Reading a paper “properly” is often imagined as beginning with the first sentence and continuing through the references without interruption. That sequence offers the satisfaction of completion, but it can spend an hour on background before establishing whether the study answers the reader's question. The opposite shortcut is equally risky: reading only the abstract and conclusion accepts the authors' compressed interpretation before examining the methods and data.
Efficient reading is not maximal skipping or faster eye movement. It is matching depth and order to purpose. A researcher mapping a new field, checking the provenance of one number, deciding whether a paper meets review criteria, learning a method, reproducing an analysis, and conducting peer review need different parts of the same article. Applying maximum scrutiny to every search result leaves less time for decisive papers. Applying abstract-level scrutiny to every paper makes evidential claims unverifiable.
S. Keshav's widely used How to Read a Paper separates three passes(1). A first pass triages relevance and basic character. A second establishes the content and evidence structure. A third attempts to reconstruct the work and challenge its assumptions. The suggested timings came from a computer scientist's practice rather than a comparative reading experiment. The durable idea is the separation of jobs. Stopping after pass one can be a successful decision; the efficiency lies in reserving pass three for papers that warrant it.
.png)
The first pass decides whether the paper deserves more attention
Before reading the content, confirm that the file represents the right publication stage and remains current. An unpublished preprint, proceedings version, author-accepted file, and final journal record are not necessarily interchangeable; analyses or conclusions can change, and later corrections or retractions are also possible. Crossmark documentation explains how participating publishers register status and important updates visible from HTML or PDF(2). Coverage is incomplete, so the publisher page and DOI record still matter. The check prevents deep analysis of an obsolete version.
Then state the present reading goal in a form that can be answered. “Learn about the field” is too broad. “Determine whether this study sampled the population in my review,” “verify the denominator behind this effect estimate,” and “decide whether this instrument can be reused” identify stopping conditions. Ten Simple Rules for Reading a Scientific Paper distinguishes orientation, fact retrieval, method learning, review, and replication and explicitly rejects one best sequence for all of them(3).
A rapid pass can inspect the title, abstract, opening context, section headings, figures and tables, conclusions, and references. Record the research question, motivation, study type, principal result, and direct relevance. Keshav frames a similar first pass around category, context, apparent correctness, contribution, and clarity. The purpose is not to endorse the result. It is to decide whether the paper belongs in the reader's next allocation of attention.
Study-type recognition is an early gate. The BMJ guide Getting Your Bearings asks whether an item is primary or secondary research and whether its question concerns treatment, diagnosis, screening, prognosis, or causation(4). Different questions require different designs. A compelling hypothesis or discussion cannot compensate for a design unable to answer the proposed question. Although the article reflects older clinical hierarchies, design-question fit remains a sound triage principle.
Experienced researchers commonly depart from printed order. A survey about how science and health researchers read IMRAD papers collected 139 complete reading-order responses(5). Of them, 98.6 percent began with the abstract, but most did not then move sequentially through introduction, methods, results, and discussion. Twenty-eight percent of abstract-first respondents sometimes stopped when it showed the paper was irrelevant. Methods came second for 21.2 percent, often to assess trustworthiness. The convenience sample was overwhelmingly European, university-based, and experienced, and the response rate was unknown. It documents reported practice, not an optimal method.
The abstract serves different functions in triage and deep reading. Duke's guide for nonscientists, How to Read and Understand a Scientific Paper, uses title and abstract to decide relevance but suggests revisiting the abstract after independently examining a paper during a deep read(6). The aim is to reduce anchoring, the tendency for an initial interpretation to shape later judgment. This is an expert anti-bias tactic, not a universally tested rule. The practical instruction is to treat the abstract's conclusion as a claim to verify, not as the paper itself.
Decide in advance what justifies stopping. A clinical perspective on reading scientific research papers recommends moving from the abstract conclusion to methods and results while asking whether a valid reason not to continue has emerged(7). A mismatched population, absent outcome, unusable design, duplicate version, or unavailable necessary data can justify setting the paper aside. Disagreement with the desired conclusion cannot. Stopping reasons should be recorded against criteria so confirmation bias does not masquerade as efficiency.
Novices also need to distinguish a background-knowledge gap from paper irrelevance. A review of disciplinary literacies in STEM argues that scientific reading varies across fields and develops with prior knowledge, strategic processing, motivation, and self-efficacy(8). Less-experienced readers commonly find the Methods and Results sections difficult and rely more on abstracts and discussions; experienced researchers give those sections greater weight. A short review article or concept lookup may supply the background needed to evaluate an important paper that initially feels inaccessible.
