How to Read Academic Papers (Keshav Three-Pass Method)
Do not start at the first sentence and grind to the last. S. Keshav's 2007 note How to Read a Paper treats academic reading as up to three passes; his 17 February 2016 revision is the version he now calls latest. The UNC Writing Center adds that you pick a purpose before you pick a depth. This page is that reading procedure. It is not how to find a scholarly source, not the CRAAP evaluation walkthrough, not how to take citable notes, and not how to write a literature review.
By CitationEasy Editorial Team22 min read
On this page
Quick Answer: Three Passes, Then Stop
Decide why you opened the PDF. Then spend five to ten minutes on a first pass (title, abstract, headings, any math, conclusion, references). If the paper is still worth your time, spend up to an hour on a second pass that privileges figures and ignores proofs. Reserve a third pass — many hours for a beginner — for papers you must review, replicate, or build on. Most assigned readings stop after pass two.
Copy-ready first-pass checklist (Keshav 2016 §2.1)
Read:
- Title, abstract, and introduction
- Section and subsection headings only
- Any mathematical content, just enough to name the theory
- Conclusions
- References you already know, marked mentally
Answer the five Cs:
- Category — measurement, analysis, prototype, review, commentary?
- Context — which papers and theories does it sit with?
- Correctness — do the assumptions look valid?
- Contributions — what is new?
- Clarity — is it well written enough to continue?
Then choose:
Stop, read background first, or continue to pass two. A first pass that ends in "not now" is a successful pass.
Format the bibliographic header while the file is still open — author, year, title, journal, DOI — then go back to reading. The CitationEasy generator can style the string later; it cannot recover a paper you never identified. Capture quotations and paraphrases with the labels in how to take research notes.
Why Academic Reading Is a Procedure
The UNC Learning Center's journal-article handout lists the properties that make a research article unlike a textbook chapter: the writer assumes expert background, the vocabulary is not glossed, the claim lives in the title and abstract, the text is built from other researchers' work, and the topic is often narrow. Highlighting every sentence does not solve those properties. A procedure does.
UNC Writing Center, “Academic Reading Strategies”
Establishing why you are reading something will help you decide how to read it, which saves time and improves comprehension.
Carey, Steiner, and Petri make the same point as Rule 1 of their 2020 PLoS Computational Biology editorial: the goal changes which sections you privilege.
Carey, Steiner, and Petri (2020), Rule 1
What you want to get out of an article should influence your approach to reading it.
A journal-club review, a methods hunt, and a Tuesday-night discussion post are three different reading jobs. The peer-review process explains why a published article can still be wrong; this page explains how far to read before you decide.
Keshav's Three-Pass Method (2007 Published, 2016 Latest)
Srinivasan Keshav published the method as a two-page note in ACM SIGCOMM Computer Communication Review (vol. 37, no. 3, July 2007, pp. 83–84; DOI 10.1145/1273445.1273458). On his Cambridge publications page he now points readers to a 17 February 2016 revision as "the latest version." The three passes and the five Cs are the same. What changed is the first-pass math glance, a note that most reviewers make only one pass (so a graphical abstract helps), and a longer budget for pass three. Harvard's CS 260r reading page still assigns the method because the systems community treats it as the default. The wording below is his.
Keshav (2016), §2, the key idea
The key idea is that you should read the paper in up to three passes, instead of starting at the beginning and plowing your way to the end.
The first pass gives you a general idea about the paper. The second pass lets you grasp the paper’s content, but not its details. The third pass helps you understand the paper in depth.
Pass 1: Five to ten minutes and the five Cs
Keshav's first pass is a bird's-eye scan, not a skim of random paragraphs. Read the title, abstract, and introduction carefully. Read headings and ignore their body text. In the 2016 revision, glance at any mathematics just long enough to name the theoretical foundation. Read the conclusions. Tick the names you already know in the reference list. Then answer Category, Context, Correctness, Contributions, and Clarity.
Keshav (2016), §2.1
The first pass is a quick scan to get a bird’s-eye view of the paper. You can also decide whether you need to do any more passes.
