Quick Read
A worked example is a fully or substantially solved mathematics problem that lets a learner inspect how a correct solution is constructed before being required to reproduce the whole process independently.
The evidence does not say students should spend mathematics lessons copying solutions.
It says something more useful:
When a learner does not yet possess a reliable mathematical route, seeing a well-designed correct route can reduce unnecessary search and help the learner construct one.
But the support must then change.
A strong learning sequence is:
Model → Complete → Fade → Execute → Vary → Retrieve → Transfer → Verify
The newest research strengthens that transition.
A 2023 mathematics meta-analysis covering 43 articles, 55 studies and 181 effect sizes reported an average worked-example effect of g = 0.48 on mathematics performance. But it also found important variation between designs and warned against assuming that every addition to a worked example improves it.
A newer 2026 randomised study involving 114 Grade 6 students found that faded worked examples—where solution steps are progressively removed—produced the largest pre-to-post effect sizes among the instructional conditions studied. Importantly, prior mathematical knowledge, rather than working-memory capacity itself, moderated the effect.
And a 2026 pair of experiments provides an important boundary: once learners have already received enough instruction to know the rule, retrieval practice can outperform continued example study for generalisation. Without that initial instructional base, worked examples can have the advantage.
So the principle is not:
Worked examples are better than solving problems.
It is:
Use the amount of guidance appropriate to the learner’s present state, then remove that guidance as capability appears.
Direct Answer
Worked examples can accelerate mathematics learning because they allow a learner to study a valid problem-solving structure without simultaneously having to discover every step through trial and error.
This is particularly valuable when the student is still constructing the relevant mathematical schema.
Instead of asking:
“Can you somehow get the answer?”
the learner can inspect:
- what information matters;
- what mathematical relationship governs the problem;
- which first move is valid;
- what must remain invariant from one line to the next;
- how intermediate states are represented;
- where conditions, signs or restrictions must be preserved;
- and how the final answer can be checked.
The worked example therefore serves as an instructional interface between explanation and independent execution.
That role is already built into the MathematicsOS learning corridor:
diagnose → prerequisite repair → correct model → guided completion → independent same-form practice → spaced retrieval → interleaved selection → transfer → timed verification → delayed independent retest.
The critical word is corridor.
A worked example is a passage through learning.
It is not the destination.
1. Why “Just Try the Question” Can Be Inefficient for a Novice
Consider a student encountering a new mathematics structure.
The student may simultaneously need to:
- understand the question;
- identify the mathematical relationship;
- recall prerequisite knowledge;
- select a method;
- decide the first step;
- manipulate symbols correctly;
- hold intermediate results;
- monitor signs and restrictions;
- judge whether the route is still valid;
- check the answer.
An expert compresses much of this.
A novice does not.
To the experienced mathematician, several lines may constitute one familiar structure.
To the novice, every line may represent another decision.
That distinction explains why asking an inexperienced learner to solve many new problems independently can sometimes train search, guessing or superficial cue-following rather than the intended mathematical relationship.
A correct worked example changes the task.
Instead of searching through the entire possibility field, the learner can examine a route that is already known to be valid.
The mental question becomes:
Why does this step follow from the previous one?
That is a much more constrained learning problem.
2. The Evidence for Worked Examples Is Substantial—but Bounded
The strongest broad mathematics evidence in the current master remains the 2023 meta-analysis by Barbieri and colleagues.
It synthesised 55 studies and 181 effect sizes and reported an average mathematics-performance effect of Hedges’ g = 0.48.
That is meaningful evidence that worked examples can improve mathematical performance relative to comparison conditions.
But the result needs boundaries.
The master correctly records that much of the evidence base involved algebra and geometry; dosage was difficult to evaluate cleanly; prior-knowledge moderation could not be estimated adequately; and instructional designs differed substantially.
Therefore:
g = 0.48 is not a guaranteed gain for a particular child.
It is not a percentage improvement.
It is not evidence that every worksheet should contain solved examples.
And it certainly does not prove that looking at more solutions automatically creates stronger mathematics.
The defensible conclusion is narrower:
Across the studied literature, worked-example instruction produced a meaningful average advantage, but the realised effect depends on the learner, task, design and subsequent transition to independent performance.
That last part has become even clearer in the 2025–2026 evidence.
3. The Important Upgrade: Examples Should Change as the Learner Changes
A mathematics lesson should not ask only:
“Does this teaching method work?”
