How Studying Works | The Example–Problem Ratio — Why Facts and Generalisable Skills Need Different Mixes of Study and Practice

HSW-0192 · How Studying Works

How many worked examples should you study before you stop looking and start solving?

One example for every problem?

One example, then ten problems?

No examples at all—just retrieve and practise?

The tempting answer is to search for one optimal ratio.

Current learning research suggests that is the wrong question.

The example–problem ratio is the balance between studying complete worked solutions and attempting problems independently. Its best value depends on what kind of knowledge is being learned, what the learner already knows, and which learning process the task needs to trigger.

A fact that must be retrieved verbatim is not the same learning problem as a general skill that must be induced across changing examples. One may benefit from repeated retrieval attempts. The other may require enough worked structure for the learner to discover what stays invariant while the surface changes.

This article owns that allocation problem. How Worked Examples Work owns the general mechanism of learning from complete expert solutions. Example–Problem Pairs owns the local “see one, solve one” bridge. Worked Example Fading owns the gradual removal of support. Completion Problems owns partly worked transitions. Here the narrower question is: how much of the learning budget should be allocated to examples versus problems, and why can the answer change with the knowledge content?

The 50-Second Read

• There is no universal 50:50 rule. The useful mix depends on the knowledge component and the learner.
• Verbatim facts and generalisable skills require different learning processes. Facts depend heavily on memory and fluency; variable skills require induction, pattern extraction and application across changing conditions.
• A 2025 experiment found a crossover. Pure practice testing produced better learning for verbatim fact content, while example-integrated practice produced better learning for general skill content in the researchers’ materials.
• Worked examples reduce search and expose structure. But too many examples can leave learners good at following and weak at generating.
• Practice creates production evidence. But too much unsupported practice can spend time on guessing, repeated errors or inefficient search.
• The ratio should move. Novices often need more structure; later learning should shift toward independent problems, mixed discrimination and transfer.
• The practical rule: classify the knowledge first, then allocate examples and attempts according to the learning process needed.

1. “More Practice” and “More Examples” Can Both Be Good Advice

Learning science contains two strong traditions.

The retrieval-practice tradition says learners often remember more when they actively retrieve rather than repeatedly restudy.

The worked-example tradition says novices often learn complex procedures more efficiently when complete solutions reduce unproductive problem-solving search.

At first glance the recommendations seem to compete.

• If retrieval is powerful, should examples be replaced by more attempts?
• If worked examples are powerful, should some practice problems be replaced by examples?

The deeper answer is that both can be correct because they can be serving different kinds of knowledge.

2. What the 2025 Study Added

Napol Rachatasumrit, Kenneth Koedinger and Paulo Carvalho published an open-access study in the International Journal of Artificial Intelligence in Education in August 2025 titled “Evidence and Theory for why the Best Example-Problem Ratio To Optimize Learning Gain Depends on Knowledge Content.” See Rachatasumrit, Koedinger and Carvalho, 2025.

The study directly compared two instructional treatments across two kinds of content.

• Practice testing: learners attempted retrieval or responses.
• Example-integrated practice: worked examples were interspersed with problem attempts.

The authors reported a crossover interaction.

For verbatim facts—content where the cue and correct response were relatively stable—pure practice testing led to better learning in their experiment.

For general skills—content that had to be applied across variable conditions—example-integrated practice produced better learning.

They also built a computational account using the Adaptive Learning architecture to model why the crossover could arise from different learning processes.

The educational significance is not that every fact should receive zero examples or every skill should receive a 1:1 alternation. It is that the best instructional event may depend on the type of knowledge being changed.

3. Facts Ask Memory to Strengthen a Stable Mapping

Consider:

Capital of Japan → Tokyo.

The correct response is stable.

Studying another worked example of someone answering “Tokyo” adds little structure after the mapping is understood.

The learner needs to make the cue reliably retrieve the response.

That is why repeated retrieval attempts can be efficient for fact-like knowledge once meaning is established.

4. General Skills Ask the Learner to Infer What Stays the Same

Now consider finding the area of a triangle.

The next problem can change:

• base;
• height;
• orientation;
• diagram;
• irrelevant measurements;
• units;
• which dimension must be inferred.

The learner cannot simply memorise one stimulus-response pair.

The learner must infer a general relation that applies across varied instances.

