← Back to blog

Learning Science

Why a clear explanation can still leave you unable to solve the problem

How worked examples help, where they stop helping, and how to move from following a solution to solving one

August 21, 2026 · 13 MIN READ

A worked solution fading into a similar unsolved problem with the next decision left blank

The instructor writes two equations on the board and solves the system by substitution. First isolate a variable. Substitute the expression into the other equation. Simplify. Solve. Put the result back into the first equation. Check both values.

Nothing seems mysterious. Each line follows from the one above it, and by the end the whole solution has the clean inevitability of a route traced in red on a map.

Then you open the exercise set. The numbers have changed. The wording is slightly different. More importantly, nobody has written use substitution above the problem. You understand the solution you just watched, yet the blank page offers no obvious first move.

This does not necessarily mean you failed to understand the explanation. You may genuinely understand why every visible step was valid. The missing ability appears one level earlier: before a step can be judged, it has to be generated; before a method can be executed, it has to be selected.

To understand difficult concepts well enough to use them, study a worked example, explain why its important decisions make sense, hide some of the support, and then solve a changed problem without being told the method. A clear explanation shows that you can follow a solution. An independent attempt shows which parts of the solution you can produce yourself.

Following a solution is not the same as building one

When you follow a solution, the main question is whether the next line makes sense. When you build one, you face a longer series of decisions. What kind of problem is this? Which details matter? What should the situation be turned into: an equation, a diagram, a causal model, a data structure? Which method applies? What should happen first? At what point should the answer start looking suspicious?

A finished example has already settled most of this. It presents a useful representation, directs attention toward the relevant features, selects a strategy, orders the steps, and usually removes the false starts. By the time you see line four, the route has been cleaned up by someone who already knows where it ends.

The distinction appears far beyond algebra. You can understand every line of a binary-search implementation and still miss that the data must be sorted. You can follow a paired-samples analysis and then misidentify a study with independent observations. You can understand confounding after the confounder has been named, then miss it in a new paper where no label is attached.

“Understanding” is therefore a poor single switch. A learner may be able to follow a solution, explain it, complete part of it, reproduce the procedure, choose the procedure, or adapt it when the conditions change. Those are related achievements, but they are not interchangeable.

Worked examples are useful for a reason

This is not an argument against examples. For a novice, an example can be the most sensible place to begin.

Imagine trying to learn a multi-step method while also searching blindly for the method. Every attempted move competes with the effort required to understand what the moves mean. A worked example reduces that search. It lets the learner inspect a successful structure before having to invent one.

In a recent experiment, Ouhao Chen and colleagues assigned 114 novice learners to study multi-step algebra through worked examples or through problem solving. The example group reported lower cognitive load at every measured solution step and later performed better on retention and near-transfer tests. A 2023 meta-analysis reached beyond one experiment: across 55 mathematics studies and 181 reported effects, worked examples produced a moderate average advantage. The benefit varied with the design, so “worked example” should not be treated as one uniform intervention.

The point is straightforward. When you do not know the route, discovering the route and learning from it are two jobs at once. A worked example temporarily performs part of the first job so that you can concentrate on the second.

The mistake begins when the support is mistaken for evidence that support is no longer needed.

The example is making decisions for you

Consider the heading Solving systems by substitution. It looks like a label. Instructionally, it is also an answer.

The student still has to carry out substitution, but the textbook has removed a major decision: Should substitution be used here? A page of twenty similar exercises removes it twenty times. On an exam, systems suited to substitution may sit beside systems that are easier to solve by elimination. The method name disappears, and classification becomes part of the task.

Classic physics research by Michelene Chi, Paul Feltovich, and Robert Glaser helps explain why this is difficult. When experts and novices sorted physics problems, experts organized them by the principles needed to solve them. Novices leaned toward literal features: inclined planes, pulleys, springs. Seeing a principle after an expert has selected it is easier than recognizing that the principle applies beneath unfamiliar surface details.

A polished solution can therefore create what we might call local clarity. Each step is clear inside a structure that has already been chosen. What remains uncertain is whether the learner could choose that structure before seeing the step.

The example did not fail. It answered several questions before you noticed they were questions.

Explain the decision, not the narration

Self-explanation is often recommended as the cure: explain the example to yourself. That advice is useful only if “explain” means more than narrating the visible operations.

“First the author subtracts four. Then the author divides by two” is a description. It does not say why subtracting four serves the goal, which feature of the equation licenses that move, what alternative could work, or when the same move would become invalid.

A better explanation targets decision points. Why is this method appropriate? What condition must be true? What is the step trying to accomplish? What would have to change for another method to become preferable? In binary search, the important explanation is not merely that the search interval is halved. It is that ordered data make one half safely disposable.

In a study of 56 bank apprentices learning compound- and real-interest calculations, Alexander Renkl and colleagues found that prompts eliciting substantive self-explanations supported transferable knowledge, especially on near-transfer tasks for learners with less prior knowledge. Simply giving learners multiple examples did not produce the expected transfer advantage. Seeing more cases and working out what connects them are different activities.

Self-explanation is not magic. A learner can confidently explain the wrong relationship, and a 2023 mathematics meta-analysis found no general bonus from adding self-explanation prompts to worked examples. Prompt design, prior knowledge, feedback, and the object being explained all matter. The useful version of self-explanation makes you reconstruct a relationship the example had supplied. It does not merely make you talk longer.

Remove support one decision at a time

Many courses move from a complete solution directly to an empty page. The learner watches one worked example, then receives a conventional problem and a motivational instruction to “have a go.”

There is a useful middle ground: completion problems and faded examples. Instead of removing the whole solution, the material removes part of it. The learner predicts the next step, supplies a missing strategic move, finishes the second half, or chooses the first operation. As competence grows, more of the reasoning is transferred.

