Résumé de lecture, chapitre par chapitre — rapporte ce que le livre dit.
THIS DOCUMENT IS NOT A CITATION SOURCE — a reading summary, ungraded.
C'est un résumé de lecture, non coté, qui rapporte ce que le livre dit. Il ne porte aucun poids probant et ne fonde aucune affirmation. Toute citation reste ancrée sur le .txt du livre, jamais sur ce résumé.
Efficiency in Learning - Evidence-Based Guidelines to Manage Cognitive Load - Clark, Ruth Colvin; Nguyen, Frank; Sweller, John (2006) (EN)| Title | Efficiency in Learning: Evidence-Based Guidelines to Manage Cognitive Load |
| Authors | Ruth Colvin Clark · Frank Nguyen · John Sweller |
| Publisher | Pfeiffer (Wiley), 2006 · ISBN-13 978-0-7879-7728-3 · LCCN 2005010921 |
| Corpus source | library/24_Sciences_apprentissage/Efficiency in Learning - … (2006) (EN).txt |
| Volume | 103 025 words |
| Structure | Introduction · 13 chapters in 5 parts · Appendix All About the Numbers · Glossary · ~120–130 references · Index present · List of figures and tables · CD-ROM contents |
| Table of contents | ✅ READ in the file, not reconstructed |
Source qualification, before reading. 247 ‰ function words — lower than narrative prose (≈290 ‰) because the book is dense with tables, figure captions and citations; ten deciles 240–264, last at 195 (references and index). 0 machine-translation banners. Chapter titles located in the body at strictly increasing positions — complete book.
⚠️ Three extraction traps in this file, found before quoting and worth recording:
efficiency, difficult, specific carry a single U+FB01 character in headings, part openers and chapter openings, but are plain fi in body text. A search must try both.±, = and ≤ signs are lost in extraction — 7 ± 2 reads 7 \x01 2, and the book's central formula E = P – ML reads E P – ML. No quotation here contains those signs.c01.qxd 9/29/05 17:53 Page 14), and running heads appear letter-spaced (T h e Ps y c h o l o g y), so they never match a plain search.Method. Three readers in parallel over the five parts, each asked for the evidence — how many studies, which by name, effect sizes where stated, and the boundary conditions. All quotations re-searched against the .txt by exact match, as written into this file.
Three additive kinds of load:
« Intrinsic load is the mental work imposed by the complexity of the content in your lessons and is primarily determined by your instructional goals. » [V]
« Extraneous load imposes mental work that is irrelevant to the learning goal and consequently wastes limited mental resources. » [V]
« Germane cognitive load is mental work imposed by instructional activities that benefit the instructional goal. » [V]
The instructional imperative: minimise extraneous, maximise germane, manage intrinsic by segmenting and sequencing.
Element interactivity is what makes intrinsic load high or low:
« Element interactivity simply means that several knowledge elements must be coordinated in memory to accomplish the task. » [V]
Schemas are the currency, and expertise is schema size, not slot count:
« Schemas are memory structures that permit us to treat a large number of information elements as though they are a single element. » [V]
« The limits of working memory capacity were first made explicit in George Miller's 1956 paper: The Magical Number Seven Plus or Minus Two. » [V]
Their revision of Miller: the number of chunks is fixed for everyone, chunk size is not. A chess board is ~24 elements to a novice and eight or nine schemas to a master (Chase & Simon, 1973). Instruction for novices supplies schema substitutes; experts do not need them.
Automaticity is the third route around the bottleneck, with a caution practitioners rarely repeat:
« Once automated, the skill can be performed with little or no resources from working memory. » [V]
— but the path there requires hundreds of repetitions, and the book states plainly that most organisational training does not allocate the time.
Efficiency, as a measured quantity — the title rests on this:
« Cognitive load theory defines efficiency in terms of two variables: learner performance and learner mental effort. » [V]
« Conceptually, the efficiency metric is calculated by subtracting mental load (ML) from performance (P) outcomes. » [V]
Performance is usually an end-of-lesson test; mental load is usually a subjective learner rating:
« The difficulty (mental load) of a lesson is assessed using a 1 to 7 or 1 to 9 scale » [V]
« Although learner estimates of mental load are subjective, studies that have compared these ratings with other physiological or psychological measures of mental load show that they are effective » [V]
PART ONE — ch. 1–2. Cognitive Load and Efficiency in Learning frames the problem economically (U.S. organisational training at 50–60 billion dollars a year, into the $300 billion range with salary time) and states the theory's five properties. The Psychology of Efficiency supplies the memory architecture and the dual visual/auditory channels.