The second pass builds an independent account of the evidence
Once a paper passes triage, read until its question, approach, result, and limitations can be explained without repeating the abstract. A perspective on active versus passive reading recommends asking what question the authors address, why it matters, what approach the reader would choose, and what evidence would support the claim(9). The reader predicts the expected result of each experiment or analysis, compares that prediction with the figures, and forms a provisional conclusion before accepting the discussion.
Figures and tables are compressed primary results, not decorations. Inspect axes, ranges, units, colors, symbols, sample sizes, error bars, legends, comparison groups, and statistical annotations. Determine whether an error bar represents a standard deviation, standard error, or confidence interval. Check whether the prose highlights only one comparison while another pattern is visible. In a small instructional study called Figure Facts, students recorded the question, technique, result, and conclusion for each figure panel in their own words(10). Fourteen students spent more time on figures after instruction, and novel-data interpretation improved between early and middle measurements. The uncontrolled sample is too small for a universal causal claim; the worksheet remains a useful active-reading device.
Methods should be read in relation to each result. Identify who or what was included and excluded, how comparison groups and assignment were formed, how variables were measured, whether measures were reliable and valid, how missing data and attrition were handled, and whether analyses fit the question and data. An introductory guide on the art of reading a journal article uses title, abstract, and conclusion for relevance, then directs attention to sample, eligibility, measurement, controls, reliability, validity, and statistics(11). It also treats the discussion as the authors' interpretation, to be compared with the data and alternatives.
A guide to critical appraisal of scientific articles adds selection, response, loss to follow-up, controls, confounding, representativeness, variation, and precision(12). Confounding arises when a third factor differs with both a suspected cause and the outcome, distorting their observed relationship. A large sample can narrow random uncertainty while preserving systematic selection or measurement error. Funding, conflicts of interest, and the currency and breadth of references also inform interpretation.
.png)
Reading order can now follow the purpose and design. A reader reproducing a method may begin with methods, supplements, protocol, and code. A reader verifying one result may trace its table, analysis population, and model. A systematic review should be navigated through search, selection, appraisal, and synthesis decisions. The PRISMA 2020 explanation directs readers to the review question and eligibility criteria; the search and information sources; how studies were selected, assessed, and extracted; the synthesis, its certainty, and possible reporting bias; plus registration, protocol, and data-availability details(13). Complete PRISMA reporting does not prove sound conduct, but it shows where information needed for appraisal should appear.
For randomized trials, the CONSORT 2025 explanation and elaboration is an information map for allocation, masking, outcomes, analysis populations, participant flow, harms, effect estimates, precision, registration, protocol and statistical-analysis-plan access, data sharing, funding, and conflicts(14). CONSORT is explicitly not a quality instrument and does not prescribe one universally correct trial method. For cohort, case-control, and cross-sectional research, STROBE identifies selection, setting, variables and confounders, bias, study size, missingness, adjusted and unadjusted estimates, precision, limitations, and generalizability(15). Reporting guidance locates evidence; it does not decide whether that evidence is credible.
The output of pass two should be structured notes, not a long plot summary. Separate the question, population or material, design, key comparison, numerical result, limitation, relevance to the current project, and references requiring follow-up. A UC Davis guide to searching and organizing research articles advises reference-manager users to capture why an article matters, its main conclusion, methods worth revisiting, relevant tags or folders, and figures they may need later(16). The note must preserve why the paper mattered and how deeply it was checked, not merely prove that it was opened.
The third pass reconstructs the paper and tests alternatives
A paper that will anchor an argument, supply a reused method, determine a design choice, or undergo peer review deserves deeper work. Try to reproduce the authors' reasoning from the data. Ask whether the same analytic choices lead to the same conclusion, what assumptions are necessary, what changes under plausible alternative assumptions, and which relevant literature is missing. Keshav's third pass resembles virtually redoing the study. Supplements, equations, code, registration, protocol, and data may be required, and several hours can be an appropriate investment.
Use a design-specific appraisal tool rather than one generic questionnaire. CASP offers separate educational checklists across study designs, including trials and reviews, qualitative work, cohort and diagnostic research, case-control designs, economic evaluations, and prediction models(17). The tools move from screening to methodological soundness, result magnitude and precision, and applicability. They preserve “can't tell” and discourage a numerical total. The Oxford Centre for Evidence-Based Medicine's critical-appraisal tools similarly organize reading around a focused question, valid methods, important results, and local applicability(18). Neither replaces statistical or design expertise.