It should ask:
“Does this teaching method fit this receiver at this point in learning?”
This is the receiver principle.
A complete worked solution can be extremely useful when a learner has little idea how to start.
The same complete solution can become redundant once the learner already understands the structure.
At that point, repeatedly exposing the student to every step may remove precisely the cognitive operation the student now needs to practise.
The instructional requirement therefore changes with capability.
Early state
The student may need:
Problem → complete correct route → explanation of key decisions
Developing state
The student may need:
Problem → partial route → learner completes missing transitions
More stable state
The student may need:
Problem only → independent reconstruction
Transfer state
The student may need:
changed problem → route selection → adaptation → verification
This is fading.
And the newest direct mathematics evidence strengthens the case for treating fading as an active transition rather than a decorative teaching technique.
4. New 2026 Evidence: Fading Worked Examples
Miller-Cotto and Medrano studied 114 Grade 6 students working on geometry problems.
Students were randomly assigned to four conditions:
- worked examples with self-explanation prompts;
- fading with self-explanation prompts;
- fading without those prompts;
- problem solving alone.
The intervention was delivered through ASSISTments across three days between pre- and post-testing.
The fading condition produced the largest reported pre-to-post effect sizes.
But an especially important result was that prior knowledge moderated the effect of fading, whereas measured working memory did not.
That matters for MathematicsOS.
It suggests we should avoid turning a general cognitive construct such as “working-memory capacity” into a simple classroom identity.
Instead, instructional support should remain sensitive to what the learner already knows about the actual mathematical structure.
This moves the design from:
low ability → give more help
towards:
insufficient current schema → temporarily expose more of the route
and then:
growing schema → remove information and require reconstruction
That is a much safer instructional architecture.
5. Model → Complete → Fade → Execute
The master runtime already contains the correct basic architecture:
Use correct worked examples, narrate decision points rather than every trivial thought, move into partially completed examples, and then require independent execution.
We can now state this more precisely.
Stage 1 — MODEL
Show a complete valid solution.
But do not merely display calculations.
Expose the important mathematical decisions.
For example:
Solve
[
3(2x-5)=4x+7
]
A useful worked solution is:
[
6x-15=4x+7
]
[
2x-15=7
]
[
2x=22
]
[
x=11
]
The instructional value is not the four lines themselves.
It is identifying the invariants:
- expansion preserves equivalence;
- the same operation must preserve equality;
- like terms can then be collected;
- the resulting value should satisfy the original equation.
A learner who merely copies these lines has observed motion.
A learner who sees why each transition is legal has begun constructing structure.
Stage 2 — COMPLETE
Now remove part of the solution.
For example:
[
3(2x-5)=4x+7
]
[
6x-15=4x+7
]
[
___________
]
[
2x=22
]
[
x=11
]
The learner must reconstruct the missing state.
The question changes from:
“Can you imitate this?”
to:
“Can you restore the route?”
Stage 3 — FADE
Remove more information.
Perhaps only the first transformation remains.
Or provide the problem and the final answer but no middle states.
The learner now carries increasing responsibility for the mathematical transition.
Support is being removed because the target capability is appearing.
Stage 4 — EXECUTE
The student receives a structurally related question:
[
5(2x+1)=3x+26
]
No solution is provided.
No missing boxes.
No hint chain.
The student must now reconstruct the route independently.
Only here do we obtain meaningful evidence that the mathematical operation is beginning to belong to the learner rather than the worksheet.
6. Then Something Else Must Happen: Variation
Independent repetition of one surface form is still not enough.
A student may learn:
“Whenever I see brackets followed by x, do these four familiar steps.”
That can produce convincing worksheet performance while leaving the underlying capability fragile.
So the next questions must change.
Change:
- coefficients;
- signs;
- variable position;
- number of brackets;
- representation;
- wording;
- direction of the problem;
- irrelevant surface information;
- connection with another topic.
This is where examples become a bridge to structure, rather than templates to imitate.
A major 2026 study by Cao and Carvalho sharpened this point.
Across two experiments, when participants first received explicit instruction, retrieval practice subsequently produced better generalisation than continued worked-example study.
When that initial instruction was removed, however, repeated worked examples outperformed repeated retrieval practice; varied retrieval practice could close the difference.
The study used controlled rule-learning tasks with adults, so it should not be transported directly into a Singapore secondary classroom as an effect-size prediction.