PART TWO — ch. 3–8, managing irrelevant load. This is the evidence-dense core.
| Ch. | Effect established | What the evidence actually is |
|---|---|---|
| 3 | Modality effect, dual coding | Mayer's programme: six experiments, median 79 % more solutions from text-plus-diagram, median effect size 1.37; animation set, four of four, median gain 80 %, median ES 1.17. ⚠️ Bounded: audio helps only when neither diagram nor words suffice alone, content is complex, learners are novices, and nothing must be re-consulted (audio is transient). |
| 4 | Split attention, signalling | Mautone & Mayer (2001), three experiments, learning 48 % to 44 % better, ES .69 / .74 / .74. Moreno & Mayer (1999): modality ES 1.06 vs. split-attention ES .48 — the two effects are not the same size. |
| 5 | Redundancy, weeding, seductive details | Chandler & Sweller (1991): diagram alone better in one-half to one-third of the time. Seductive-details replication ES .55; redundant on-screen text ES .53; audio-then-text sequencing ES 1.18 under instructional pacing — and no redundancy effect at all under learner pacing. |
| 6 | (no new effect) | ⚠️ The weakest chapter, and the authors say so. No new experiments; every study recycled from ch. 3–5; the guidelines are stated as predictions. |
| 7 | Segmenting, pretraining, pacing | Mayer, Mathias & Wetzell (2002), three experiments, ES .91 / 1.54 / 2.16 — and Experiment 3 has the right control: pretraining after the lesson gives the same exposure and does not help. |
| 8 | Worked example effect, completion, backwards fading | Sweller & Cooper (1985): training time 32.0 s vs. 185.5 s. ⚠️ Backwards fading, the basis of Guideline 19, has the weakest numbers in the book — ES .23 near transfer, .27 far — and the authors report them as low. |
PART THREE — ch. 9, spending the freed capacity. Varied-context worked examples for transfer; self-explanation (Chi 1989: 15.5 vs. 2.75 explanations per example between high and low scorers; Atkinson 2003 prompts, ES .42 / .37); automaticity (Shiffrin & Schneider, over 2 000 sessions); mental rehearsal — which works only after a basic schema exists.
PART FOUR — ch. 10–11, the pivot. See §4.
PART FIVE — ch. 12–13. Ch. 12 integrates the guidelines into a design sequence and criticises authoring tools for violating the redundancy principle by default. Ch. 13 is Sweller's retrospective — see §5.
« Many of the instructional methods that are effective for novices either have no effect or, in some cases, depress the learning of learners with more expertise. This outcome is called the expertise reversal effect. » [V]
« expertise reversal predicts that a given instructional method that works well for novice learners is not only not useful for individuals with more expertise, but also results in depressed learning outcomes! » [V]
Prior knowledge is the only individual difference the book accepts as interacting reliably with method — and learning styles are dismissed as "unproductive instructional mythology."
Reversals appear in five distinct places across Part Two, each with its own study: audio explanation of diagrams (Kalyuga et al. 2000 — the version that gave the best initial learning lost its effectiveness over time); signals becoming redundant; procedure segmenting yielding nothing with advanced learners; worked examples depressing learning at later stages (Kalyuga et al. 2001); and the mixed-audience instruction that tells you to deliberately break an earlier guideline:
« we recommend violating the split-attention guideline and placing explanatory text underneath the diagram » [V]
« In a second experiment that was similar in all respects except that the learners were more advanced, there were no differences in outcomes. » [V]
« the version that led to best initial learning lost its effectiveness over time » [V]
« As learners gain expertise, worked examples can actually depress learning because it requires more mental effort for an experienced learner to study a worked example than to simply work a problem herself. » [V]
⚠️ And the honesty extends to declining to generalise a result they like:
« We need more research on text coherence before we recommend writing low coherent texts for more experienced learners. » [V]
Ch. 11 (rapid testing) reports a validated instrument — a twelve-item rapid test in 118 seconds against 574, correlating .92 with the traditional test, ES 2.23 — and then issues no guideline at all, on the grounds that the research is too new and the domains too structured. A book that measures something well and still declines to recommend it is doing something unusual.