Worked examples can support this inductive job because they keep the successful solution structure visible while the learner compares cases.

5. This Is a Knowledge-Content Problem, Not a Method Popularity Contest

Students often ask which method is “best”:

• flashcards;
• practice questions;
• worked examples;
• videos;
• notes;
• teaching someone else.

A mature answer begins one level earlier:

What knowledge change does this target require?

If the target is stable recall, another passive example may have low marginal value.

If the target is induction of a general procedure from changing cases, repeated unsupported attempts may have low marginal value.

The method should match the learning process.

6. Worked Examples Buy Structure

A worked example does several useful things at once.

• shows a valid route;
• reduces blind search;
• makes intermediate states visible;
• exposes how representations change;
• lets the learner compare decisions across cases;
• preserves cognitive capacity for understanding rather than guessing.

That is why examples can be powerful for unfamiliar multi-step skills.

But the example’s value falls if the learner merely reads the surface and never extracts the transferable structure.

7. Practice Problems Buy Evidence

A practice problem forces the learner to generate.

Now the study system can observe:

• whether the method can be selected;
• whether prerequisites are retrievable;
• whether execution is accurate;
• where the first weak step appears;
• whether confidence matches performance;
• whether the skill works without the example visible.

Practice therefore does two jobs: it can strengthen learning, and it creates diagnostic evidence.

But a problem attempted before the learner has a workable model can produce expensive trial-and-error rather than useful practice.

8. The Ratio Is Not Only About Count

“Four examples and four problems” sounds precise.

But count can hide important differences.

• One example may take two minutes; one problem may take fifteen.
• One example may reveal an entire method; another may differ only cosmetically.
• One problem may test one step; another may integrate six skills.
• A learner may process an example actively or skim it passively.

A better ratio can be measured in learning opportunities, not only items.

9. Example–Problem Ratio vs Example–Problem Pairs

Example–Problem Pairs owns a specific local sequence: study a worked solution and immediately solve a structurally similar problem.

The ratio problem is broader.

A learner may need:

• zero examples after a fact is understood;
• several contrasting examples before one difficult problem;
• one example for every problem at the beginning;
• one example after every two failed attempts;
• almost no examples near final independent performance.

The ratio is the allocation policy across a learning phase; the pair is one possible unit inside that policy.

10. Example–Problem Ratio vs Worked Example Fading

Worked Example Fading owns the systematic removal of solution support as expertise increases.

Fading often changes the ratio indirectly: fewer full examples, more partial completions, then more independent problems.

The ratio article asks the allocation question before and alongside fading: what kind of content deserves more examples or more practice at this stage?

11. Example–Problem Ratio vs Completion Problems

Completion Problems are neither fully worked examples nor fully independent problems.

They are useful because the example–problem decision is not truly binary.

A good allocation system can use a continuum:

full example → completion problem → guided problem → independent problem → mixed transfer.

12. Prior Knowledge Moves the Ratio

The same knowledge content can require a different ratio for different learners.

A novice may need examples because the problem contains too many unknown decisions.

An experienced learner may gain little from another full solution and more from retrieval, method selection and transfer.

This is one reason static worksheets are inefficient. They assume every learner needs the same mixture at the same time.

13. What New 2026 Evidence Adds

A 2026 Educational Psychology Review study, “Striking the Balance: How Variability Shapes Retrieval Practice and Worked Examples for Transfer Learning,” compared worked examples and retrieval practice under repeated and varied materials. In one experiment, retrieval practice outperformed worked examples even under varied practice, contrary to the authors’ predicted interaction. The study also noted that retrieval practice took more learning time and that prior explicit instruction may have raised participants’ initial knowledge enough to reduce a worked-example advantage.

See the 2026 study.

This is useful precisely because it prevents a simplistic reading of the 2025 crossover.

Content type matters. Prior knowledge matters. Whether rules were already explicitly taught matters. Learning time matters. The final task matters.

The ratio is conditional, not a new magic number.

14. Mathematics: Separate Fact Retrieval From Method Induction

A mathematics chapter contains multiple knowledge types.

Target	Likely learning job	Possible ratio direction
Formula recall	Stable cue-response memory	More retrieval
Recognising when formula applies	Conditional discrimination	Examples + mixed problems
Multi-step method	Procedure induction and execution	More examples early, then fade
Exam transfer	Independent selection under changed surface	Mostly problems

One chapter can therefore need several ratios at once.

