Learning Sciences: How People Learn

A teacher explains fractions with a lovely diagram. The class nods, the quiz that afternoon goes well, and a month later half the room cannot say which of two fractions is larger. Nobody was lazy and the explanation was good. What failed was a belief almost everyone holds: that understanding something once means you will still have it.

The learning sciences are the research fields that test that belief: cognitive and developmental psychology, classroom experiments, and the design of instruction. Their best-supported results are surprisingly stable and surprisingly unlike intuition. Practice that feels smooth tends to work worse than practice that feels effortful. A sentence you reread is more familiar, not better known. A beginner and an expert are not the same kind of learner with different amounts of knowledge; they need different teaching.

This piece goes beyond the basics of memory. The library already covers the forgetting curve, active recall and flashcards, so here we ask the next questions: what is learning made of, why does the bottleneck matter, when does a good technique stop working, and how big are the effects really?

Learning is a change in long-term memory, built from what you already know

Strip away the jargon and learning, for most school purposes, means that something is stored in long-term memory that was not there before, and that you can use it later in a new place. Kirschner, Sweller and Clark put it bluntly in 2006: if nothing has changed in long-term memory, nothing has been learned. This does not make learning "just memorising". Understanding is a rich structure of linked memories, which psychologists call a schema.

The consequence is that prior knowledge does most of the work. In 1973 William Chase and Herbert Simon showed chess masters and beginners boards for a few seconds. Masters rebuilt real game positions far better, but on boards with the pieces placed at random their advantage mostly vanished. They did not have bigger memories; they had stored patterns. In a 1988 study of reading by Recht and Leslie, children who were poor readers but knew a lot about baseball understood a baseball passage better than good readers who knew little about the game. What you can learn next depends on what you hold now.

A graduate's habit: before asking "was my explanation good?", ask "what could this learner hook it to?" Many wrong answers are not random; they are reasonable ideas built on an earlier, partial schema (children who think a heavier ball must fall faster have watched feathers and stones). Good teaching finds the existing idea and works from it.

Working memory is the bottleneck, so load is the design problem

Cognitive Psychology (35/22) gave the number: working memory holds roughly four chunks at once, and only briefly. John Sweller turned that into an instructional theory in the 1980s, cognitive load theory. It separates three kinds of load. Intrinsic load is how many elements of the task must be held in mind together (it depends on the task and on what you already know). Extraneous load is effort wasted by how the material is presented, such as a diagram on one page and its labels on another, which forces you to hold one while hunting for the other (the split-attention effect). Germane load, effort spent building schemas, is the one you want, and later versions of the theory fold it into the first two.

The early signature result is the worked-example effect. Sweller and Cooper (1985) found that beginners in algebra learned more from studying solved problems than from solving the same number of problems, because searching for a method eats the working memory that learning needs. Then comes a twist, the expertise reversal effect (Kalyuga and colleagues, 2003): the same worked example that helps a novice can slow an expert, because the expert must now process guidance they do not need. The right amount of help depends on the learner.

The numbers are bookkeeping, not measurements: element counts and chunk sizes are chosen for illustration, and real load is not measured in slots. The direction of each effect is what the research supports.

Retrieval beats rereading, but the win shows up only after a delay

The library's memory course covers the testing effect. The detail worth keeping is when it appears. In 2006 Henry Roediger and Jeffrey Karpicke had students study prose passages, then either restudy or take recall tests without feedback, and tested everyone after 5 minutes, 2 days or 1 week. On the 5-minute test, restudying was better. On the 2-day and 1-week tests, the students who had been tested remembered far more, even though restudying made them feel more confident. If you measure learning right after a lesson, the weaker method looks fine.

That is the central trap of the field: performance during or right after practice is not learning. Robert Bjork named the useful opposite desirable difficulties (1994): conditions that slow immediate performance and improve lasting learning, such as retrieving instead of rereading. The word "desirable" matters. Difficulty that does not require retrieval or thought, like a badly printed handout, is just difficulty.

The curves are invented (exponential decay with made-up starting points and speeds) so the crossing is visible. The real experiments found the same crossing, with different numbers.

Spacing has a best gap, and it depends on when you need the memory

Spacing beats cramming; the memory course shows that. The next question is how far apart to space. Nicholas Cepeda and colleagues (2008) taught more than a thousand people facts and varied both the gap between two study sessions and how long they waited for the final test. The best gap grew as the test got further away, but as a share of the delay it shrank: roughly a tenth to a fifth of the delay for tests weeks or months ahead, and a smaller share for tests a year ahead. An earlier study by Harry Bahrick and colleagues (1993) had followed language learners for years and found relearning every 56 days beat every 14 or 28 days for long-term retention.

The reason is a trade-off. A longer gap means the memory has faded more, so a successful retrieval strengthens it more, but if it has faded too far the review fails and becomes relearning from scratch. The toy below is a deliberately small model of that trade-off. Move the delay and the gap and watch the best gap move.

Toy model: the memory fades with a time constant of 5 days. A review that succeeds (chance equals what is left) multiplies that time constant by up to 9; one that fails gives no boost. All constants are invented. A single review cannot carry a memory for a year, which is why real plans use several, as the memory course shows.