But architecturally it gives us an important result:
Once the learner possesses enough structure, continuing to show the structure may become less useful than requiring the learner to retrieve and use it.
That gives MATHCIV-013 its correct stopping rule.
7. Worked Examples Must Have an Exit Condition
The teaching question is therefore not simply:
When should we use worked examples?
It is also:
When should we stop using them?
A reasonable exit test is whether the learner can:
- identify the relevant structure;
- explain the critical decision;
- complete a partially faded example;
- solve a matched problem independently;
- solve a varied problem independently;
- verify the solution;
- reproduce the capability after a delay.
If these succeed consistently, additional full worked examples may be redundant.
The instructional system should move forward.
MODEL
“I can understand a correct route.”
↓
COMPLETE
“I can restore parts of it.”
↓
FADE
“I can reconstruct most of it.”
↓
EXECUTE
“I can build it myself.”
↓
VARY
“I recognise the structure despite surface change.”
↓
RETRIEVE
“I can recover the route without seeing it.”
↓
TRANSFER
“I can adapt it.”
↓
VERIFY
“I can tell whether my result makes mathematical sense.”
That is much closer to capability than copying ten model solutions.
8. Why Self-Explanation Needs More Precision Than “Explain Your Thinking”
Worked examples are often paired with prompts such as:
“Explain why this step works.”
That sounds universally useful.
The evidence is more complicated.
The 2023 mathematics meta-analysis found that conditions containing self-explanation prompts did not produce a simple additional advantage; within that synthesis, the moderator went in the opposite direction compared with worked-example conditions without such prompts.
This does not prove that explaining mathematics is harmful.
It demonstrates why instructional labels are too crude.
A useful prompt should target a mathematical relation:
- Why is this transformation equivalent?
- What condition must remain true?
- Why was this representation selected?
- Which previous line justifies this one?
- How could you verify this answer?
A vague prompt—
“Explain your thinking”
—can add demand without exposing useful structure.
So explanation should also pass the receiver test.
9. Correct Examples and Erroneous Examples Are Different Tools
The master makes another important distinction.
The 2023 mathematics meta-analysis found stronger effects for correct examples alone than for designs involving incorrect examples in the comparisons it synthesised.
However, a newer 2025 meta-analysis specifically devoted to erroneous examples synthesised 42 papers and 177 effect sizes and reported a small positive overall effect, g = 0.136, with the way learners were asked to explain errors acting as an important moderator.
These findings do not require us to choose one and discard the other.
They concern different research questions and comparison structures.
The practical distinction is:
Correct worked example
Best treated primarily as an acquisition/model interface.
Erroneous example
Can become an error-detection and discrimination interface when the learner has enough knowledge to analyse what went wrong.
The second should not automatically replace the first.
A novice who is still unsure what a correct equation transformation looks like may not benefit from being flooded with deliberately incorrect ones.
This topic belongs more fully in MATHCIV-018: Errors as Sensors.
10. Worked Examples in Additional Mathematics
The same architecture becomes particularly useful as mathematical dependency rises.
Consider differentiation.
A student may see:
[
y=(3x^2+1)^5
]
and know individually that:
- powers can be differentiated;
- composite expressions contain an inner function;
- the chain rule exists.
Yet still fail to coordinate them.
A useful worked example should expose the dependency:
[
y=(3x^2+1)^5
]
Outer derivative:
[
5(3x^2+1)^4
]
Inner derivative:
[
6x
]
Therefore:
[
\frac{dy}{dx}
30x(3x^2+1)^4
]
But the next task should already begin fading.
Then:
[
y=(2x^3-4)^6
]
should require more learner-generated work.
Then:
[
y=\sqrt{5x^2+3}
]
changes the surface form.
Eventually the student should identify the composite structure without being told:
“Use chain rule.”
That transition—from seeing the route to selecting the route—is the actual educational objective.
11. The Biggest Failure Mode: Worked Examples That Manufacture Familiarity
A student can become extremely comfortable watching mathematics.
That comfort can be mistaken for learning.
Solutions look obvious after someone else has created them.
This is hindsight familiarity.
The learner recognises each line but cannot generate the next one when the model disappears.
That creates a dangerous measurement error:
recognition is recorded as capability.
MathematicsOS prevents this by requiring independent reconstruction.
The master explicitly states that worked examples support acquisition but do not prove transfer, and that examples should not remain indefinitely.
Therefore, after sufficient modelling:
close the example.
