Worked examples (studying solutions vs solving problems)

Verdict

For a novice learning well-structured content — how to solve a class of algebra, geometry, or physics problems — studying worked-out solutions is more efficient than being handed problems to struggle through, because unguided problem-solving overloads a novice's working memory before a schema has formed. The best current estimate (math meta) is g≈0.48 on aligned tests (~half that on a standardized scale). But this is the textbook case of an aptitude-treatment interaction: the same worked examples that help novices hurt more knowledgeable learners (the expertise reversal effect) — once you have the schema, the worked-out steps are redundant and add load. So the finding is not "use worked examples" but "use worked examples for novices and fade them as competence grows."

Note this sits in productive tension with productive failure: worked examples favor showing the method first, while productive-failure evidence shows that letting older learners grapple before instruction can deepen conceptual understanding. The reconciliation is the same variable everywhere — match the amount of guidance to the learner's prior knowledge and the goal (fluency/procedural → worked examples first; conceptual/transfer with older learners → some struggle first).

What the evidence shows

The expertise reversal is the load-bearing pattern and it replicates in authentic classroom subjects (Bokosmaty's high-school geometry): novices benefit from fully-guided worked examples, but for students who have partly learned the material the extra step-by-step guidance becomes redundant and depresses performance. Mechanistically this is the redundancy effect / element interactivity relative to the learner's schemas (Kalyuga 2003; Chen, Kalyuga & Sweller 2017).

One live inconsistency worth flagging: adding self-explanation prompts helps in lab studies (Atkinson/Renkl) but negatively moderated the effect in the aggregate math meta (Barbieri) — so prompt quality matters, and "add self-explanation" is not a free win.

Hereditarian-lens assessment

Risk: low. Domain problem-solving outcomes; no g. The genetics review made one clarification that matters for a school builder: the expertise reversal is a prior-knowledge (schema) interaction, not an ability/g one — "high prior knowledge" is not "high ability." So worked examples don't widen heritable ability gaps; if anything they give an ordinal working-memory-load benefit to novice learners (narrowing gaps among novices) while only the already-competent should be moved off them. The direction of the prior-knowledge interaction across studies is mixed, so treat "fade with competence" as the robust rule, not a precise dosing formula.

Lab-to-classroom & durability

The worked-example literature is more classroom-embedded than retrieval/spacing (Barbieri: ~22 teacher-run studies in authentic algebra/geometry), which is a point in its favor. But outcomes are researcher-designed aligned accuracy, mostly immediate — delayed-retention evidence is thin, and a standardized/durable estimate is likely ~half the headline. Far transfer is weak unless deliberately engineered (fading + prompts).

Boundaries & what critics say

• Expertise reversal: the biggest boundary — worked examples are a novice tool; keep using them on competent students and you waste time or do harm.
• Standardized/delayed magnitude is likely ~half the g≈0.48 headline; durability is under-tested.
• Self-explanation prompts are quality-dependent, not automatically beneficial.
• Applies best to well-structured procedural domains (math, science); less clear for ill-structured tasks like essay writing.

Practical guidance

• Start novices with worked examples, not cold problem-solving, for well-structured procedural topics.
• Fade systematically: worked example → completion problem (some steps removed) → full problem, as competence grows. Don't keep fully-worked examples in front of students who've got it.
• Correct worked examples are the default; studying incorrect examples can help but is a targeted technique, not the baseline.
• For conceptual/transfer goals with older students, consider some struggle before instruction (productive failure) rather than worked examples first.

Open questions

• The precise prior-knowledge dosing (when exactly to fade) varies by study; only the qualitative "fade with competence" rule is robust.
• Why self-explanation prompts help in the lab but not in the aggregate math meta is unresolved.