Review: learning science

This page lists each source on what makes notes and documents learned, with what it found. Its eight strands run from surveys of the field to practice.

Grouped by strand. Surveys first, then the seminal work, then later syntheses.

1. Surveys and maps of the field

  • Dunlosky 2013 rates ten techniques: practice testing and distributed practice high; elaborative interrogation, self-explanation, interleaving moderate; summarising, highlighting, rereading low (R).
  • Donoghue 2021 re-pools ten techniques (MA): spacing 0.85, testing 0.74, elaborative interrogation 0.56, self-explanation 0.54, rereading 0.47, interleaving 0.47, highlighting 0.44, summarising 0.44.
  • In the same pool, surface outcomes average 0.60 and deep outcomes 0.26.
  • Noetel 2022 pools 1,189 multimedia studies (MMA): contiguity and signalling are largest; design matters more for complex material.

2. Retrieval practice and spacing

  • Roediger 2006: restudy wins at five minutes, testing wins at two days and one week.
  • Rowland 2014 (MA): testing vs restudy g = 0.50; 0.73 with feedback, 0.39 without (secondary).
  • Yang 2021 (MA): g = 0.50 against all controls over 222 classroom studies; g = 0.33 against restudy.
  • Adesope 2017 (MA): about 0.6 against all controls (secondary).
  • Latimier 2021 (MA): spaced vs massed retrieval g = 0.74; expanding vs uniform gaps g = 0.03, not significant.
  • Cepeda 2008 (S, n above 1,350): best gap is 20-40% of the time left at a one-week horizon, 5-10% at one year.
  • Rawson 2013: successive relearning to criterion raised course exam scores on key terms (RCT in courses).
  • Brunmair 2019 (MA): interleaving g = 0.42; maths 0.34; heterogeneous.
  • Bisra 2018 (MA): self-explanation prompts g = 0.55; 0.41 in the five studies that equalised time on task.
  • St. Hilaire 2023 (MA): prequestions g = 0.54 on questioned material, 0.04 on the rest; Pan 2023 (R) reviews the same effect.
  • Hamaker 1986 (R): questions placed after a passage help questioned and related content.

3. Cognitive load, worked examples, subgoals

  • Sweller 2019 (R): the CLT framework and its effects.
  • Barbieri 2023 (MA): worked examples beat problem solving in maths, g = 0.48, about 18 percentile points.
  • Correct examples beat incorrect ones; added self-explanation prompts lowered the gain in that pool.
  • Renkl 2004 (RCT): removing solution steps one by one beats a jump to whole problems.
  • Kalyuga 2003 (R) and Tetzlaff 2025 (MA, 60 studies, 5924 participants): novices gain from more assistance (d = +0.51, CI 0.26 to 0.75), advanced learners from less (d = -0.43, CI -0.65 to -0.21). Heterogeneity is high (I2 near 90%), and giving assistance to novices helps more than withholding it from experts.
  • Tetzlaff 2025 adds that most studies are STEM (38 of 60) and only 10 are humanities and language, where the evidence is least clear. Its limits include dichotomised knowledge and few studies per moderator level.
  • Pollock 2002 (RCT): showing complex material as isolated elements first, then together, helps novices.
  • Paas 1994 (RCT): varied worked examples give better transfer with less effort.
  • Costa 2022 (MA, 12 studies): whole 4C/ID programmes, built on simple-to-complex task classes, d = 0.79, larger in higher education.
  • Catrambone 1998 and Margulieux 2016 (RCT): abstract subgoal labels transfer better; labels in both text and example beat either alone. No pooled estimate found.
  • Alemdag 2025 (MA): erroneous examples g = 0.14, larger with a prompt to explain the error.
  • Yeo 2019 (RCT, 3 experiments): for a new procedure, studying examples beat testing on an immediate test; at one week testing equalled studying with varied problems and beat it with identical ones.
  • Mayer 2002 (RCT, 3 experiments): pre-training on names and traits helped transfer in all three and retention in two.
  • Kester 2006 (RCT): procedural information given just in time beats up-front.
  • Klauer 1984 (MA): stated objectives raise targeted learning and lower the rest, net slightly positive.

4. Concept maps, notes, organisers

  • Nesbit 2006 (MA, 55 studies, 5,818 participants): against activities without a map, constructing one g = 0.82, studying one g = 0.37; 0.47 for constructing in the North American studies alone.
  • Schroeder 2018 (MA, 142 effects): constructing 0.72, studying 0.43.
  • Karpicke 2011 and Blunt 2014 (RCT): free recall beat elaborative mapping at one week; a map drawn from memory equals free recall.
  • Robinson 1995 (RCT): graphic organisers beat outlines for relations among concepts, not always.
  • Kiewra 1985 (RCT): a skeleton halves what students write and raises recall; best with the skeleton before the lecture and full notes after.
  • Ponce 2022 (MA): own highlighting g = 0.36 memory, 0.20 comprehension (n.s.); instructor-provided highlighting 0.44 on both.
  • Kobayashi 2005 (MA): notes without review, weighted d = 0.22 (unweighted 0.29; 131 samples from 57 studies) against no note-taking, whether listening or reading. Recognition tests gave only 0.18.
  • Mueller 2014 reported a laptop deficit; Urry 2021 and Morehead 2019 did not replicate it; Flanigan 2024 (MA) finds handwriting g = 0.25 on achievement.

