39Cognitive Architecture and Instructional Design: 20 Years Later

John Sweller, Jeroen J. G. van Merriënboer and Fred Paas, Educational Psychology Review 31, pp. 261–292, 2019

Linked to 23 chapters: supports 6 · contradicts 7 · mechanism 19 · same claim 2 · pairing 3

The capacity limits of working memory had been in the literature for forty years when Sweller, van Merriënboer and Paas published Cognitive Architecture and Instructional Design in 1998, and almost no instructional recommendation of the period made use of them. Their paper set out what follows for teaching from two facts: that novel information is handled by a working memory limited in how much it holds and for how long, and that information already in long-term memory returns to working memory in quantity with no such limit. From that pair it derived seven instructional effects, each a comparison between a conventional way of teaching something and a way designed around the limits, and went on to collect more than five thousand citations. Twenty years later the same three authors report what has changed. The architecture now rests on Geary's division of knowledge into what humans evolved to acquire and what a culture decided they should learn. Germane load, the third category, has been redefined so that it no longer imposes a load at all. Nine further effects have been added, four of them of a new kind whose content is the point at which the other effects stop working. And the measurement of load has advanced less than twenty years of effort would suggest.

The architecture of 1998

Miller measured the limits of working memory in 1956 and Peterson and Peterson in 1959, but that those limits apply only to novel information stayed implicit in most treatments. Long-term memory fared worse: associated with rote learning, it was assumed to play little part in learning with understanding. The authors trace their own emphasis on it to De Groot, whose study of chess players, first published in 1946, found that chess skill could be explained entirely by remembered board configurations and the best move for each. Skill turned out to be storage. From this the theory takes the relation that carries everything else: working memory is severely limited with novel material, and those limits effectively vanish when what it handles has come from long-term memory, a point Ericsson and Kintsch made in 1995 under the name long-term working memory. Expertise does not merely add facts; it changes what a person can process. The business of instruction is therefore to get information into long-term memory in a form that respects the limits it must pass on the way in.

That has a consequence the authors press hard: complexity is not a property of the material. They measure it as element interactivity, the number of elements that must be held and related at once. For a reader of this paper the written word characteristics is one element retrieved without effort; for someone learning to read English the same squiggles are many interacting elements and can overwhelm working memory on their own. A measure of complexity that ignores what the learner knows predicts nothing about learning.

Intrinsic load is what element interactivity imposes, and it changes only when what is to be learned changes or the learner does. Extraneous load comes from how the material is presented and what the instruction makes the learner do, and almost every effect in the theory is an effect on it. In 1998 element interactivity was taken to belong to intrinsic load alone; Sweller concluded in 2010 that it determines extraneous load equally, so that a poor method is poor precisely because it adds interacting elements the subject matter did not require.

Germane load has had the largest revision. In 1998 it was a third load, the effort devoted to learning itself, added to the total by substituting for extraneous load. That produced a consequence which failed: if germane load took the room extraneous load vacated, cutting extraneous load would leave the total unchanged, and experiment after experiment found the total falling. The 2019 account keeps the name and drops the load. Germane load now means resources redirected to intrinsic rather than extraneous demands, and only two categories remain to be measured.

Through the prism of evolutionary psychology

The strengthening of the theory's basis comes from Geary's distinction between two categories of knowledge. Biologically primary knowledge is knowledge humans evolved to acquire: listening and speaking, recognising faces, basic social functioning, solving unfamiliar problems, transferring what one knows to a new situation, planning. These are difficult accomplishments, acquired without instruction and largely unconsciously, which is why they cannot be taught. Each such skill evolved in its own epoch and stands apart from the others. Biologically secondary knowledge is knowledge a culture has decided matters, which is almost everything taught in schools, and it forms one system rather than many. The contrast the authors keep returning to is listening against reading: children learn to listen without tuition, while writing was invented thousands of years ago and almost nobody learned to read until schooling was organised to teach it.