Give a fresh problem.
Change something.
Return later.
Remove the tutor.
Remove the answer key.
If AI was involved, remove the AI.
Then observe what remains.
12. The MathematicsOS Worked-Example Protocol
A practical implementation can therefore use this control sequence.
WE-1 — SENSE
Determine what the learner currently knows.
Do not infer this from one overall mark.
Inspect actual working.
WE-2 — LOCATE
Find the mathematical transition the learner cannot reliably generate.
WE-3 — MODEL
Provide the smallest sufficient correct worked route.
Emphasise mathematical decisions and invariants.
WE-4 — COMPARE
Ask the learner to distinguish essential structure from incidental numbers or wording.
WE-5 — COMPLETE
Remove one or more steps.
Require reconstruction.
WE-6 — FADE
Progressively remove support as evidence improves.
WE-7 — EXECUTE
Require a fresh independent solution.
WE-8 — VARY
Change surface form or representation.
WE-9 — RETRIEVE
Return after a delay without reopening the example.
WE-10 — VERIFY
Require an independent check.
WE-11 — WRITE BACK
Only update capability if performance survives without the worked solution.
That preserves the project’s fundamental distinction:
Delivery ≠ Receipt ≠ Retention ≠ Transfer ≠ Independent Capability.
13. What Parents Should Look For
A parent does not need to ask:
“Does the tuition centre use worked examples?”
Almost every mathematics classroom does.
A better set of questions is:
What happens after the example?
Does the student eventually have to solve without seeing it?
Are examples gradually removed?
Does the teacher inspect the learner’s first invalid step?
Are later questions varied?
Is the student tested again after a delay?
Can the student explain why the route works?
Can the student check an answer independently?
If the answer to those questions is yes, the worked example is functioning as scaffolding.
If the student continuously receives solutions whenever difficulty appears, the support may be concealing the very capability the lesson is supposed to build.
14. What Students Should Do With a Worked Example
Do not simply read down the page.
For every important transition, ask:
What changed?
Why was it allowed to change?
What stayed invariant?
Could I predict the next step before uncovering it?
What would change if the numbers changed?
How would I know the final answer was wrong?
Then close the example.
Reconstruct it.
Then solve another question.
That is how a worked solution stops belonging to the book and begins becoming part of your usable mathematical capability.
15. The CivilisationOS Connection
There is a larger reason worked examples matter.
Civilisations do not preserve capability merely by storing finished answers.
They preserve methods by making procedures inspectable and teachable.
A mathematical derivation, algorithm, engineering calculation or proof can function as an external record of reasoning.
But a stored procedure becomes living capability only when another receiver can:
read it → reconstruct it → use it → adapt it → verify it → transmit it again.
That is why worked examples matter beyond one worksheet.
They are a small educational instance of a much larger civilisation process:
capability regeneration.
But the boundary remains important.
A solved page is stored information.
It becomes learner capability only when the next receiver can operate without the page.
Final Answer
Worked examples work best not because mathematics should be made passive, but because guided observation can prepare independent action.
For a learner who does not yet possess a reliable mathematical schema, a correct route can reduce unnecessary search and expose the structure that needs to be learned.
Then the architecture must change.
Model.
Complete.
Fade.
Execute.
Vary.
Retrieve.
Transfer.
Verify.
The newest 2026 evidence strengthens precisely this receiver-dependent interpretation: faded guidance can outperform persistent full support in developing learners, and once sufficient instruction has established the relevant structure, active retrieval may become more valuable than continued example viewing.
So the principle is not:
Give students more answers.
It is:
Show enough of a correct route for the learner to construct the structure—then progressively give the mathematics back to the learner.
That is the difference between seeing mathematics being done and becoming able to do mathematics independently.
Research boundary
The study findings and the MathematicsOS interpretation are related but not identical. The 2023 meta-analysis provides broad evidence across mathematics studies; the 2026 fading study concerned 114 Grade 6 students and geometry; the 2026 variability study used controlled rule-learning tasks with adults. None independently proves that one fixed worked-example sequence is optimal for every Singapore student, topic or stage.
Evidence status: Strong overall evidence for worked examples as an instructional strategy; newer evidence strengthens fading and receiver-state adaptation; exact dosage, fading rate and transfer design remain context-dependent.
Next article: MATHCIV-014 — Retrieval Practice in Mathematics: Reconstruct, Do Not Merely Re-read.