At the end of the first pass, you should be able to answer the five Cs:
Glance at the mathematical content (if any) to determine the underlying theoretical foundations
The UNC Learning Center gives the same entry points in classroom language: start with the abstract or first page and the conclusion or last page, then convert headings into questions. You may stop after this pass because the paper is off-topic, because you lack the background, or because the assumptions look unsound. Keshav is explicit that a typical reviewer — and most later readers — will make only one pass, which is why headings, abstracts, and (in the 2016 note) a single well-chosen graphical abstract have to carry the argument.
UNC Learning Center, “Reading Journal Articles”
Begin with the abstract (or first page) and the conclusion (or final page) for an overview of topics.
Feel free to read sections out of order or to focus only on a few sections.
Worked first pass: Keshav (2007) itself
Here is a first pass on the paper that teaches the method. The bibliographic string is the published CCR note, not a made-up article. Use the 2016 PDF when you want his later first-pass step and the longer third-pass clock.
Bibliographic header (APA 7)
Keshav, S. (2007). How to read a paper. ACM SIGCOMM Computer Communication Review, 37(3), 83–84. https://doi.org/10.1145/1273445.1273458
Five Cs after ten minutes
Category:
A short methods note / introductory survey (ACM subject A.1), not an empirical IMRaD study.
Context:
Keshav points readers to Roscoe on writing systems reviews, Schulzrinne and Whitesides on writing technical papers, and Peyton Jones on research skills. The 2016 related-work list adds Iain H. McLean's downloadable review matrix for experimental psychology. The note assumes conference-paper culture in computer systems.
Correctness:
The assumptions are pedagogical (researchers waste time reading linearly; most readers make one pass). They are not statistical claims. They travel well, but humanities essays will not have the figures he tells you to inspect.
Contributions:
A three-pass reading procedure plus a three-step literature-survey recipe (recent papers → shared citations and author sites → top-venue proceedings). The 2016 pass adds the math glance and a longer third-pass budget.
Clarity:
Two pages, numbered steps, five named questions. Worth a second pass if you need the survey method or the third-pass "virtual re-implementation" test.
That first pass already tells you not to treat Keshav as a methods section you can replicate in a lab, and not to expect IMRaD headings. Carey, Steiner, and Petri's Rule 2 is the same checkpoint in biomedical language: name the article type before you judge the evidence.
Pass 2: Up to an hour, figures before proofs
On the second pass, read with care but skip proofs and other machinery you cannot yet reconstruct. Mark key points in the margin. Follow figures, diagrams, and graphs: labeled axes, error bars, and whether the caption matches the claim. Flag unread references that look foundational. When the hour is up, you should be able to explain the main thrust and its evidence to someone else. That is enough for a paper in an adjacent field. It is not enough for a review you signed, or for a method you will run.
Keshav (2016), §2.2
In the second pass, read the paper with greater care, but ignore details such as proofs.
You should be able to summarize the main thrust of the paper, with supporting evidence, to someone else.
If pass two still leaves the paper opaque, Keshav lists three honest exits: set it aside, return after background reading, or continue to pass three. Unfamiliar acronyms, a technique you have never seen, a badly written argument, and being tired at midnight are all on his list. They are reasons to change the plan, not proof that you are a bad reader.
Pass 3: Virtually re-implement the paper
The third pass is for reviewers and for the few papers that sit under your own project. You try to recreate the work from the authors' assumptions, then compare your reconstruction with theirs. Hidden leaps, missing citations, and weak experiments show up in the gaps. The 2007 CCR note budgeted "about four or five hours for beginners, and about an hour for an experienced reader." The 2016 revision is franker: many hours for beginners, and more than an hour or two even when you are experienced. At the end you should be able to rebuild the structure from memory and name the strong and weak points.
Keshav (2016), §2.3
The key to the third pass is to attempt to virtually re-implement the paper: that is, making the same assumptions as the authors, re-create the work.
This pass can take many hours for beginners and more than an hour or two even for an experienced reader.
Do not schedule a third pass for every PDF in a 30-item bibliography. That is how a literature review dies. Use pass one to drop papers, pass two to understand the ones you will cite, and pass three only when you must defend or extend the work.