From this the paper draws a conclusion aimed at a century of educational enthusiasm. Generic cognitive skills matter more to human functioning than domain-specific ones, which is why educators keep proposing to teach them, and they cannot be taught for the same reason that they matter: evolution already supplies them. Sala and Gobet's 2017 review is the authority here, that domain-specific skills are eminently teachable while "attempts to teach them lead to dead-ends". The test the authors require is a randomised controlled trial with a far transfer measure, far because the whole claim for a generic skill is that it works outside the domain in which it was practised. The 1998 paper was partly a reaction to the teaching of general problem-solving skills; that campaign has passed and others have replaced it with no better result. Doctors are taught the SBAR protocol for reporting situation, background, assessment and recommendation in an emergency precisely because their ability to speak to each other needs no teaching.

The architecture that processes secondary knowledge is built from primary processes, and the paper sets it out as five principles of a natural information processing system that mimics the way evolution handles information. Long-term memory is the information store. Most of what it holds comes from other people, by the borrowing and reorganising principle, which is why explicit instruction is the default rather than a concession. When nothing can be borrowed, novel moves are generated and tested at random, the randomness as genesis principle. The narrow limits of change principle names the restriction on how much novel information working memory can take at once, which keeps random generation from wrecking a store that took a lifetime to build. The environmental organising and linking principle names the traffic the other way: stored knowledge returns to working memory without limit and governs action suited to the environment.

The effects

Seven effects were reported in 1998. The oldest, from Sweller and Levine in 1982, replaces a problem that asks for one quantity, and so invites the working-memory-hungry search called means-ends analysis, with one that asks for as many quantities as can be found. Split attention, reported by Tarmizi and Sweller in 1988, arises when a geometry diagram and its solution statements are each unintelligible alone and must be integrated in the head; integrating them physically on the page removes the cost. Redundancy, from Chandler and Sweller in 1991, is the opposite case: a diagram of the flow of blood through the heart and lungs plus a text saying the same thing was worse than the diagram alone, because the learner must process both to discover that they coincide. The variability effect is the exception: Paas and van Merriënboer found in 1994 that practice tasks differing on every dimension improved transfer, which raises load rather than cutting it and works only where load was low to begin with. That finding forced the distinction between load that is wasted and load that is productive.

The nine effects added since, together with element interactivity promoted from a background condition to an effect in its own right, are of two kinds. Five are simple effects of the familiar sort. Four are compound effects, whose content is the boundary of another effect, and the authors treat their arrival as the sign of a maturing theory: one that states when its recommendations hold is worth more than one that does not.

Expertise reversal has the widest reach. As knowledge grows, what were many interacting elements become one stored element, so material designed for a novice starts to carry information the learner no longer needs; the worked example that helped at the start becomes redundant and then harmful, and practice at solving problems takes its place. Guidance fading is the same finding read across a course, and implies that first-year and third-year students should be taught differently for no reason but what they already know. Transient information bounds the modality effect: Leahy and Sweller found in 2011 that short audio-visual segments beat visual-only presentation, while long ones lose to it, because spoken words vanish and must be held. Element interactivity is the most general of the four, and Chen and colleagues demonstrated it between 2015 and 2017 by obtaining a worked example effect with mathematical problem solving and its reverse with mathematical definitions, where students made to generate the answer learned more.

The collective working memory effect treats a group as one processing system built from several limited working memories, able to divide a task between them at the cost of communicating and coordinating. Kirschner and colleagues found in 2011 that the trade runs one way at high load and the other at low: groups learned more efficiently from demanding tasks, individuals from undemanding ones, where the transaction costs bought nothing. The same authors extended the theory from the lesson to the course through four-component instructional design, in which a programme is built from real-work learning tasks arranged in levels of rising complexity with support fading inside each level, from information about how the domain works given before the task because its element interactivity is high, from procedural instructions delivered just in time because a rule can only be built while the information is active, and from part-task practice for whatever is worth automating.

