State of the art

Upper bounds: what has been achieved. Lower bounds: what is shown not to work or to cost something.

Upper bounds, per sub-question

Sub-question Best result Method Measured how
Overview first graphic organizer g = 0.70 memory, 0.53 comprehension (Ponce 2025; publication bias, so possibly inflated) meta-analysis memory and comprehension tests
Context timing topic before the passage helps, after does not (Bransford and Johnson 1972) randomised experiment comprehension and recall
Goal-set depth prequestions g = 0.54 on prequestioned items, 0.04 on the rest (St. Hilaire 2023) meta-analysis memory tests
Order within a unit example then problem beats problems only (van Gog 2011) randomised experiment similar and transfer problems
Order of lessons alternating concepts and procedures beats concepts first (Rittle-Johnson 2015) review and experiments procedures and transfer
Order across units whole 4C/ID programmes d = 0.79 (Costa 2022, sequencing not isolated); interleaving g = 0.42, reversed for words (Brunmair 2019) meta-analyses performance tests (Costa); transfer and retention (Brunmair)
Hypertext order most-related-next helps low-knowledge readers (Salmerón 2005, 2006) randomised experiments comprehension after hypertext reading
Proofs self-explanation training d = 0.95 (Hodds 2014) randomised experiments proof comprehension tests
Where to stop a few correct recalls, then returns diminish (Rawson and Dunlosky 2011) randomised experiments delayed recall

Lower bounds and costs

Limit Source Measured how
Organizer effect small on immediate tests (d = 0.21, Luiten; delayed 0.19 to 0.38), mean 0.66 (median 0.48) on delayed tests (Stone, read); unfamiliar-material condition supported by Mayer's vote count, not by Stone's effect sizes (tiny k) Luiten 1980; Stone 1983; Mayer 1979 meta-analyses; theory with experiments
Skimming gives only moderate comprehension Rayner 2016 review
Goal-bounded depth lowers learning of the non-goal part Klauer 1984; McCrudden 2005 meta-analysis; experiment
Withholding assistance beat giving it for advanced learners (d = -0.43, CI -0.65 to -0.21, high heterogeneity); the gain for novices is larger (+0.51) Tetzlaff 2025 meta-analysis
Parts trained apart transfer badly; fixed difficulty steps fail Wickens 2013 meta-analysis
Felt understanding overstates real understanding Rozenblit and Keil 2002 experiments
Overconfident learners stop too early Dunlosky and Rawson 2012 experiments

Holder of the record

  • No study holds a record for breadth-first against depth-first reading of the same hierarchical material: none was found.
  • The best-supported components are separate: overview first, goal-set depth, self-test as the stop.

Most adopted

A record of practice, kept apart from the research record above.

Sub-question Most adopted Best evidence
Conclusion first BLUF, Pyramid Principle, executive summary no controlled test of BLUF found
Importance order in news inverted pyramid, the newsroom default Nildam 1998 favours narrative for Thai news readers, Sternadori 2008 is mixed; both theses
Conclusion first for bad news taught as an exception explanation first rated slightly more readable in one experiment, with no comprehension difference (Jansen 2010)
Web scanning front-load the key facts (Nielsen 2006) descriptive eyetracking, n = 232; no learning outcome
Labelled structure structured abstracts more information, no recall gain at equal length (Hartley 2004)