Understanding hard material

The proof runs a page and a half. The mechanism has seven arrows. The case-law paragraph has one sentence nine lines long. You read it, you follow it, you nod, and then you turn the page and cannot say what it said.

Lessons 3 to 5 gave you tools for holding on to things you already understand. This lesson is about the step before that: getting a hard thing into your head at all. The research says three things, and each comes with a limit. For a beginner learning a procedure, studying a solution beats attempting the problem unguided. That stops being true as you improve. And the difference between reading an explanation and understanding it is a specific habit you can learn.

The core idea

Start in an algebra classroom. Sweller and Cooper (1985) gave Year 9 algebra students pairs of similar problems. One group solved both. The other group studied the first as a worked solution and then solved the second.

Look at the design before you take a lesson from it. This wasn't "reading instead of doing". It was a worked example immediately followed by a problem of the same type. And there's a boundary: the advantage did not extend to problems that used the same algebraic moves in a different order. The worked example taught the pattern shown, not the general skill.

The second finding is that the benefit reverses as you gain expertise. Kalyuga and colleagues (2003) named this the expertise-reversal effect. Guidance that helps a novice becomes redundant for a more knowledgeable learner, and it can actively harm one who already has the pattern, because processing an explanation you no longer need costs working memory that could go on the task.² The same worked example that builds the schema in week one is dead weight once the schema exists.

The third finding is about what happens while you read. Chi and colleagues (1989) recorded eight university students with no college physics working through a mechanics text and its worked examples, thinking aloud. The students who went on to do well explained each step to themselves as they read, asking what it was for and why it followed, and they noticed accurately when they were lost.

That study couldn't tell whether the habit caused the difference or whether better students simply had it. So Chi and colleagues (1994) prompted ordinary eighth-graders to explain a text on the circulatory system to themselves, sentence by sentence. The fourteen who were prompted learned more than the ten who read the text twice.⁴ Bisra and colleagues (2018) pooled the later studies of prompted self-explanation and found a moderate overall benefit, g = 0.55 across 69 effect sizes.¹⁰

There's a related habit, elaborative interrogation, which comes from Pressley and colleagues (1987): as you read a claim, ask "why would that be true?" and answer it.⁵ Dunlosky and colleagues (2013) rate both self-explanation and elaborative interrogation as moderate utility, mainly because neither has been adequately tested in real classrooms. They add a condition for elaborative interrogation: its effects grow with prior knowledge and are uncertain for learners who have little.⁶ Keep that condition in mind. It comes back later.

Why it works: cognitive load

Lesson 2 gave you the machinery. Working memory holds about four chunks; long-term memory is where the chunks live; learning is building new chunks out of what passes through the small space.

Now put a novice in front of a problem. She doesn't have the pattern yet, so she has to search: try a move, check whether it got closer to the goal, back up, try another. Every attempt occupies working memory. Sweller's account is that searching for a solution and building the schema compete for the same space, and a novice mid-search has little working memory left for the second.¹⁷ She may even reach the answer, but she spent her working memory on getting there rather than on noticing how, so she can't do it again.

A worked example removes the search. The pattern is on the page: here is the move, here is why, here is the next move. Working memory is free to do the one thing that matters, which is to build the chunk. That is the worked-example effect, and the cognitive-load account explains its boundary too. The chunk built is the one shown. A problem that requires a different move exposes that the learner has a pattern, not yet a method.

It also explains the reversal. A learner who already has the chunk gains nothing from reading a full explanation: it means processing information that is already in long-term memory and reconciling it with what she knows, which is work with no payoff. For her, the problem is the better exercise, because attempting it now exercises the schema rather than searching for one.

Self-explanation works on a different gap. Every explanation ever written skips steps. The author, who understands, doesn't see the gaps; the reader, who doesn't, falls into them without noticing, because a skipped step doesn't announce itself. Asking "why does this line follow from the last one?" forces you to reconstruct the missing step, and the moment you can't is the moment you have found where your understanding actually stops. Chi's 1989 study can't say whether the good students were cleverer; what it recorded was that they kept checking, and the 1994 study showed that students told to check learned more.