Active deep-reading courses illustrate what reconstruction can involve. CREATE labels five actions: first Consider, then Read. Readers next Elucidate hypotheses before they Analyze and interpret data; finally, they Think of the next experiment. In one CREATE course study, students annotated papers, redrew figures, made concept maps, predicted experiments, and compared their reasoning with author responses(19). Self-assessed reading ability and epistemological beliefs improved, but the evidence came largely from one uncontrolled course. A separate nonrandom comparison found improvement in both an adapted CREATE course and a traditional active literature-discussion course, with no significant between-group advantage(20). Active engagement shared by both may matter more than a branded sequence.
The wider teaching literature is heterogeneous. A review identified 74 approaches to teaching primary scientific literature, differing in duration, modality, target skill, and assessment(21). Studies concentrated in biology and four-year institutions; many lacked assessment, almost all involved one institution, and few measured both cognitive and affective outcomes. No universal evidence-backed reading order emerges. Defining the learning goal, choosing an appropriate paper, scaffolding difficult sections and figures, and evaluating the activity are more defensible principles.
Critical reading is not a search for reasons to reject an author. Generate alternative explanations and compare them fairly with the authors' account. Check whether the study answers its own question before criticizing it for not answering another one. Then assess whether applying the finding to the reader's question introduces indirectness. Also monitor confirmation bias: favored results should not receive lighter scrutiny than inconvenient ones.
Not every paper needs a third pass. Field mapping may rely on triage plus structured notes. Representative theory papers, directly cited results, adopted methods, conflicting evidence, and design-defining studies require deeper passes. Citing a number from an abstract, treating discussion language as a measured result, or reusing a causal conclusion without inspecting methods is not efficient selection. It is omitted verification.
Record reading depth explicitly: triage only, core content checked, methods and results appraised, or reconstruction attempted. Record why a paper was stopped, such as irrelevance, design mismatch, duplicate version, or inaccessible full text. When the research question changes, those notes show whether a previously excluded paper deserves a new pass.
“Do not read every paper from start to finish” does not mean “skip whatever is difficult.” It means choose the purpose, triage before investing, build understanding from methods and data rather than summaries, and reserve reconstruction for the evidence on which decisions depend. Completion is not the measure of adequate reading. The question is whether the source has been read deeply enough to support the claim being made from it.
Sources
1. How to Read a Paper - https://doi.org/10.1145/1273445.1273458
2. Crossmark Documentation - https://www.crossref.org/documentation/crossmark/
3. Ten Simple Rules for Reading a Scientific Paper - https://doi.org/10.1371/journal.pcbi.1008032
4. Getting Your Bearings - https://doi.org/10.1136/bmj.315.7102.243
5. How, and Why, Researchers Read IMRAD Papers - https://doi.org/10.1371/journal.pone.0297034
6. How to Read and Understand a Scientific Paper - https://arc.duke.edu/how-to-read-and-understand-a-scientific-paper-a-guide-for-non-scientists/
7. How to Read a Scientific Research Paper - https://doi.org/10.4187/respcare.09541366
8. Disciplinary Literacies in STEM - https://doi.org/10.1080/23752696.2021.1882326
9. Active versus Passive Reading - https://doi.org/10.1093/nsr/nwaa130
10. Figure Facts - https://doi.org/10.1187/cbe.11-07-0057
11. Art of Reading a Journal Article - https://doi.org/10.4103/0973-029X.110733
12. Critical Appraisal of Scientific Articles - https://doi.org/10.3238/arztebl.2009.0100
13. PRISMA 2020 Explanation and Elaboration - https://doi.org/10.1136/bmj.n160
14. CONSORT 2025 Explanation and Elaboration - https://doi.org/10.1136/bmj-2024-081124
15. STROBE Explanation and Elaboration - https://doi.org/10.1371/journal.pmed.0040297
16. How to Search and Organize Research Articles - https://biotech.ucdavis.edu/news/how-search-and-organize-research-articles
17. CASP Checklists - https://casp-uk.net/casp-tools-checklists/
18. CEBM Critical Appraisal Tools - https://www.cebm.ox.ac.uk/resources/ebm-tools/critical-appraisal-tools
19. The CREATE Approach - https://doi.org/10.1187/cbe.11-03-0027
20. CREATE versus Traditional Primary-Literature Analysis - https://doi.org/10.1128/jmbe.v14i2.506
21. Teaching Students to Read Primary Scientific Literature - https://doi.org/10.1187/cbe.22-10-0211