Experiments by Alexander Renkl, Robert Atkinson, and Cornelia Große found that gradually fading solution steps could improve the transition from example study to problem solving. Learners acquired the principles attached to the steps they had to complete. Interestingly, the benefit was not explained by a dramatic rise in self-explanation; fading also reduced unproductive activity. Difficulty itself was not the prize. The removed step gave the learner responsibility for a manageable part of the procedure.

The strategically important part is not always the final algebraic line, which happens to be easy to erase from a worksheet. Sometimes the right thing to hide is the first step, the choice of formula, the interpretation of a result, or the check that an assumption still holds. Fade decisions, not merely ink.

Nor should all guidance disappear on a fixed schedule. Prior knowledge matters, but so does the kind of task. In a statistics study, students with little prior knowledge learned conceptual material best from worked examples, while better-prepared students benefited more from constructing arguments. Yet in a less-structured legal-reasoning task, worked examples remained useful even for advanced law students. There is no universal graduation date from examples.

Practice choosing the method

A student can independently execute a method and still depend on an invisible cue that selects it.

Blocked practice supplies this cue. A chapter introduces the quadratic formula, demonstrates it, then gives a page of quadratic-formula problems. Repetition may improve execution, but each item quietly says: use the method from this page again.

Interleaved practice removes some of that help by mixing problem types. In an eight-week study of 350 university physics students, Joshua Samani and Steven Pan compared conventionally blocked homework with assignments that alternated topics. Blocked homework felt easier, produced better performance during practice, and led students to believe they had learned more. On two later surprise tests containing novel, harder problems, median performance after interleaved practice was 50 percent higher on the first test and 125 percent higher on the second.

One plausible reason is that interleaving creates repeated opportunities to choose a strategy. The authors did not isolate that mechanism; spacing, retrieval, and comparison may also have contributed. The advantage also did not appear on later high-stakes midterms, after students had done substantial additional studying. So “mix everything” is not a law of learning.

The practical idea is narrower and better. Once you know the basic methods, practice deciding among them. Change more than the numbers. Change the wording, the representation, or one condition that determines whether a method applies. Put similar-looking cases beside each other and ask what makes one a substitution problem, another an elimination problem, and a third a case where either method is reasonable.

A changed number mostly tests execution. An unlabeled, mixed problem tests choice.

Use the first wrong decision

When a new problem goes badly, “I don’t understand this topic” is usually too broad to be useful. Find the first decision at which your reasoning diverged from a workable solution.

Sometimes the first failure is a missing prerequisite. A proof can look coherent while the theorem doing the real work is not yet secure; a statistical procedure can make sense line by line while the variables it depends on remain unclear. Before rereading the whole explanation, ask which smaller concept or operation it assumes you can already use. Practice that piece separately, then return.

Perhaps the situation was represented incorrectly. Perhaps you missed the condition that distinguished paired from independent observations. Perhaps the strategy was right but the sequence failed. The calculation may be wrong even though the method was sound. Or the computation may be flawless while the scientific conclusion outruns the study design.

These failures need different repairs. A strategy-selection error calls for mixed and contrasting cases. A missing step calls for a completion problem. Repeated arithmetic or syntax errors call for narrow procedural practice. A monitoring failure calls for explicit checks: does the answer fit the scale, satisfy the original equation, preserve the algorithm’s invariant, or support the causal claim being made?

Recent work on self-assessment after worked examples reinforces this diagnostic use of independent solving. Across two experiments, learners judged their understanding more accurately when they based the judgment on problem-solving performance rather than on how well they felt they could explain the supplied solution, unless the explanation-based judgment received substantial extra support. The empty page is therefore more than practice. It is a measurement instrument.

Turn one example into independent practice

A useful progression looks like this:

  1. See the solution and identify its overall goal.

  2. Name the reason or condition behind each important decision.

  3. Predict the next step before revealing it.

  4. Hide one strategic part and complete it.

  5. Solve a similar problem without the example.

  6. Compare it with a near miss where one changed condition requires another method.

  7. Mix problem types and remove their labels.

  8. Locate the first wrong decision and practice that decision again.

The aim is to move the boundary between what the support decides and what you decide.

The same principle applies to AI explanations used as tutoring rather than answer keys. In a 2025 field experiment involving nearly 1,000 high-school mathematics students, access to GPT-4-based tutors raised scores while the tools were available. But students using the relatively unrestricted version later scored 17 percent below the no-AI group on an unaided exam. A guarded tutor that supplied teacher-designed hints and resisted giving away complete answers largely removed the harm, though it did not create a later advantage over the control group.

That result does not establish that AI damages learning in general. It shows that support design changes the work left to the student. A tool may improve the answer on the screen while reducing how much of the solution the learner constructs.

Use AI, textbooks, videos, or Quizpace to bring principles, conditions, and distinctions back as questions. Ask for a hidden next step, a similar case where the method fails, or two solutions to compare. Flashcards designed around one clear target can help you remember when a method applies and which assumptions matter. They cannot replace the full act of building a multi-step solution.

The instructor’s explanation was clear. It may also have taught you something real. Learning continues when the choices inside that explanation begin to survive its disappearance.

A clear explanation shows why one path works. Being able to solve the problem means choosing a path before anyone has drawn it for you.

References

More to study

Study Systems

Cramming is bad for long-term learning. It can still be rational triage

The night before an exam is too late to master an entire course. It is still enough time to identify the material most likely to matter, test what you can retrieve, repair costly errors, and protect enough sleep to use what you studied.

Notes & Materials

Why taking more notes can leave you with less to remember

What useful notes preserve, what they transform, and how to turn lectures and textbooks into material you can actually recall.