Mixing problem types feels worse and works better

Textbooks usually give ten problems of one kind, then ten of the next. That is blocked practice. Interleaved practice mixes the kinds. During practice, blocked looks better: you know which method to use because the last problem used it. Mixed practice forces you to decide which method fits, and that decision is the skill an exam, or life, will test. In a 2007 study by Doug Rohrer and Kelli Taylor, students who practised volume formulas in mixed order did worse during practice and, on a test a week later, roughly three times as well as the blocked group. Kornell and Bjork (2008) found the same with paintings: people learned artists' styles better from mixed sets but, on average, believed the blocked sets had worked better.

Interleaving is not free. It helps most when the categories are easily confused and the learner can already do each kind, and it is a poor fit for a first encounter with something brand new. Dunlosky and colleagues (2013) rated it moderate in utility, with less classroom evidence than retrieval or spacing.

The sequence and the count of switches are computed live. The bars show the shape of the published finding (better during practice when blocked, better later when mixed) with round invented numbers, not the data of any one study.

Feeling that you learned is not learning

If effortful methods work better but feel worse, learners will often choose the wrong ones. This is why metacognition, judging your own learning, is part of the course. Fluent, familiar material produces an illusion of competence: you recognise it, so you think you could produce it. Self-testing exposes the gap, which is a second benefit of retrieval: it corrects your own forecast.

The same trap sits under the best-known myth in education, learning styles: the idea that people learn better when taught in their preferred mode (visual, auditory, kinaesthetic). Many people do report preferences, and most teachers have been told to teach to them. But the claim that matters is the matching hypothesis: that people given their preferred format learn better than people given another. Pashler, McDaniel, Rohrer and Bjork reviewed this in 2008 and found almost no studies designed to test it properly, and the few that were did not support it. The test has a clear shape, shown below. What does help is matching the format to the content: a map is better shown than described, whoever is looking.

Evidence has grades, and effect sizes need honest reading

"Research shows" is not one thing. The Education Endowment Foundation's Teaching and Learning Toolkit (England) and the US What Works Clearinghouse both grade approaches by strength of evidence, size of effect and cost. Dunlosky and colleagues (2013) did it for ten study techniques. Putting the results together gives a rough map:

An effect size such as Cohen's d says how far apart two groups' averages are in standard deviations. John Hattie's Visible Learning (2009) pooled hundreds of meta-analyses and used d = 0.40 as a "hinge point" for what is worth doing. Critics, including Robert Slavin, Dietmar Terhart and Bergeron and Rivard (2017), argue that comparing d values across studies of different designs and ages is misleading and that some of the averaging is statistically unsound. Matthew Kraft (2020) suggested that for well-designed studies on standardised achievement tests, 0.05 is small, 0.05 to 0.20 is medium and above 0.20 is large, much lower than the old rule. Growth-mindset messages are a live example: Sisk and colleagues' 2018 meta-analyses found small average effects, and the 2019 US national experiment by Yeager and colleagues found an average gain of about a tenth of a grade point, concentrated among lower-achieving students in supportive schools. Their authors and critics read those numbers differently.

Computed live from the normal distribution: the chance a random treated student beats a random control is Φ(d/√2), and the average treated student sits at the Φ(d) percentile of the control group. It assumes equal spread and bell-shaped scores.

What a graduate just sees

• Learning is a lasting change in long-term memory; a clear explanation that leaves no trace has taught nothing.
• Prior knowledge is the hook; the same lesson lands differently on different learners.
• Working memory is the bottleneck, so cut waste, split steps and show worked examples to beginners.
• Help that supports a novice can hamper an expert (expertise reversal).
• Testing yourself beats rereading, and the difference shows after a delay, not in the next ten minutes.
• The best gap between reviews grows with how long you need the memory.
• Mixed practice feels worse and transfers better.
• Feeling fluent is not knowing, and "learning styles" fail the one test that would matter.
• Every "it works" has a grade and an effect size, and both should be asked for.

Where it shows up next

• Curriculum & Instructional Design (36/54): worked-example fading, sequencing and the guidance debate all follow from load and prior knowledge.
• Teaching Reading & Mathematics (36/56): decoding, fluency and number sense are schemas that must be automatic before comprehension and problem solving can happen.
• Assessment & Measurement (36/55): if performance is not learning, tests have to be built to tell them apart, and testing itself is a learning tool.
• Classroom Practice & Educational Technology (36/60): flashcard apps, adaptive tutors and AI helpers are judged by these ideas, for example whether an AI that gives the answer removes the retrieval.
• In the world: language apps, driving schools, medical training, pilot simulators and corporate courses all face the same bottleneck and the same temptation to measure what is easy.

If you want to go deeper

• 📘 Textbooks. Susan Ambrose and colleagues, How Learning Works: Seven Research-Based Principles for Smart Teaching (Jossey-Bass, 2010), the usual university teaching-centre text; and Dale Schunk, Learning Theories: An Educational Perspective (Pearson, recent editions) for the theories side. Check for the current edition.
• 📘 Gentler. Peter Brown, Henry Roediger and Mark McDaniel, Make It Stick (Harvard, 2014), and Daniel Willingham, Why Don't Students Like School? (Jossey-Bass, 2009; second edition 2021).
• 🌐 Free text. National Research Council, How People Learn: Brain, Mind, Experience, and School (2000), free to read at nap.nationalacademies.org.
• 🌐 Free course. MIT OpenCourseWare 5.95J Teaching College-Level Science and Engineering (Fall 2015), which applies this research to teaching.