State of the art

An upper bound here is the best result achieved, by whom, measured how. A limit is a result showing that a part does not deliver what it is assumed to deliver. A trend is a description of practice, with no comparison of formats.

Sub-question Record holder Method Measure Kind
How widespread is IMRaD in medicine? Sollaci 2004 1,297 articles, 4 journals, 1935-1985 above 80% in the 1970s, sole pattern in the 1980s upper bound
How widespread elsewhere? Lin 2012; Moskovitz 2024 hand-coded corpora, 39 disciplines; STEM by study type plain IMRD not the default; under half in engineering and social science limit
Does a checklist raise completeness? Turner 2012 Cochrane review, 50 non-randomised evaluations, 16,604 trials allocation concealment 45% vs 22% (RR 1.81, 99% CI 1.25 to 2.61) upper bound
Is reporting then complete? Turner 2012 same "remains sub-optimal" in endorsing journals too limit
Does endorsement help for guidelines other than CONSORT? Stevens 2014 systematic review, 101 guidelines, 26 evaluations of nine of them PRISMA summed score only (SMD 0.53, 99% CI 0.02 to 1.03, endorsing against non-endorsing journals); insufficient evidence for the rest limit
Does a structured abstract hold more information? Taddio 1994 300 abstracts before and after the switch, 33 criteria quality 0.57 vs 0.74 upper bound
Same, outside medicine Budgen 2008 randomised paired judgement, 64 judges completeness +6.65 on an 18-item questionnaire, clarity +2.98 on 1-10 upper bound
Does it speed up search? Hartley 1996 simplified abstracts, timed true/false questions faster and more accurate; a second study with a printed database found the same upper bound
Same, with real database abstracts studies 15 and 17 in Hartley 2004 timed search on MEDLINE abstracts no speed advantage limit
Does searching by labelled field help? Booth 1997 simulated database, ten searches structured fields gave higher precision with the same or lower recall in three of five questions per database; one question per database found nothing (p. 165); Table 4 shows higher precision in two of five limit
Is it easier to read? Hartley 1997 eight studies: formulas, sentence reordering, reader ratings formula scores higher only when authors rewrote the abstract; readers rated structured ones easier limit
Is it more accurate? Pitkin 1999; Hartley 2000 abstract data checked against the body 18% to 68% of medical abstracts deficient; no format difference in psychology limit
Does a visual abstract draw attention? Oska 2020 crossover, 40 articles, three tweet forms in random order mean views 1,351 vs 639 (citation) vs 732 (key figure) upper bound
Does it improve understanding? none found none no controlled comprehension study located open
Does results-blind review change what is published? Scheel 2021 71 Registered Reports vs 152 standard reports, first hypothesis 44% vs 96% positive results upper bound
Is the quality judged higher? Soderberg 2021 353 reviewers, 29 Registered Reports vs 57 comparisons, 19 criteria methodological rigour +0.99 on a -4 to +4 scale; novelty no difference upper bound
Does a code-and-text bundle reproduce? Pimentel 2019 re-ran 863,878 Python GitHub notebooks with a valid execution order 24.11% ran without error, 4.03% gave the same results limit
How long does a reading last? Tenopir 2008 repeated surveys since 1977, last-reading incident 48 minutes (1977) to 31 minutes (2005) trend
Where do experienced readers look? Hubbard 2017 survey, 260 undergraduates and researchers in biological sciences at one UK university, section ranking figures and results first, introduction last trend
Is scientific prose getting harder? Plavén-Sigray 2017 709,577 abstracts, FRE and NDC FRE falls with year (r = -0.93) trend
Is there evidence for plain-language summary criteria? Stoll 2022 systematic review, 90 articles, 33 empirical evidence on criteria scarce limit

Most adopted (a practice record)

The two records are defined in the method.