15. English: Model Paragraphs and Independent Writing Are Different Learning Events

For writing, examples can reveal quality that is difficult to infer from abstract rules alone.

Students may compare model paragraphs to learn:

• how evidence is integrated;
• how claims are qualified;
• how transitions carry argument;
• how sentence form changes emphasis.

But writing skill cannot be acquired through examples alone.

The ratio must eventually move toward planning, drafting, revising and producing new responses under independent conditions.

16. Science: Facts, Models and Inquiry Need Different Mixes

Science includes:

• terms and symbols;
• causal mechanisms;
• graph interpretation;
• experimental design;
• evidence evaluation;
• quantitative procedures.

A terminology target can use heavy retrieval after meaning is understood.

Experimental design may need multiple examples showing different fair-test structures before students can independently design a new investigation.

Do not make one practice philosophy govern every knowledge component in the subject.

17. The Example Budget

Think of examples as expensive but information-rich assets.

An example deserves time when it adds structure the learner cannot yet generate.

Ask:

• Does this example reveal a new subgoal?
• Does it show a boundary condition?
• Does it expose a different representation?
• Does it contrast with a misleading near-case?
• Does it repair an error that repeated practice is not fixing?

If the answer is no, another example may have low marginal value.

18. The Practice Budget

A practice problem deserves time when it produces useful evidence or strengthens production.

Ask:

• Does it require unaided retrieval?
• Does it force method selection?
• Does it vary a feature that should not change the underlying rule?
• Does it discriminate between neighbouring methods?
• Will the result change what the learner does next?

If not, more volume can become repetition without information.

19. The Dynamic-Ratio Rule

A practical sequence for a new generalisable skill is:

1. Orient: one clear worked example.
2. Compare: one or two examples that vary important features.
3. Attempt: solve a closely related problem.
4. Repair: return to an example or completion problem if the first weak link is structural.
5. Fade: reduce worked support as successful decisions become stable.
6. Mix: add neighbouring methods so the learner must discriminate.
7. Transfer: use unfamiliar problems with no example available.

The ratio changes because the learner changes.

20. The Center-to-Edge Route

Allocate examples and problems from the centre outward.

1. Centre: classify the knowledge component—stable fact, concept, conditional rule, variable skill or integrated performance.
2. First ring: estimate prior knowledge.
3. Second ring: choose enough examples to make the hidden structure visible.
4. Third ring: add enough problems to prove independent production.
5. Edge: remove examples and test mixed, delayed and unfamiliar transfer.

This turns the ratio from a fixed recipe into a controlled transition.

21. The School Route: Homework Should Have a Learning Job

Ten practice questions are not automatically better than five questions plus two worked comparisons.

Five examples are not automatically better than one example plus four retrieval attempts.

Before assigning volume, ask what must change:

• memory strength;
• fluency;
• schema formation;
• method discrimination;
• execution accuracy;
• transfer.

The homework form should follow the learning job.

22. The Systems Route: Different Workloads Need Different Pipelines

In engineering, one processing pipeline is rarely optimal for every input type.

Stable records may need fast lookup and repeated access. Variable cases may need modelling, comparison and transformation.

Learning is similar in one limited sense: different knowledge structures recruit different learning operations.

The analogy helps explain why a universal example–problem ratio is unlikely.

23. The Financial Route: Allocate Study Time by Marginal Return

Each additional example and each additional problem has a marginal return.

At the beginning of a complex method, another example may be worth more than another blind attempt.

Later, the eighth similar worked example may add almost nothing while one independent transfer problem produces high-value evidence.

Allocate the next ten minutes to whichever activity changes capability or uncertainty most.

24. The Learning Route: Diagnose Why the Problem Failed

After an incorrect problem, do not automatically assign another problem.

Classify the failure.

Failure	Likely next move
Forgot fact	Retrieval + feedback
Could not see structure	Worked/contrasting example
Knew structure, execution error	Targeted practice
Confused two methods	Mixed discrimination
Could follow example but not start	Completion → independent problem
Works only on familiar surface	Variable transfer problems

The ratio should react to the diagnosis.

25. The Education Route: Teach Students to Choose Their Own Ratio

Independent learners should eventually decide when to inspect an example and when to stop looking.

That requires metacognitive control.