IMRAD Tells You Where to Look
Most reports of original research are not essays. The ICMJE Recommendations (updated January 2026) name the four-part skeleton and say why it exists.
ICMJE Recommendations (January 2026), “Preparing a Manuscript,” §A.1
The text of articles reporting original research is usually divided into Introduction, Methods, Results, and Discussion sections. This so-called “IMRAD” structure is not an arbitrary publication format but a reflection of the process of scientific discovery.
ICMJE is equally clear that meta-analyses, case reports, narrative reviews, and editorials may use other shapes. Keshav's note is one of those other shapes. So is Carey, Steiner, and Petri (2020). If you cannot find a Methods heading, do not invent one — name the genre and read for that genre's job. PLOS research articles often print Results before Materials and Methods; the reading job does not change. The January 2026 Methods update also tells you what to look for when a study used AI: tool, version, and prompts.
ICMJE Recommendations (January 2026), Methods
The Methods section should aim to be sufficiently detailed such that others with access to the data would be able to reproduce the results.
Authors who used AI technology to conduct the study should describe its use in the Methods section in sufficient detail to enable replication of the approach, including the tool used, version, and prompts where applicable.
| Section | The author's job (ICMJE / Carey) | Your reading job |
|---|---|---|
| Abstract | ICMJE: context, purpose, basic procedures, main findings, principal conclusions; many readers never go further | Decide relevance. Do not cite an abstract claim you have not checked in the body. |
| Introduction | ICMJE: background, specific purpose or hypothesis; cite only pertinent work | Extract the question. Note which prior papers the authors treat as the conversation. |
| Methods | ICMJE: clarity about how and why; detailed enough that others with the data could reproduce the results; AI tool, version, and prompts if used | Skim on pass two unless you must judge validity or reuse the procedure. |
| Results | Carey Table 2: objective observations; data live in figures and tables | Read the figures and captions before the prose. Check that the text matches the axes. |
| Discussion | Carey Table 2: the authors' interpretation, next steps, and (often incomplete) limitations | Separate finding from spin. Write down the limitation they admit and the one they skip. |
| References | Carey Table 2: the papers and datasets the work required | Keshav's survey move: shared citations and repeated author names are the next downloads. |
Carey, Steiner, and Petri's Rule 3 turns those sections into six questions you can ask of the whole paper and of each figure.
Carey, Steiner, and Petri (2020), Rule 3
When reading, ask yourself: (1) What do the author(s) want to know (motivation)? (2) What did they do (approach/methods)? (3) Why was it done that way (context within the field)? (4) What do the results show (figures and data tables)? (5) How did the author(s) interpret the results (interpretation/discussion)? (6) What should be done next?
Worked first pass: Carey, Steiner, and Petri (2020)
A second real paper, this time an open-access editorial rather than a CCR note. Use it to practice naming the genre before you hunt for a p-value that is not there.
Bibliographic header (APA 7)
Carey, M. A., Steiner, K. L., & Petri, W. A., Jr. (2020). Ten simple rules for reading a scientific paper. PLoS Computational Biology, 16(7), e1008032. https://doi.org/10.1371/journal.pcbi.1008032
Five Cs after ten minutes
Category:
PLOS "Ten simple rules" editorial / mentoring essay. Not a research article and not IMRaD.
Context:
Written from a University of Virginia infectious-disease lab for undergraduates through early-career faculty. Cites Kirsch on expectancy, Karanicolas on blinding, and Fiorella and Mayer on learning by teaching.
Correctness:
Advice, not a trial. The six questions and the section-intent table are checkable against ICMJE. The Lego-wall metaphor is pedagogical, not a finding.
Contributions:
Ten rules, a table that maps reading intention to which sections to privilege, and a table of what each research-article section is for.
Clarity:
Six pages, open access, CC BY. Worth a second pass for Rule 4 (figures) and Rule 6 (published papers are not stone).