Worked example: an equation, three times

The full cycle on a piece of algebra. You can almost certainly do this already, which means you won't feel the novice's search; what you can see is the shape, and the shape is what you'll reuse on material you can't yet do. Stage 1 and Stage 2 together are one example-problem pair, the unit Sweller and Cooper actually tested.

Stage 1: study the worked example. Solve for x: 3(x − 4) = 2x + 1.

• Expand the bracket, because you can't collect x-terms while one of them is locked inside a bracket: 3x − 12 = 2x + 1.
• Get every x on one side. Subtract 2x from both sides: x − 12 = 1.
• Get every plain number on the other side. Add 12 to both sides: x = 13.
• Check by substituting back: 3(13 − 4) = 27; 2(13) + 1 = 27. Done.

Read that as Chi's good students would. Why expand first? Because the bracket hides an x. Why subtract 2x rather than 3x? Either works; subtracting the smaller keeps x positive, one less thing to go wrong. Why check? It's the only step that catches an arithmetic slip. Any "why" with no answer is a gap, and you fill it now, before attempting anything.

Stage 2: attempt a near-transfer problem. Solve 5(x − 2) = 3x + 4.

The moves are the same in the same order. Expand first: 5x − 10 = 3x + 4. Now stop and do the next two moves yourself, on paper, before you read on.

If Stage 1 went in, that felt almost mechanical. That's the Sweller and Cooper result: faster, fewer errors, on problems that share the studied pattern.

Stage 3: attempt a varied problem. Solve (x + 3) / 2 = x − 1.

Now the pattern from Stage 1 stalls. There is no bracket to expand; there is a fraction. A learner who has only the studied pattern will try to collect x-terms and find the x trapped in a numerator. The move needed, multiply both sides by 2 to clear the denominator, is not in the example. Once you make it (x + 3 = 2x − 2, so x = 5), the rest is the familiar pattern, but the first step had to come from somewhere else.

In 1985 the advantage ran out on problems that used the same moves in a different order; a problem needing a new move, like this one, is further out still. Worked examples didn't fail; the method is study examples of each pattern you need, not study one example and expect the skill. A varied problem you can't start is the signal to find a worked example of that variation, explain it, and return.

Stage 4: fade. After a few rounds, complete worked examples stop helping, and Kalyuga's effect says they start to cost. Move to completion problems: a solution with one or two steps blanked out for you to fill in. Then problems alone. Then mixed problems of several types, which is lesson 5's interleaving arriving on schedule. Guidance goes from full, to partial, to none.

When to fade is a decision, and lesson 1 warned you not to make decisions like this by feel. So use a count. Fade one step when you have solved three near-transfer problems in a row without looking at the example; fade the next when the completion problems go three for three; move to varied and mixed problems only after that. This is the same three-correct-recalls criterion that Rawson and Dunlosky set for retrieval practice (lesson 3), applied to guidance. If reading the full solution has started to feel like a chore, that's a prompt to run the test, not the test itself.

A harder example: a paragraph you can't get past

The wrinkle: most hard material isn't a problem with a solution. It's prose. Here's a dense passage in the style of the cognitive-load literature (written for this lesson, not quoted from anyone), and a self-explanation pass on it.

Instruction should manage intrinsic load, which arises from the element interactivity of the material, and reduce extraneous load, which arises from how the material is presented; germane load is now understood not as a third source of load but as the working-memory resources a learner devotes to dealing with intrinsic load.

Read once, it slides past. Now take it a clause at a time with three prompts: what does this add? why is that true? how does it connect to what came before? The prompts are this lesson's, modelled on the sentence-by-sentence prompting in Chi's 1994 study. One honesty note: the "why would that be true?" evidence from Pressley's work comes from single factual sentences, and Dunlosky's review says its reach on long, complex text is untested.⁵⁶ On dense prose the better-tested habit is self-explanation, and that is what the pass below is.

"Instruction should manage intrinsic load, which arises from the element interactivity of the material." What does it add? A definition: intrinsic load is set by how many things you have to hold at once to understand the material. Why "manage" rather than "reduce"? This is the first gap. You can't reduce the number of interacting elements in, say, a chemical equilibrium without changing what's being taught. You can only sequence them, teach parts before the whole. So "manage" is doing real work. If you'd read on without stopping, you would have missed the distinction the whole sentence is built on.