Sub-question Record holder Method Measure Kind
Is a reporting checklist required of every submission in ML? NeurIPS (since 2019), ACL venues (since 2020) venue rules mandatory at submission; NeurIPS 2026 has 16 questions adoption
How many papers come with code in ML? Pineau 2021 2019 NeurIPS submissions about 40% at submission, 74.4% at camera-ready adoption
How many NLP submissions link code? Magnusson 2023 10,405 checklist responses 45.9% adoption
How many models carry a card? Liang 2024b 74,970 Hugging Face repositories 44.2% by one count in the preprint, 90.5% of downloads adoption
Are artifacts common where evaluation exists? Winter 2022 decade of SE and PL conferences about 90% of eligible 2020 PLDI articles adoption
How many publishers give AI guidance? Ganjavi 2024 top 100 publishers and journals, 2023 24% and 87% adoption
How many statements say "on request"? Gabelica 2022 3,416 BMC statements 42% adoption
How transparent are industry AI developers? Bommasani 2024 14 developers, 100 indicators mean 58 of 100 adoption
How many STM publishers have an author policy on AI chatbots? Bhavsar 2025 audit of 162 publishers, 2023-09 to 2023-12 56 (34.6%) adoption
How many health journals instruct on plain-language summaries? Gainey 2022 scoping review of 534 journals 27 adoption
How many ML papers share code and data? Coakley 2026 56,800 papers, five AI venues 11% in 2014, 64% in 2024 (preprint) adoption
How many AI start-ups publish? Loisel 2026 317 firms, 1998-2025 more than half (52.4%) have no qualifying output (preprint) adoption

Best evidence (a research record)

Sub-question Record holder Method Measure Kind
Does an ML checklist change reporting? Magnusson 2023 pre-checklist sample of 50 papers against self-reported answers 8 of 10 items up upper bound (confounded)
Are checklist answers accurate? Pineau 2021 reviewers checked code claims 71.5% confirmed limit
Did documentation improve before the checklists? Coakley 2026 same corpus, trend by year gains predate the checklists; the authors infer an open-science norm limit (preprint)
Does a prompt at a set editorial point raise compliance? Evanko 2025 before and after, one publisher RRIDs in published articles 33% to 65%; code capsules 1.4% to 4.6% upper bound (uncontrolled, owner's data)
Do artifact badges raise citations? Winter 2022 corpus, decade of AE no significant gain limit
Same, with correlation only Heumüller 2020 789 ICSE papers small positive correlation upper bound (observational)
Does a badge raise sharing? Kidwell 2016 before and after with comparison journals under 3% to 39% upper bound
Does a data mandate raise sharing? Hardwicke 2018 interrupted time series 25% to 78% upper bound
Does a statement predict real availability? Danchev 2021 487 trials, declared against actual 68.6% declared, 0.6% public limit
Does sharing on request work? Stodden 2018 204 papers, request and re-run 44% obtained, 26% reproduced limit
Does a datasheet change what an engineer notices? Boyd 2021 think-aloud, 23 participants more ethical issues, earlier upper bound (small)
Does a model card raise use? Liang 2024b difference in differences, 42 models +29.0% in one batch upper bound (one batch)
Does a plain summary beat a text summary? Buljan 2018 three randomised trials no knowledge difference limit
Which summary type raises perceived understanding? Bredbenner 2019 randomised, 538 participants video and plain language ahead upper bound
Is a significance statement simpler? Markowitz 2024 34,584 PNAS papers d near 0.3 on word simplicity trend
How much biomedical writing is LLM-processed? Kobak 2025 excess vocabulary, 15 million abstracts at least 13.5% (2024) lower bound
Same, by field and venue Liang 2025 mixture model, 950,965 papers up to 17.5% in computer science upper bound (estimate)
Does STAR Methods improve reporting? none found none no study located open
Do release blog posts or technical reports change reading or citing? none found none no study located open
Does authors' declared AI use match estimated use? none found none no study located open
Do reviewers follow an AI rule? Kim 2026 randomised policy and survey, N = 1,486 reviewers and chairs 22.5% under a ban said they used an LLM (self-report, preprint) limit

Reading the table

  • Every record concerns an abstract, a checklist, a review format or a bundle; none compares a whole article written in two shapes.
  • Any claim that IMRaD helps readers of full articles is therefore graded convention, untested (position.md).
  • The trends describe readers and prose; they do not say which shape serves those readers best.
  • The upper bounds for compliance come from a venue or publisher measuring itself (Evanko 2025, ICML 2026b); an outside measurement of the same is rare.