top of page
Search

Math Fact Fluency Intervention: What the Record Shows

Writer: Andrea Chernin
Andrea Chernin
7 days ago
23 min read

Math Fact Fluency Intervention: What the Record Shows


You have ten minutes at the end of a session and a student who still counts on their fingers for 7 × 8. Everything you have read says the same thing: build automaticity, because automatic facts free up working memory for the harder work. It is one of the most repeated claims in math fact fluency intervention, and almost nobody who repeats it has read the sentence it comes from.


The sentence exists. It is in a federal practice guide. And the guide that published it labelled it, in its own technical appendix, as the opinion of the panel.


That is not a reason to stop doing fact work. The practice has better evidence behind it now than it did in 2009, and this post will walk through exactly what that evidence is attached to.


But the reason most often given for doing it is not the reason the record supports, and the two have drifted far enough apart that a practitioner quoting the familiar version is overstating what anyone has shown.


The research on fact practice is better than it was. The argument practitioners use to justify fact practice has not kept up with it.

The short version


Four things are true at once, and holding all four is the whole job:


  • "Fluency" and "automaticity" are different words for different things, and the two bodies that define them formally both define fluency without speed in it.

  • The working-memory argument is panel opinion, said so in print by the guide that made it, and its 2021 successor quietly rephrased it.

  • Timed fact activities have a Strong federal evidence rating — attached to a bundle of practices, not to a timer.

  • The measured effect shrinks the further the outcome gets from the facts themselves, from g = 0.89 on fact fluency measures to g = 0.25 on word problems.


Everything below is the sourcing for those four.

Three documents define fluency, and only one of them mentions speed


Start with the vocabulary, because the vocabulary is where this goes wrong first.


The 2021 IES practice guide on mathematics intervention (WWC 2021006) carries a glossary. Two entries matter:


Fluency – the ability to perform mathematics accurately and with ease.

Automaticity – the ability to produce answers quickly without much mental energy.

Read those twice. The federal definition of fluency contains no clock. "Accurately and with ease" is a statement about competence, not about seconds. The word "quickly" appears only in the definition of automaticity, which is a different entry for a different construct.


Now the National Council of Teachers of Mathematics position statement on procedural fluency, dated January 2023. Its full text opens by naming the problem directly: "Unfortunately, the term fluency continues to be (incorrectly) interpreted as remembering facts and applying standard algorithms or procedures."


NCTM's own definition: "the ability to apply procedures efficiently, flexibly, and accurately; to transfer procedures to different problems and contexts; to build or modify procedures from other procedures; and to recognize when one strategy or procedure is more appropriate to apply than another." Four components. Speed is not one of them — "efficiently" is about strategy choice, which the statement illustrates with 98 + 35 solved as 100 + 35 − 2.


So the two organisations that bother to define the word both define it without a stopwatch, and the construct that does carry a stopwatch has its own name.


The phrase "math fact fluency," the one every product is sold under and every practitioner uses, sits in neither glossary in the sense the market uses it. It has come to mean “fast accurate recall of single-digit combinations,” which is the definition of automaticity with a different label on it.


This matters practically, not just pedantically. When a parent asks whether their child is "fluent," and a program reports a fluency score, and a manual says fluency is a goal, those three uses may not be the same claim. It is the same problem this site has found in what "Orton-Gillingham based" actually means and in what "aligned" means against the IDA standards: a word everyone uses and nobody adjudicates.

The automaticity claim, and the one sentence it rests on


Here is the familiar argument in the field's usual phrasing: when facts are automatic, working memory is freed for the demands of multi-step problems.


The 2009 federal guide, Assisting Students Struggling with Mathematics: Response to Intervention (RtI) for Elementary and Middle Schools, is where this enters the practitioner literature with federal weight. Its Recommendation 6 opens: "Quick retrieval of basic arithmetic facts is critical for success in mathematics."


It then explains why, and the explanation is specific: weak retrieval "is likely to impede understanding of concepts students encounter with rational numbers since teachers and texts often assume automatic retrieval of facts such as 3 × 9 = and 11 – 7 = as they explain concepts such as equivalence and the commutative property."


Note what that paragraph does and does not say. It says weak retrieval is "likely to impede understanding," and it grounds that in a specific, checkable mechanism — teachers and textbooks assume the facts while explaining something else. It does not say that automatic facts free working memory. It does not use the phrase at all.


Then, in Appendix D of the same document, under the technical information for that recommendation, the panel writes this:


The panel acknowledges that the broader implication that general mathematics proficiency will improve when fact fluency improves is the opinion of the panel.