5. Multimedia principles (Mayer)

  • Mayer 2020 summarises the principles: multimedia, coherence, signalling, contiguity, segmenting, pre-training, modality.
  • Mayer 2001 (RCT): sixteen learner-paced segments beat one continuous lesson on transfer, not recall.
  • Ginns 2006 (MA): spatial contiguity, d = 0.72 (37 effects); 0.85 pooled with temporal (50 effects). Larger for high element interactivity (d = 0.78, 43 effects) than low (d = 0.28, 7 effects, CI -0.03 to 0.58). All learners were novices.
  • Schneider 2018 (MA, 103 studies): signalling retention g = 0.53 (CI 0.42 to 0.64), transfer 0.33 (CI 0.22 to 0.43), N = 12,201. The cues are organizational signals, colour and labels, not only headings. The paper found no important prior-knowledge moderator, with a descriptively larger effect for high prior knowledge, which contrasts with Richter 2016.
  • Richter 2016 (MA): signalling r = 0.17, strongest for low prior knowledge.
  • Sundararajan 2020 (MA, 58 papers, 68 effects, 7,521 participants): seductive details hurt learning, g = -0.33 (CI -0.48 to -0.18), strongly heterogeneous (I2 = 89%).
  • Höffler 2007 (MA): animation over static pictures d = 0.37; larger for realistic (0.76) and procedural-motor (1.06) content, 0.40 for representational graphics.
  • Berney 2016 (MA): g = 0.23, high heterogeneity.
  • Ploetzner 2020 (R): animation helps when the change shown is what must be learned.
  • Tversky 2002 (R): animations are often too fast or complex to perceive; interactivity can help.
  • Clark 1991 (R): dual coding theory applied to education.

6. Hypertext and navigation load

  • DeStefano and LeFevre 2007 (R, 38 studies): each link is a decision that loads working memory.
  • Novices are helped only by a hierarchy matching the domain; restricting links helps.
  • Amadieu 2009 (S, n = 24): a network concept map did worse than a hierarchical one, most for low-prior-knowledge learners.
  • Doignon 1985 (theory): prerequisites as a quasi-order; the outer fringe is what a learner is ready for.
  • Fang 2019 and Sun 2021 (MA): an adaptive system built on it equals classroom teaching (g = 0.05, Sun); Sun finds 0.43 as a supplement, Fang no such difference.
  • Kulik 1990 (MA): mastery learning about 0.52 on exams (secondary), larger for weaker students, at a cost in time.

7. Comparison and transfer

  • Alfieri 2013 (MA, 57 experiments, 336 tests): case comparison d = 0.50 on 336 tests (95% CI 0.44 to 0.56); d = 0.60 on the 57 experiments, heterogeneous (I2 = 59). The comparison against sequential study of the same cases gives d = 0.37 (CI 0.31 to 0.44); the effect is larger against control (0.69) and nonanalogous cases (0.82). Domain moderates inconsistently, so the maths figure of 0.23 is not a stable effect and is not given.
  • Alfieri 2013, moderators: asking for similarities only beat similarities and differences; perceptual content beat procedural; immediate tests beat next-day tests; learner experience and age were not moderators.
  • Gick 1983 (RCT): one case plus a principle did not transfer; two cases did.
  • Gentner 2003 (RCT): two cases compared beat the same two studied apart.
  • Catrambone 1989 (RCT): comparison must be prompted.
  • Rittle-Johnson 2007 (RCT): solution methods side by side beat sequential on procedures.
  • Schwartz 1998 (RCT): contrasting cases before telling give better transfer from the lecture.
  • Sinha 2021 (MA): problem before instruction g = 0.36, reversed for young learners.
  • Ainsworth 2006 (R): translating between representations is hard and needs support.
  • Fyfe 2014 (R): concreteness fading, with later doubt.

8. Practice

Graded convention, untested unless a study measured the effect. A company page is its owner's claim about itself.

  • Open Spaced Repetition 2026a (tool, community benchmark): 9,999 Anki collections, 349,923,850 evaluated reviews. Log loss without same-day reviews: FSRS-6 0.3460, FSRS v4 0.3726, HLR 0.4694, an LSTM neural model 0.3332. It scores recall prediction, run by the FSRS community.
  • Open Spaced Repetition 2026b (tool docs): the project reports 20-30% fewer reviews than SM-2 at equal retention (below the quality bar: an example, not evidence). The page states the figure is from simulation.
  • Duolingo 2026 (company blog): half-life regression error "nearly half" that of the Leitner method, over 12 million sessions. It reports prediction, retention and activity, no learning gains.
  • Khan Academy 2024 (company): about 350K students; 30+ minutes a week is associated with about 20% greater-than-expected gains, reached by about 9%. The page calls its design non-randomised.
  • Oreopoulos 2026 (academic): a two-year cluster RCT in 18 Tennessee middle schools. Khanmigo assignment gave 0.06 to 0.08 SD per school year, about 0.14 SD per full year of active use.
  • Newton 2020 (R, 37 samples, 15,405 educators): 89.1% weighted belief in matching teaching to learning styles, measured in 34 of the 37 samples. The abstract counts 15,405 educators and the Discussion 15,045.
  • Newton 2017 (S, 114 UK academics): 58% believe, 33% used styles in the past year, 32% (31.6%) would keep using them after seeing the evidence.
  • Zettelkasten, second brain: widely used, and the only trial found is Malashenko 2023 (see against.md).