Worked first pass: Battaglia et al. (2013), a real IMRaD article
The two papers above teach reading. They are not original-research IMRaD. Jennifer Raff uses Battaglia, Grugni, Perego, and colleagues (2013) as the paper she drew methods diagrams from the same week she wrote her guide. It is open access. The headings we opened are Introduction, Results, Discussion, then Materials and Methods — PLOS print order, not the ICMJE mnemonic.
Bibliographic header (APA 7)
Battaglia, V., Grugni, V., Perego, U. A., Angerhofer, N., Gomez-Palmieri, J. E., Woodward, S. R., Achilli, A., Myres, N., Torroni, A., & Semino, O. (2013). The first peopling of South America: New evidence from Y-chromosome haplogroup Q. PLoS ONE, 8(8), e71390. https://doi.org/10.1371/journal.pone.0071390
Five Cs after ten minutes on the live PDF
Category:
Original population-genetics research article. IMRaD content in PLOS Results-before-Methods order, plus a long "state of the art" introduction.
Context:
Y-chromosome haplogroup Q in Native American men from Mexico to South America, compared with Mongolia and Kamchatka. The introduction stages Acosta (1590), Hrdlička / Clovis, Greenberg's three-wave model, and recent mtDNA coastal-versus-ice-free-corridor work.
Correctness:
More than 400 haplogroup-Q Y chromosomes from a screen of more than 2,000 men of paternal Native American ancestry, plus 27 from Mongolia and Kamchatka; 23 Q-specific SNPs and 37 Y-STRs. Competing interests: two authors were employed by Ancestry.com. That disclosure is a first-pass fact, not a verdict.
Contributions:
Two founding lineages, Q1a3a1a-M3 and Q1a3a1-L54(xM3); a southern Siberian origin and Beringian differentiation; arrival in Mesoamerica, then a rapid Pacific-coast move into the Andes; later local patterns tied to agriculture and the large empires.
Clarity:
Open access, CC BY, phylogeny in Figure 1. Worth a second pass if your question is about Native American Y-chromosome history or if you are practicing Raff's "draw the methods" step. Stop here if you only needed a reading-procedure example.
Figures Before Prose
On an original-research article, the data are the point. Carey, Steiner, and Petri say so directly, and they allow two orders: figures first, or results text and figures together. Either way, you owe each panel an x-axis, a y-axis, a color scheme, a statistical note, and a take-home sentence. Dive back into Methods whenever the caption is thinner than the plot.
Carey, Steiner, and Petri (2020), Rule 4
Scientists write original research papers primarily to present new data that may change or reinforce the collective knowledge of a field. Therefore, the most important parts of this type of scientific paper are the data.
Keshav's second-pass figure test is more specific: are the axes labeled, and do the graphs show error bars so the conclusions look statistically meaningful? A beautiful figure that hides the sample size is a reason to slow down, not a reason to trust the abstract. If the statistics themselves are beyond you, read the authors' plain-language claim in the Results and Discussion, then decide whether you need a textbook chapter before a third pass. You do not have to re-derive every test to use the paper's main finding in a literature review.
Humanities Papers Are Not Broken IMRaD
A literary or historical article often has an introduction that states a thesis, unlabeled body sections that stage evidence, and a conclusion that widens the claim. There may be no Methods heading and no figure. Keshav's five Cs still work if you translate them: Category is the kind of argument (close reading, archival recovery, theoretical intervention); Context is the critical conversation in the notes; Correctness is whether the evidence can bear the thesis; Contributions is what changes if the thesis holds; Clarity is whether you can restate the claim in two sentences.
The Harvard Library Interrogating Texts guide (Susan Gilroy) is the better first-year procedure for that genre: preview, annotate, outline or summarize, watch repetitions, contextualize, and compare the reading with the rest of the syllabus.
Harvard Library, Interrogating Texts, “Annotate”
Annotating puts you actively and immediately in a "dialogue" with an author and the issues and ideas you encounter in a written text.
Throw away your highlighter: Highlighting can seem like an active reading strategy, but it can actually distract from the business of learning and dilute your comprehension.
The UNC Learning Center makes the same highlighter warning for journal articles: wait until a section is finished, then mark only the main idea and the support. A yellow page is not a second pass.