"…and reduce extraneous load, which arises from how the material is presented." What does it add? A second kind of load, and a verb change: this one can be cut. Why? Because presentation is the author's choice, not the subject's nature. How does it connect? It contrasts with the first clause: one load is in the material, the other is in the delivery. That contrast is the point of the sentence.

"…germane load is now understood not as a third source of load but as the working-memory resources a learner devotes to dealing with intrinsic load." What does it add? A correction to an older idea. Why is that true? Second gap: the passage doesn't say, and you can't reconstruct the reason from the sentence. You have to go to the source. Here it is, from Sweller's 2010 paper and the 2019 review: the two loads you've just met can both be defined by element interactivity, and a third load that couldn't be defined that way sat outside the theory's own terms, so it was folded into the first.⁷¹¹ How does it connect? It closes the count: two loads, not three, and your effort should go on the first.

Two gaps in one sentence, both invisible on a straight read. The first you could fill by thinking; the second needed a source, and notice that when the source supplied it, the answer was short. That's normal. Self-explanation doesn't guarantee understanding; it guarantees you find out where it stops, which rereading never does.

Notice also what the pass required: you had to already know what working memory is. Without lesson 2, "why is that true?" has nothing to grip. That's the prior-knowledge condition Dunlosky attaches to elaborative interrogation.⁶ On material where you have no base at all, the honest order is: get some facts and cases in first (retrieval practice on the basics), then come back and interrogate.

What people get wrong

"Struggling first is always best." It's the most common misreading of desirable difficulties. Lesson 1 gave the rule: a difficulty is desirable only if you can meet it. A novice searching for moves she doesn't have is not meeting a difficulty; she's spending working memory on search. Worked examples first. The struggle comes later, on purpose, when you fade. There is a designed version of struggling first, Kapur's "productive failure", in which the problem is chosen to expose what you don't know and is always followed by a full explanation; whether that beats examples-first for understanding is a live research question, discussed below, and it is not what 40 unguided minutes on a problem set is.

"If I can follow the solution, I could produce it." This is lesson 1's fluency illusion in a new coat. Following a worked example raises retrieval strength for the next few minutes and feels like understanding. Chi's poor students were confident. The only test is the near-transfer attempt, done without looking, and then the varied one.

"Putting it in my own words is self-explanation." Paraphrase restates the step; self-explanation says why the step is there and what would go wrong without it. You can paraphrase a line you don't understand. You can't explain why it follows from the last one without either understanding it or discovering that you don't, and the discovery is the point.

"Summarising and highlighting are elaboration." They aren't. Dunlosky and colleagues rate summarisation and highlighting low utility, and in one study highlighting hurt on tasks that needed inference.⁶ Both let you process the surface of a text without ever asking why a sentence is true. Elaboration is the "why" question and its answer, connected to something you already know.

"Guidance is training wheels; real learners don't need it." Expertise reversal cuts both ways: learners who have the pattern don't need full guidance, novices do. Not needing worked examples is a stage you reach in each subject separately, not a trait.

The contested question: guided or discovery?

One argument in this area is not settled, and you should know its shape.

Kirschner, Sweller and Clark (2006) argued that instruction with minimal guidance (discovery, problem-based and inquiry learning, in the forms they targeted) does not work for novices. Their case was the cognitive-load reasoning above, that without the pattern in long-term memory learners search rather than learn, and the studies they reviewed, including earlier reviews of pure discovery learning and of problem-based medical curricula that found no knowledge advantage.⁸ Hmelo-Silver, Duncan and Chinn (2007) replied that the problem-based and inquiry approaches actually used in good classrooms are not minimal guidance at all; they are heavily scaffolded, with structure, prompts, and expert modelling built in. They also argued that the two sides were measuring different things: inquiry designs aim at scientific reasoning, self-direction and motivation as well as content, and they pointed to classroom studies showing content gains too.⁹ Sweller, Kirschner and Clark (2007) answered that if problem-based learning as practised is heavily scaffolded then it is not what they criticised, the label is misleading, and the burden is on its advocates to show in controlled comparisons that the scaffolded versions beat explicit instruction.¹²

Part of the disagreement is definitional: the two sides don't mean the same thing by "guidance". Part is empirical and open: how much structure, of what kind, at which stage. What would settle it is trials that specify the scaffolding precisely, measure prior knowledge, and measure the outcomes each side cares about, including delayed transfer and reasoning as well as immediate content tests, so that "problem-based" stops being one label for unguided and richly guided designs alike.

