Read Research Papers Properly
Be able to pick up a research paper in my field, understand what it claims and whether to believe it, and place it in the wider literature.
This is a starting point — make it yours
Use this goal to build your own roadmap — tailored to you and starting fresh.
The method
Reading a paper front-to-back like a textbook is why most people give up. This phase replaces that with a technique.
Learn the three-pass method and use it on one paper~3h
Keshav's three-pass approach: a 5-minute skim (title, abstract, headings, conclusions), a 1-hour read ignoring proofs, then a full read only if the paper deserves it.
The point is that most papers should be abandoned after pass one, and that's a success, not a failure.
Done when: you have run all three passes on one paper and can state, in a sentence, what it claims and what it assumes.
Learn the anatomy of a paper and what each section is really for~2h
Abstract sells, introduction frames, related work positions, method is the actual contribution, results are the evidence, discussion is where authors are most optimistic. Limitations are usually the most informative paragraph in the paper.
Done when: you can name what each section is for, and you read limitations sections first out of habit.
Set up a paper-finding and paper-keeping workflow~3h
Pick one reference manager (Zotero is free and good enough) and one search habit. Learn to use Google Scholar's "Cited by" to walk forwards in time from a landmark paper, and the reference list to walk backwards.
Done when: you have 10 papers in a library with tags, and you can find everything that cited a given paper in under a minute.
Learn to tell a landmark from a footnote~2h
Citation count, venue, whether later work builds on it or merely mentions it, and whether the results have been replicated. Learn the main venues in your own field — a paper at the top conference is not automatically right, but it has survived scrutiny an arXiv preprint has not.
Done when: you can name the three most important venues in your field and explain why one paper on your list matters more than another.
Reading for real
Method without mileage is useless. This phase is deliberately repetitive — fluency comes from volume, not from more technique.
Read a survey paper in your field and map it~5h
Surveys are the cheapest way into a field: someone has already done the reading and organised it. Pick a recent survey and draw the map it implies — the sub-areas, who works on what, which problems are considered solved.
Done when: you have a one-page diagram of your field's sub-areas with two or three key papers named under each.
Read three papers from the survey's reference list~8h
Choose ones the survey called foundational. Three-pass each. Expect the first to be hard and the third to be noticeably easier — the vocabulary starts repeating.
Done when: all three are read and annotated, and you can say how each relates to the other two.
Write a one-page summary of each, in your own words~4h
No copied sentences. Problem, approach, evidence, and the one thing you would ask the authors. If you cannot write it without looking back at the paper, you have not understood it yet — that is the test doing its job.
Done when: three summaries exist and each fits on one page.
Find a paper that disagrees with one you've read~5h
Real fields contain arguments. Use "Cited by" and look for papers that critique, fail to replicate or propose an alternative. Read both and work out what the disagreement actually turns on — it is usually an assumption, not a fact.
Done when: you can state the crux of the disagreement in two sentences and say which side you find more convincing, and why.
Reading critically
Understanding a paper and believing it are different skills. This phase is the second one.
Learn the standard failure modes of empirical claims~4h
p-hacking, selective reporting, tiny samples, unrepresentative baselines, test-set leakage, benchmarks chosen after the fact, and results that only hold on the authors' own data.
Done when: you can name six ways a result can be technically true and practically meaningless, with an example of each.
Interrogate the evidence in a paper you already like~4h
Go back to your favourite paper from phase two and attack it. Is the baseline fair? Would the result survive a different dataset? What is conspicuously not reported?
Attacking work you like is the exercise — anyone can criticise a paper they already doubt.
Done when: you have written three specific, concrete objections, not vague scepticism.
Reproduce one result, or establish why you can't~8h
Take the smallest reported result and try to get it yourself — rerun the code, recompute the statistic, or redo the calculation from the stated numbers. Failing to reproduce something is a genuinely useful outcome and teaches more than success.
Done when: you have either matched the reported number or written down precisely what was missing.
Explain one paper to someone who hasn't read it~4h
A friend, a study group, a blog post, or a written explainer for yourself in six months. Teaching is where you discover the parts you had been skating over.
Done when: you have delivered the explanation and answered at least one question you did not anticipate.
Staying current
The frontier moves. This phase turns reading from a project into a routine you can hold indefinitely.
Set up a filter for new work in your area~2h
arXiv subject alerts, a handful of researchers whose work you rate, and one curated newsletter. Deliberately narrow — the point is a manageable trickle, not a firehose you will start ignoring in a fortnight.
Done when: new relevant papers arrive somewhere you actually look, at a rate under ten a week.
Run a weekly one-hour reading slot for a month~4h
Same time each week. One pass-one skim of everything that arrived, then one paper read properly. An hour a week sustained beats a heroic weekend that never repeats.
Done when: four weeks have happened and you have four papers you can describe.
Write the literature-review paragraph for your topic~5h
The paragraph that would open a report or dissertation: what the field currently believes, who established it, and where the open question is. This is the artefact all the reading was for.
Done when: the paragraph exists, cites at least six papers you have genuinely read, and names one open problem.
Join one place where the papers get discussed~3h
A reading group, a lab meeting, a seminar series, or an online community that goes deeper than headlines. Reading alone plateaus — you keep your own blind spots.
Done when: you have attended twice and said something at least once.