UNC Learning Center, “Reading Journal Articles”
Limit your highlighting and underlining. Many students use this strategy, but it does not require much thinking or help you remember what you read.
When Guides Disagree
The Harvard SEAS CS 260r page assigns Keshav, Michael Mitzenmacher's How to Read a Research Paper, and Jennifer Raff's How to Read and Understand a Scientific Article together, then warns that they do not agree. Raff's own 25 August 2013 guide (updated 2014 and 2018) is the source of the abstract-last rule.
Harvard SEAS CS 260r, “On reading research papers”
These guides are great. They also contradict. (Keshav says to read the abstract first; Raff says to never read the abstract until the end.) There’s no one right way to read a paper.
Jennifer Raff (2013), steps 1 and 10
Begin by reading the introduction, not the abstract.
When I’m choosing papers to read, I decide what’s relevant to my interests based on a combination of the title and abstract. But when I’ve got a collection of papers assembled for deep reading, I always read the abstract last. I do this because abstracts contain a succinct summary of the entire paper, and I’m concerned about inadvertently becoming biased by the authors’ interpretation of the results.
Resolve the contradiction with UNC's purpose test, not with a loyalty oath. Read the abstract first when you are triaging a stack — that is also how Raff chooses what belongs in the stack. Delay it when you are training yourself to reconstruct a results story from figures before you hear the authors' spin. Elisabeth Pain's 21 March 2016 Science Careers compilation (DOI 10.1126/science.caredit.a1600047) is useful for the same reason: working scientists disagree on order, and they say so. Either choice is a method. Reading with no choice is how a two-hour block disappears.
When You Do Not Understand the Paper
The old version of this guide was right on one failure mode and vague on the fix. You will not understand every technical detail of every paper in your field. That is normal. What fails is pretending a first pass was a third pass.
- Build the background Keshav assumes you might lack. The UNC Learning Center says to start with a shorter source, a textbook chapter, a review article, or a paper the target cites repeatedly — then return.
- Read out of order on purpose. Introduction and Discussion often teach the vocabulary the Methods assume. Carey's Rule 8 is the extra mile: look up the term, open the supplement, read a cited paper.
- Stop highlighting and start annotating. Harvard's symbol system (a key idea, a surprise, a question you will bring to class) survives a week later. A highlighter stripe does not.
- Talk before you reread. Carey's Rule 9 and UNC's after-reading list both treat a two-minute explanation to a classmate as a comprehension test. If you cannot say the claim without the PDF, you do not have it yet.
- Keep the research question in view. Interesting asides are how a reading list becomes a second topic. Flag them and return to the question you are actually writing.
Critical Reading Without Performing Cynicism
Carey, Steiner, and Petri's Rule 6 is the shortest honest sentence in the reading literature. Pair it with their Rule 7: the authors are human, and a misnumbered supplement panel is not a fraud signal. On Battaglia et al., that means reading the Ancestry.com employment line and the sample description before you decide what the phylogeny can bear — not discarding an open-access paper because a company appears in the competing-interests note.
Carey, Steiner, and Petri (2020), Rule 6
Published papers are not truths etched in stone. Published papers in high impact journals are not truths etched in stone. Published papers by bigwigs in the field are not truths etched in stone.
Ask whether another explanation fits the same figure. Ask whether you like the paper because it agrees with you. Watch for conclusions that outrun the data, correlation sold as cause, a sample too small for the claim, methods too thin to reproduce, and conflicts of interest that appear only in the fine print. Those checks belong here, while the PDF is open. The later accept / caution / reject decision, once you leave the page, is how to evaluate sources.
From a Stack of PDFs to a Literature Survey
Keshav's §3 is the part most classroom handouts drop. To survey an unfamiliar area, do not download fifty PDFs and start at page one.
- Search Google Scholar or a subject database for three to five recent papers. Do one pass on each. Read their related-work sections. If you find a recent survey, read that first.
- In those bibliographies, mark shared citations and repeated author names. Those are the key papers and the key labs. Check the researchers' recent venues.