Measuring cognitive load

The instrument that has carried most of the theory's empirical weight is the subjective rating scale Paas introduced in 1992, on which learners report invested mental effort after each task. Comparisons with physiological measures have found it as valid, as reliable, easier to use and more sensitive to small differences. Its weakness is its strength: being one number it gives the total and cannot separate intrinsic from extraneous load. In practice the scale has drifted. Researchers have relabelled it task difficulty rather than invested mental effort, cut it from nine categories to five or seven, and taken one rating at the end rather than averaging ratings after every task. Van Gog and colleagues in 2012 and Schmeck and colleagues in 2015 found the single later rating always higher than the average taken along the way, a discrepancy the authors suggest may itself be a sign of depletion. Leppink and colleagues built a ten-item questionnaire in 2013 and 2014 with several indicators for each type of load, and concluded that intrinsic and extraneous can be told apart by it. The search for objective measures has produced secondary-task techniques, such as the rhythmic foot-tapping method of Park and Brünken in 2015, which impose load themselves and do not distinguish the types, and physiological measures, chiefly electroencephalography and eye-tracking: van Gerven and colleagues found in 2004 that pupil dilation tracked load in young adults but not in older ones. Twenty years after the problem was posed, the position of the review is that the simple rating scale still does most of the work.

Depletion, regulation, stress and movement

Four directions are named for the future, and the first contradicts an assumption the theory has held from the start. Chen and colleagues reported in 2018 that working memory resources deplete after sustained effort and recover with rest, and used this to explain the spacing effect, the oldest instructional effect psychology has produced and the one with the least agreed cause: massed practice reduced measured capacity where spaced practice allowed it to recover. If that holds, capacity is not fixed for an individual with long-term memory as its only determinant, and the authors accept that the fixed-capacity assumption must go. They keep their distance from ego depletion, noting that avoiding sweets on a diet is unlikely to load working memory as learning mathematics does, and citing meta-analyses that conflict. The second direction joins the theory to work on self-regulated learning, where whether regulation can be taught is the primary-secondary question again. De Bruin and van Merriënboer propose Koriat's account of cues as the bridge: students judge their learning from cues and favour invalid ones, so that a text which reads easily is judged understood and a problem that took great effort is judged to have taught a great deal, when the effort may have gone entirely on means-ends analysis. Third, stress, emotion and uncertainty compete with the task for working memory, and the usual advice is to remove them; but for nurses attending the dying, officers facing violence and doctors deciding on incomplete information they are part of what must be learned, which makes them intrinsic rather than extraneous load. The fourth direction extends the finding that people learn motor tasks better from animation than from statics, which the authors attribute to an evolved capacity to learn by watching others act, to movement made rather than observed: gesturing offloads information during problem solving, and tracing a worked example with a finger improved learning in geometry.

Since 2019

The paper closed by naming the theory's strengths, among them that it generates testable hypotheses whose failure modifies it and that most of its evidence comes from randomised controlled trials; "the ultimate test is empirical", the authors write of the architecture, and the effects are that test. On the practical side the prediction has held. Cognitive load theory is now one of the few pieces of cognitive psychology that ordinary teachers cite, carried into staffrooms by education-department reports of the kind the paper mentions and by popular treatments written since, and its advice on worked examples and split attention is routine in teacher training.

The contested parts are the ones the paper leans on hardest. The evolutionary foundation is an inference rather than a measurement: no independent test identifies a piece of knowledge as biologically primary, and the argument that generic skills cannot be taught rests on the failure of attempts to teach them, which is equally consistent with the teaching having been poor. The claim that guidance must come first meets a rival in the productive-failure research, where learners who struggle with problems before being taught the solution outperform those taught first, a result its proponents attribute to what the struggle prepares rather than to what it teaches; the two camps agree on much of the evidence and disagree about which measures count. Measurement remains the weak point, and instruments built since to separate the categories have had their factor structure disputed much as Leppink's was. The depletion direction is the most exposed: the ego-depletion literature beside which it sits failed a large multi-laboratory replication in 2016, and the authors' own caution about the analogy now reads as the safer half of their argument. What has held up with little argument is the part that was never controversial: that novel material overloads a small working memory, that expertise dissolves the problem by storing as one element what was once many, and that most of the difficulty in teaching something is imposed by how it is presented rather than by what it is.