Interleaving (mixing problem types vs blocking)

Verdict

Interleaving — mixing different problem types or categories within a practice set instead of practicing one type at a time (blocking) — is real but much narrower than its popular reputation. It helps in two specific situations: (1) learning to tell confusable categories apart (the mechanism is discriminative contrast — you learn which type you're looking at by juxtaposing it with others), and (2) mixed math-problem practice, where students must also learn which strategy a problem calls for. It is null for expository text and actually negative for memorizing discrete facts or vocabulary (block those). And the celebrated math effect is substantially confounded with spacing — much of interleaving's benefit is that mixing automatically spaces each type out.

What the evidence shows

The math evidence is genuinely strong in design terms — Rohrer 2020 is a preregistered cluster RCT that meets What Works Clearinghouse standards without reservations. But three deflations keep this at moderate: (1) the outcome is a proximal, researcher-designed test of the same problem kinds practiced, so a standardized equivalent is likely ~0.4; (2) at full field scale (the 62-school NewGlobe trial) the effect is only 0.29; and (3) the benefit confounds discrimination + spacing + retrieval — the authors' own per-problem-kind gradient tracks how much each kind was spaced.

The material-dependence is the headline practical fact: interleaving helps when categories are confusable and you must learn to discriminate (paintings 0.67; math strategy-selection), does nothing for expository text, and backfires for rote verbal learning (words −0.39, blocking wins). For category learning specifically, the benefit transfers to genuinely novel items as strongly as to studied ones (Firth 2021) — so there it's more than teaching-to-the-test.

Hereditarian-lens assessment

Risk: low. Domain discrimination-skill outcomes, randomized designs, no g. Effects are somewhat larger for younger/novice learners; like worked examples, an expertise interaction is plausible (experts already discriminate), but this is a prior-knowledge interaction, not an ability/g one, and it doesn't widen heritable gaps.

Boundaries & what critics say

• Interleaving vs spacing confound: in category learning the benefit needs immediate succession of contrasting items (not spacing), but in the classroom math studies it's bundled with spacing — so "interleaving works" partly means "spacing works."
• Material matters more than the technique: null for text, negative for discrete facts/vocab.
• New skills need an initial block: don't interleave from the very first exposure; Rohrer and the null (Ostrow) both point to a small initial block plus adequate delay.
• At-scale shrinkage: the honest real-world math effect is closer to 0.3 than 0.8.

Practical guidance

• Interleave practice problems in math (and similar problem-solving subjects): mixed problem sets and cumulative reviews so students must choose the method, not just execute it.
• Interleave to teach discrimination wherever students confuse categories (e.g. distinguishing similar concepts, species, artistic styles, grammatical cases).
• Block, don't interleave, for memorizing discrete facts and vocabulary.
• Give a short initial block when a skill is brand-new, then mix.
• Recognize that much of the benefit is spacing — the two travel together.

Open questions

• The standardized/far-transfer math effect is unmeasured (all classroom effects use aligned tests).
• How much of the math benefit is discrimination vs embedded spacing vs retrieval is not cleanly decomposed.