enough for most technical fields.
Why linear reading fails
A research paper packs its most important claim — the contribution — into the abstract and introduction, then spends the bulk of its pages on methodology, proofs, and experiments that only make sense once you know what the paper is for. Read linearly, you hit the hardest, most detailed material before you have the context to know why it matters or whether you even need it. The fix is to invert the process: extract the big picture first, and descend into detail only where the payoff justifies it.
The three-pass method
The most durable advice on this — from Srinivasan Keshav's widely circulated note "How to Read a Paper" — is to read in three passes, each with a different goal, stopping whenever you have learned what you needed.
- 1
Pass 1 — the bird's-eye view (a few minutes)
Read the title, abstract, and introduction; read the section headings; glance at the figures; read the conclusion. Goal: understand what the paper claims to do and decide whether to read further. Many papers can be triaged out right here.
- 2
Pass 2 — the substance (about an hour)
Read the paper more carefully but skip heavy proofs and derivations. Study the figures, diagrams, and results tables closely — they often convey the method faster than the prose. Goal: grasp the approach and the evidence well enough to summarise it to someone else.
- 3
Pass 3 — the deep read (several hours)
Only for papers you must fully master or build on. Work through the details as if you were re-implementing or re-deriving the work, questioning every assumption. Goal: understand it deeply enough to identify its weaknesses and reproduce its results.
The point of the structure is permission to stop. Most papers deserve only pass one; a few earn pass two; very few need pass three. Beginners burn out by giving every paper a pass-three effort. Reading well is as much about allocating attention as absorbing detail.
Read the figures before the prose
In machine-learning papers especially, the architecture diagram and the key results table often communicate the core idea faster and more clearly than the surrounding text. When you feel lost in a section, jump to its figures — they are frequently the most information-dense part of the paper.
Find the one contribution
The most useful habit on a first read is to force yourself to answer a single question: what is the one thing this paper is claiming to contribute? Nearly every paper has a central contribution — a new method, a new result, a new way of framing a problem — and everything else is scaffolding to support it. If you can state that contribution and the problem it solves in a sentence or two, you have understood the paper at the level most conversations about it operate. Struggling to name the contribution is itself informative: sometimes it means you need another pass, and sometimes it means the paper's claim is thinner than its length suggests.
Read it critically, not credulously
A paper is an argument, not a verdict, and reading it well means reading it skeptically. As you go through the method and experiments, keep a running set of questions:
Checklist
0/4
This critical stance is not cynicism; it is how you turn reading into understanding. The limitations section (and the questions a paper does not answer) is often where the most useful insight lives, both about the work and about the open problems around it.
Where to find papers and build context
Most AI research appears as preprints on arXiv, which is free and open. But a single paper is a snapshot of a conversation: it builds on earlier work and is refined by later work. Two habits build the surrounding context. Reading backward through a paper's references shows you the foundation it stands on; noticing what later work cites and revises shows you how its claims held up. Over time this is how a field's structure becomes legible — you stop seeing isolated papers and start seeing the lineage of ideas connecting them.
Practical takeaway
Stop reading papers front to back. Take a first pass for the big picture from the abstract, introduction, figures, and conclusion, and decide from there whether to go deeper. On a second pass, chase the method and the evidence, leaning on the figures. Reserve the full, line-by-line deep read for the few papers you truly need to master. Throughout, hunt for the single core contribution, read the claims and experiments critically, and use references to place the work in its lineage. Reading research is a skill that compounds: the more papers you read this way, the faster each new one becomes — and the sooner primary sources turn from intimidating into genuinely useful.
Sources & Further Reading
- 01How to Read a Paper — S. Keshav (Stanford / University of Waterloo)The classic short guide that popularised the three-pass method.
- 02arXiv — Cornell UniversityThe open-access preprint server where most AI research first appears.
- 03Attention Is All You Need — Vaswani et al., 2017A well-known paper to practise the three-pass method on.
Editorial note — A conceptual, methodological guide to reading research papers. The three-pass method is attributed to the cited Keshav note; no specific paper's results, statistics, or findings are quoted beyond citing the example paper's existence.