There is a second, more recent form of the same dispute, and it bears directly on this lesson. Kapur and colleagues argue for "productive failure": novices first attempt a carefully designed problem, usually in groups, and only then receive the full explanation. Sinha and Kapur (2021) pooled 53 studies that compared the two orders, problem-solving then instruction against instruction then problem-solving.

The cognitive-load side disputes this, and the two literatures differ in what they measure: the worked-example studies measure procedural performance on the taught problem type; the productive-failure studies mostly report conceptual understanding and transfer. Whether a designed attempt before instruction can beat examples-first for understanding is open, and it depends on how the attempt is structured and what is tested.

What isn't seriously disputed is the piece this lesson rests on. For a novice learning a procedure, studying worked examples beats unguided attempts at the same kind of problem, and guidance should fade with expertise. Neither Kirschner, Sweller and Clark nor Hmelo-Silver, Duncan and Chinn dispute that novices need support or that it should fade. The argument is about the rest.

Practice

Connections

This lesson is why lesson 2 came before lesson 3. Retrieval, spacing and interleaving all assume there is something in memory to retrieve, space and mix. Worked examples and self-explanation are how it gets there. The fade is where lesson 5's mixed practice begins, and once the schema exists, every problem attempted without the example is the schema being used rather than looked up, which is where lesson 3's retrieval practice takes over.

You should also now see how the lessons on this site are built. Each gives a worked example with the reasoning shown, then a harder one with a wrinkle that exposes the edge, then practice you do before the quiz. That order is Sweller and Cooper, then Kalyuga, then Chi. The wrinkle is there because the 1985 benefit didn't extend to varied problems, so a second, varied example is the least a lesson owes you. The practice is there because the fade has to happen, or you'll be left following solutions and believing you could produce them.

Lesson 7 takes the same logic to skills: the worked example becomes a demonstration, and the fade becomes coaching that withdraws.

Go deeper

• Sweller, van Merriënboer & Paas, "Cognitive Architecture and Instructional Design: 20 Years Later", Educational Psychology Review 31 (2019). The theory's authors restating it with its effects, including worked examples and expertise reversal, in one paper; free to read at the link.
• Chi, Bassok, Lewis, Reimann & Glaser, "Self-Explanations", Cognitive Science 13 (1989). The original protocol study. The transcripts of good and poor students are the clearest picture of what self-explanation actually looks like.
• Dunlosky, "Strengthening the Student Toolbox", American Educator (Fall 2013). Free online. The plain-English ratings of all ten techniques, including why elaborative interrogation and self-explanation are moderate rather than high.
• Ollie Lovell's interview series with John Sweller: the theory's author answering a teacher's practical questions, free as text and audio; the worked-examples and expertise-reversal instalments cover this lesson's ground.
• The Educational Psychologist 42(2) exchange (2007): Hmelo-Silver, Duncan & Chinn; Schmidt, Loyens, van Gog & Paas; Kuhn; and the Sweller, Kirschner & Clark reply, following Kirschner, Sweller & Clark (2006). Read the whole exchange rather than a single pair, to see a scientific disagreement conducted well and to decide for yourself how much of it is about definitions.