• If I do not understand the method, find one high-quality example.
• If I understand but cannot retrieve, stop reading and test.
• If I repeatedly fail at the same structural step, compare examples.
• If I can solve familiar problems, remove examples and vary the surface.
• If I can solve only one method at a time, mix neighbouring methods.

The long-term goal is not permanent instructor control of the ratio. It is learner control based on evidence.

26. The Training Route: Ratio A/B Test

For two comparable subtopics, deliberately try different mixes.

1. Subtopic A: more example-heavy early practice.
2. Subtopic B: more retrieval/problem-heavy early practice.
3. Keep total time approximately comparable.
4. Test both after a delay.
5. Include one familiar item and one transfer item.
6. Compare not only score but time, error type and independence.
7. Adjust the future ratio from evidence.

This is not a clean scientific experiment unless many other variables are controlled. It is a learner audit designed to improve allocation, not prove a universal theory.

27. The Improvement Route: Track Ratio Drift

Students tend to drift toward comfort.

• Some keep reading examples because examples feel safe.
• Some keep grinding problems because visible volume feels productive.

Once a week, ask:

• Am I still studying examples after I can independently perform?
• Am I attempting problems without enough model of the method?
• Is my error rate falling?
• Can I choose the method without labels?
• Can I transfer beyond the practised surface?

The answers tell you whether the ratio should move.

28. The World Route: Professionals Alternate Observation and Performance

Apprenticeship in many fields alternates between observing expert performance and carrying progressively more responsibility.

A junior clinician studies cases and then makes supervised decisions. A trainee engineer studies designs and then produces analyses. A musician listens and watches, then performs.

The exact ratio differs by domain and risk, but the architecture is familiar:

see enough structure to act intelligently → act enough to prove the structure is yours.

29. A Delayed Independence Check

After an example-heavy learning session, wait a day and test without examples.

1. Retrieve the governing rule or method.
2. Solve one standard problem.
3. Solve one varied problem.
4. Explain why the method applies.
5. Identify one case where it would not apply.
6. Record whether you needed to reopen an example.

If the learner needs the example to begin, the ratio has not yet produced independence.

30. Parent and Tutor Guide: Do Not Confuse More Questions With Better Practice

When a child struggles, parents and tutors often respond with more questions.

First ask what kind of failure is happening.

• If the child cannot remember a stable fact, more worked solutions may be unnecessary.
• If the child cannot see how a multi-step method is organised, another unsupported problem may be wasteful.
• If the child can follow a model but cannot start independently, use partial support and fade it.
• If the child is already accurate on familiar questions, stop adding similar examples and test transfer.

The right question is not “How many questions did you do?” but “What did the next example or problem change?”

31. What Not to Do

• Do not turn the 2025 crossover into a universal rule that facts never need examples and skills always need them.
• Do not assume 1:1 is a magic ratio.
• Do not leave novices in unsupported search when the method is not yet represented.
• Do not let examples replace eventual independent production.
• Do not count items without considering time and information value.
• Do not ignore prior knowledge.
• Do not use worked-example evidence to eliminate retrieval practice, or retrieval-practice evidence to eliminate examples.
• Do not keep the ratio fixed while the learner changes.

32. Evidence Boundary

The 2025 study provides direct experimental evidence for a content-treatment interaction in the researchers’ fact and skill materials, supported by computational modelling. It is valuable because it gives a principled explanation for why literatures favouring retrieval practice and worked examples can both be right under different knowledge conditions.

It does not establish a universal numeric ratio for all subjects, ages or learners. The classification of knowledge components, prior instruction, task variability, problem complexity, feedback and final-test demands can all alter outcomes. The 2026 variability study further shows that retrieval practice can remain advantageous even in conditions where one might expect varied worked examples to win.

The responsible educational conclusion is conditional: classify the knowledge and learner state, choose a plausible example–problem mix, and verify delayed independent performance.

33. Return: Study the Route Until You Can Carry It—Then Walk Without It

A worked example is borrowed competence.

A practice problem is a request to carry the competence yourself.

Good studying knows when each is worth the time.

For stable facts, retrieve. For unfamiliar general skills, study enough structure to induce the rule. Then shift the ratio toward independent problems, mixed decisions and transfer until the examples are no longer carrying the learner.

Continue through Worked Examples, Example–Problem Pairs, Completion Problems, the How Studying Works Numbered Series Reading Index and the How X Works Hub.