« The origins of cognitive load theory can be found in the results of a failed experiment. » [V]
« Over many years, we realized that a successful experiment almost never leads to either theoretical or practical advances. » [V]
« In our case, advances seemed to come from failed experiments. Successful experiments did little more than confirm what we already knew. » [V]
He concedes, in order: that worked examples failed in geometry and physics, which is what produced the split-attention effect; that the redundancy effect was not theirs — it "had been demonstrated, forgotten, and then demonstrated again on several occasions over many decades"; that the early claim that intrinsic load was immutable was wrong, and that reducing it costs understanding temporarily:
« We had to modify the theory to say that you can reduce intrinsic cognitive load but you cannot simultaneously maintain full understanding. » [V]
Germane load is not his. He attributes it to Paas and van Merriënboer, arising from the variability effect — high-variability worked examples raised load and improved learning:
« because it was a load that was germane to schema acquisition and automation » [V]
And on evolutionary psychology — the part most often quoted from him — he states its limit himself: the ideas explain the theory's direction but do not directly provide instructional procedures.
His methodological position is the sharpest sentence in the book:
« No instructional recommendation should ever be accepted without that recommendation having been tested using controlled experiments » [V]
« Without controlled experimental testing, we face an unending list of instructional fads. » [V]
The efficiency metric. Performance and mental load use incomparable scales, so both are converted to Z scores (mean 0, SD 1). The efficiency value is then (mean performance Z − mean difficulty Z) ÷ √2, the divisor being the geometry of a distance to a diagonal. The efficiency graph plots mental effort Z horizontally against performance Z vertically, with the diagonal marking equality: upper-left is efficient, lower-right is not.
Effect size:
« The effect size tells you how many standard deviations the test group is from the control group. » [V]
« statistical significance does not necessarily translate into practical significance » [V]
⚠️ The book gives two different small-effect thresholds, and does not notice. Chapter 1:
« As a general guideline, effect sizes less than or equal to .30 are considered small and are of negligible practical importance. » [V]
The Appendix, attributing the scale to Hojat & Xu (2004):
« As a general guideline, effect sizes less than or equal to .20 are considered small and are of negligible practical importance. » [V]
Identical sentence, different number. On its own it changes nothing about the guidelines; in a book whose argument is that practitioners should read effect sizes, an unflagged .30 / .20 discrepancy in its own definition of "negligible" is the kind of thing worth catching.
Their caution about generalising is conditional and replication-based, which is the correct form:
« Any one experiment—even one with a high effect size—is likely to have limited applicability to your instructional environment » [V]
« some of the guidelines we offer are more recent and therefore do not yet have a large number of experiments to support them » [V]
The evidence base, as they describe it:
« The guidelines and examples in this book are based on over twenty-five years of valid experimental research conducted by John Sweller and his associates throughout the world. » [V]
« Cognitive load theory is based on dozens of experiments conducted over the past twenty-five years by instructional scientists in Australia, Europe, and the United States. » [V]
« All of the research we summarize uses random assignment of participants to an experimental lesson and a comparison lesson. » [V]
⚠️ That last sentence is the only stated selection criterion in the whole book. There is no count of studies, no database, no search strategy, no inclusion/exclusion protocol — this is a narrative review by the theory's own author, not a systematic one, and the reader should hold it as such. Table 1.1 ("over sixteen studies" on the modality effect, 1995–2003) is the only enumerated body.
Well-supported here: the modality effect (largest replicated base, ES ~1.1–1.4); split attention; redundancy; worked examples; pretraining (ES .91–2.16, with a real control); expertise reversal, which is supported by both cross-sectional and staged designs.
Thin, and flagged by the authors themselves: performance aids (ch. 6 — no dedicated research, guidelines stated as predictions); whole-task and problem-based designs ("almost no controlled experiments"; Hmelo-Silver's review reported honestly, including that PBL students score slightly lower on multiple-choice); audio-then-text sequencing ("very recent"); narrating text with no visual (two studies with opposite results, reconciled post hoc, with an explicit call for more research); backwards fading (ES .23/.27).
Why this book belongs beside the other three. Hanna gives case histories and no numbers. Rosenberg gives no traceable source for his three most-quoted figures and a bibliography that stops in 1994. Porges reports original data with an n and a statistic in four places out of nineteen chapters. Clark, Nguyen and Sweller report effect sizes, name the boundary conditions, mark the guidelines whose evidence is thin, and devote a whole chapter to the conditions under which their own advice reverses.
That is not a claim that the theory is right. It is a claim about what a defensible evidence base looks like when a book has one — and it makes the difference legible, in the corpus, without argument.
Evidence rating is not done here. Assigning an E-level requires applying the framework in the domain module, against the higher-evidence literature first. This file records what the book says and how well it supports it; the grid is applied inkb/24_Sciences_apprentissage.