Sources

1. Sweller, J. & Cooper, G. A., "The use of worked examples as a substitute for problem solving in learning algebra", Cognition and Instruction 2, 59–89 (1985). Year 9 students given example-problem pairs versus problem pairs. Experiment 3: test problems solved in 43.6 s versus 78.1 s, with 0.18 versus 1.64 errors per problem; Experiment 2 found the error difference but no significant time difference. The advantage did not extend to "dissimilar" problems, which used the same algebraic operations in a different order. Also the source of the claim that problem-solving search and schema acquisition compete for attention.
2. Kalyuga, S., Ayres, P., Chandler, P. & Sweller, J., "The expertise reversal effect", Educational Psychologist 38, 23–31 (2003). Guidance that helps novices becomes redundant or harmful as expertise grows.
3. Chi, M. T. H., Bassok, M., Lewis, M. W., Reimann, P. & Glaser, R., "Self-Explanations: How students study and use examples in learning to solve problems", Cognitive Science 13, 145–182 (1989). Eight university students with no college physics, split into good and poor by later problem-solving success. Good students self-explained worked examples and monitored accurately; poor students seldom detected comprehension failures and reread the examples during problem solving, searching for a line to use.
4. Chi, M. T. H., de Leeuw, N., Chiu, M.-H. & LaVancher, C., "Eliciting self-explanations improves understanding", Cognitive Science 18, 439–477 (1994). Fourteen eighth-graders prompted to self-explain a circulatory-system text learned more than ten who read it twice.
5. Pressley, M., McDaniel, M. A., Turnure, J. E., Wood, E. & Ahmad, M., "Generation and precision of elaboration: Effects on intentional and incidental learning", Journal of Experimental Psychology: Learning, Memory, and Cognition 13, 291–300 (1987). Origin of elaborative interrogation ("why would that be true?"), tested on single factual sentences.
6. Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J. & Willingham, D. T., "Improving Students' Learning With Effective Learning Techniques", Psychological Science in the Public Interest 14, 4–58 (2013). Self-explanation and elaborative interrogation rated moderate utility because the evidence is limited, in particular not adequately evaluated in educational contexts; elaborative-interrogation effects grow with prior knowledge and are less certain for low-knowledge learners; the prior-knowledge moderator for self-explanation is under-studied. Summarisation and highlighting rated low; highlighting hurt inference performance in one study (Peterson 1992), which the authors say needs replication.
7. Sweller, J., van Merriënboer, J. J. G. & Paas, F., "Cognitive Architecture and Instructional Design: 20 Years Later", Educational Psychology Review 31, 261–292 (2019). Cognitive load theory: intrinsic load from element interactivity, extraneous load from presentation, and the reconceptualisation of germane load as working-memory resources devoted to intrinsic load; means-ends search described as exceptionally expensive of working memory and unrelated to knowledge construction.
8. Kirschner, P. A., Sweller, J. & Clark, R. E., "Why Minimal Guidance During Instruction Does Not Work", Educational Psychologist 41, 75–86 (2006). The case, from cognitive load and from reviewed studies, against minimally guided instruction for novices; learners with high prior knowledge exempted.
9. Hmelo-Silver, C. E., Duncan, R. G. & Chinn, C. A., "Scaffolding and Achievement in Problem-Based and Inquiry Learning: A Response to Kirschner, Sweller, and Clark (2006)", Educational Psychologist 42, 99–107 (2007). Reply: problem-based and inquiry learning as practised are heavily scaffolded, not minimal guidance; they target reasoning, self-direction and motivation as well as content, and classroom studies show content gains.
10. Bisra, K., Liu, Q., Nesbit, J. C., Salimi, F. & Winne, P. H., "Inducing Self-Explanation: A Meta-Analysis", Educational Psychology Review 30, 703–725 (2018). Meta-analysis of self-explanation prompts: g = 0.55 across 69 effect sizes.
11. Sweller, J., "Element Interactivity and Intrinsic, Extraneous, and Germane Cognitive Load", Educational Psychology Review 22, 123–138 (2010). Defines extraneous load in terms of element interactivity and germane load in terms of intrinsic load, so that all load is tied to element interactivity.
12. Sweller, J., Kirschner, P. A. & Clark, R. E., "Why Minimally Guided Teaching Techniques Do Not Work: A Reply to Commentaries", Educational Psychologist 42, 115–121 (2007). Reply to Hmelo-Silver et al. and others: scaffolded problem-based learning is not what was criticised; calls for controlled comparisons that vary one thing at a time.
13. Sinha, T. & Kapur, M., "When Problem Solving Followed by Instruction Works: Evidence for Productive Failure", Review of Educational Research 91, 761–798 (2021). Meta-analysis of 53 studies, 166 comparisons: problem-solving followed by instruction beat the reverse order, g = 0.36 (95% CI 0.20 to 0.51), with larger effects (g = 0.37 to 0.58) when the design followed productive-failure principles closely.