- Scan the latest proceedings or issues of those venues. The papers you now hold — plus any key paper everyone cited that you still lack — are the first version of the survey. Make two passes. Iterate when a new shared citation appears.
That is how reading becomes source gathering without turning every PDF into a third pass. When you write from the stack, synthesize instead of summarizing one paper per paragraph; the page for that job is the literature review.
What to Write Down (and What Not to Duplicate)
A reading pass that produces no artifact is a mood. Write a two- or three-sentence summary after pass one, a one-paragraph account of claim plus evidence after pass two, and — only for third-pass papers — the reconstruction notes, hidden assumptions, and future-work ideas Keshav asks for. Record the full citation the first time you touch the file.
Do not turn this page into a second notes manual. Quotation marks, page numbers, and the Q / PARA / SUM / MINE labels live on how to take research notes. File structure lives on how to organize research. The sentence-level choice between restating and quoting lives on paraphrasing vs quoting.
Frequently Asked Questions
- How long should it take to read an academic paper?
- Keshav's 2016 revision still budgets five to ten minutes for pass one and up to an hour for an experienced reader's pass two. Pass three is now "many hours for beginners and more than an hour or two even for an experienced reader" — longer than the 2007 CCR "four or five hours / about an hour" line. UNC and Carey both say the clock follows the goal: a discussion-post paper should not get a reviewer's third pass.
- Should I read every word of a research article?
- No. Keshav's first pass ignores body text under the headings. His second pass ignores proofs. ICMJE's Methods section is for reproducibility, not for every reader's first hour. Read every word only on a third pass, and only for papers you must defend or rebuild.
- How many times should I read a paper?
- Up to three, with a different job each time. Carey's Rule 8 mentions a related three-read habit (unpressured, then understanding, then notes). The number is a ceiling, not a virtue. Most papers in a bibliography deserve one pass; a few deserve two; almost none deserve three.
- What if I do not understand the statistics?
- On pass two, read the figure, the caption, and the sentence in Results that states the finding in words. You can use that finding in a literature review without re-deriving the test. If the paper is central to your method, the third pass is where you open a statistics text or ask a labmate — not where you pretend the p-value was decorative.
- Should I read the abstract first or last?
- Keshav and UNC start with it. Raff delays it on papers she has already chosen for deep reading, so the authors' summary does not overwrite her reading of the figures; she still uses title plus abstract to decide what belongs in that stack. The CS 260r page treats that disagreement as expected. Use abstract-first to triage; delay it when you are practicing figure-first reading on a paper you already know you must finish.
- Is the three-pass method only for computer science?
- Keshav wrote it for systems papers, but the five Cs and the stop-after-one-pass rule travel. Translate Category and the figure test when you move to a humanities essay or a clinical trial. ICMJE IMRAD is the biomedical map; Harvard's six habits are the essay map. Do not force a Methods heading onto a close reading. The Battaglia first pass on this page is the genetics-article version of the same five questions.
- Should I print papers or read on screen?
- Carey’s Rule 8 says the ritual matters more than the device: whatever your version of printing it out is, do it so you can mark questions, circle terms, and draw arrows from figures to the Discussion. UNC notes that printing a digital text often makes annotation easier. Searchability still favors a PDF. Pick one and annotate it.
- How is reading a paper different from taking research notes?
- Reading decides relevance, argument, and depth. Notes store a citation, a locator, and a labeled quotation or paraphrase you can cite later. This page is the first job. The second job is how to take research notes. A highlight with no bibliographic header is neither.
- What did Keshav change in the 2016 revision?
- The five Cs and the three-pass jobs are unchanged. The 17 February 2016 PDF adds a first-pass glance at any mathematics, reminds writers that most reviewers make only one pass (a graphical abstract helps), lengthens the third-pass clock, and points to McLean's review matrix. Use the 2007 CCR article for the published citation; use the 2016 PDF for the procedure he now endorses.
Cite the paper you just decided to keep
A first pass that ends in "I will use this" should end with a complete reference, not a filename. Generate the citation in the style your syllabus names, then return to pass two.