Why mathematics is hard — and what actually helps.

Educational research is harder than it looks. Effect sizes are small, replication is patchy, you can't double-blind a classroom, and the same dataset usually supports several different stories depending on what you came in believing. The "Math Wars" are largely a long conversation between philosophical positions — traditionalists who emphasise procedural fluency and direct instruction, and reformers who emphasise conceptual understanding, problem-solving, and equity. Most of the actual research finds both matter, but advocates on each side reliably read the same studies as supporting their priors.

Take Jo Boaler — a Stanford professor, founder of YouCubed, prominent advocate of growth mindset, and one of the loudest voices for de-tracking and against timed maths tests. Her empirical claims are contested. The "Railside" study underpinning much of her advocacy was the subject of a sustained methodological critique by James Milgram and Wayne Bishop, who argued that the data, properly read, do not support her conclusions. Stanford reviewed the dispute and described it as "academic debate" rather than research misconduct, but the core empirical questions remain genuinely open.

Our position: be cautious about Boaler's strongest empirical claims, but do use her tasks. The rich, low-floor / high-ceiling problems on YouCubed are excellent classroom material — accessible to a struggling student, deep enough to challenge the strongest. Good problems and contested empirical claims can coexist; we take what works in the classroom and stay sceptical about the slogans.

This is how we read most research: provisionally, with attention to who funded it, who replicated it, and whether the effect size justifies the policy claim. We follow effect sizes, not slogans, and update when the evidence does.

The cognitive-science techniques teachers reach for most often — retrieval practice (testing yourself instead of re-reading), spaced practice (revisiting topics over weeks rather than cramming), and interleaving (mixing topics rather than blocking by topic) — sit under Robert Bjork's umbrella term "desirable difficulties." They make the immediate experience of practice harder, but they produce better retention weeks and months later. The catch: cognitive load theory (Sweller) says working memory is finite, so if we pile too much difficulty on at once, learning collapses. The practitioner's job is the sweet spot — productive struggle, not overwhelm.

Anxiety is the lever that closes that sweet spot fastest. An anxious student is already loading working memory with worry before any maths arrives — so a task that would be a desirable difficulty for a calm student becomes overload for an anxious one. That is why we treat managing classroom culture (where getting things wrong feels ordinary, not catastrophic) as a precondition for desirable difficulty to work — not as a competing priority. The two go together.

The recent meta-analyses are usefully nuanced. A 2025 mathematics-specific meta-analysis (Educational Psychology Review) finds spaced practice yields g ≈ 0.28 over massed practice — small-to-medium and consistent — and practice testing yields g ≈ 0.50 in lab settings, dropping to g ≈ 0.33 in classroom settings. Format matters: multiple-choice retrieval g ≈ 0.70, short-answer g ≈ 0.48. A 2024 STEM paper provocatively titled "is the glass half full or half empty?" makes the case that effect sizes in real classrooms are routinely smaller than the lab-derived headline figures imply. Carl Hendrick's 2025 essay "Making Retrieval Practice Actually Work" lays out the seven implementation conditions that separate effective retrieval practice from a quiz-bombing approximation — keeping stakes low to manage anxiety is high on his list, which is why we treat anxiety management and desirable difficulty as inseparable rather than competing. We use these techniques because the evidence supports them — but we don't expect them to produce a Bloom-style 2σ shift in a single term.

A caveat on cognitive load theory itself: measurement is subjective (post-task self-rating scales), and the classical separation of intrinsic, extraneous, and germane load is itself contested — some researchers argue these components can't be cleanly added, and the theory has been reformulated more than once in response. We use it as a practical frame for thinking about task design, not as an exact physics of working memory.