That is the federal government's own evidence appendix, telling you that the downstream claim — the one the whole practice is usually justified by — is the panel's judgement and not a finding. The recommendation itself was rated Moderate, on seven studies.


And the summary of evidence in the body is blunter still: "These positive effects did not, however, consistently reach statistical significance, and the findings cannot be extrapolated to areas of mathematics outside of fact fluency and operations."


"Cannot be extrapolated" is doing a great deal of work in a sentence almost nobody quotes.


The claim that fact fluency lifts general mathematics proficiency is in the federal record. It is in the federal record as an opinion, flagged as one, by the people who held it.

What the 2021 guide did to that sentence


The 2021 guide is an update of the 2009 one. It says so: "this guide is an update of the 2009 guide," narrower in scope, "focusing only on practices and principles underlying effective small-group interventions in grades K–6."


It also says, of this specific recommendation, that "Recommendation 5 (Word Problems), and Recommendation 6 (Timed Activities), were included in the 2009 guide. In contrast to how they were presented a decade ago, this updated guide frames them differently and highlights new evidence and ideas for intervention."


"Frames them differently" is accurate and uninformative. Here is what actually changed, measured by text search of both PDFs:


  • The phrase "fact fluency" appears 28 times in the 2009 guide and zero times in the 2021 guide.

  • "Memorization" appears nine times in 2009 and zero times in 2021.

  • "Timed activities" appears zero times in 2009 and 32 times in 2021.

  • The recommended duration fell from "about 10 minutes in each session," with a stated floor of "a minimum of 5 minutes a session," to timed activities that "last between 1 and 5 minutes."

  • The evidence rating rose from Moderate to Strong.


So the newer document raised the evidence rating and cut the recommended dose by half to nine-tenths, and it did both without saying so. A practitioner who learned "ten minutes a session" from the 2009 guide and never read the 2021 one is working from a figure the current federal guidance does not carry.


And the working-memory sentence? The 2021 guide's version reads:


Automatic retrieval gives students more mental energy to understand relatively complex mathematical tasks and execute multi-step mathematical procedures.

"More mental energy," not working memory. The phrase "working memory" appears three times in the entire 165-page guide — once in the front matter describing who the guide is for, and twice in the reference list. "Cognitive load" appears once, also in the reference list. The guide does not make the cognitive-architecture argument at all. It makes a softer, vaguer one, and it footnotes it.


This is the same pattern this site found in what the oral reading fluency table does not say: the number travels, the document's own hedging does not.

What the Strong rating is actually attached to


The 2021 recommendation is worth stating in full, because the wording is careful and the careful parts get dropped. The recommendation is: "Regularly include timed activities as one way to build students' fluency in mathematics."


One way. And immediately after:


These timed activities last between 1 and 5 minutes and are not the entire focus of the intervention. Instead, they are one component embedded within a multi-component intervention. Add timed activities to intervention once students have been working on a concept over many lessons. Do not use timed activities to introduce and teach mathematics concepts and operations.

The rating rests on 27 studies — 21 meeting WWC group design standards without reservations, six with reservations — covering 4,336 students in 403 schools, with interventions running from roughly eight days to 19 weeks. By any standard that is a solid evidence base, and it is the reason the rating moved up.


But Appendix C states the limiting condition plainly:


The 27 studies supporting this recommendation demonstrated positive effects... with all studies testing interventions that included the recommended practices as a major component.

A major component. Not the independent variable. The 2009 guide said the same thing about its own smaller base — "In some cases, fact fluency instruction was one of several components in the intervention, and it is difficult to judge the impact of the fact fluency component alone" — and the 2021 base is built the same way.


What has Strong evidence is a package: explicit strategy instruction, previously-taught content, immediate corrective feedback, student progress charting, and a short timed activity inside it.


The appendix is explicit about which parts came from studies and which from the panel. Nineteen of the 27 studies taught students an efficient strategy before the timed activity. Seventeen had students chart their own progress.


Twenty-six delivered immediate feedback. And then: "The first and second How-to Steps advise teachers on how to select fluency-building topics and how to choose the activity and materials for implementation. These steps are based on panel advice."


Two of the five how-to steps are opinion, labelled as such. This is the same shape as the dosage finding in intervention session frequency and duration and the grouping finding in reading program group size: a published specification whose parts have very different warrants, presented as one block.

Nine outcome domains, and the one that is not there


Appendix C lists the domains in which the 27 studies produced findings. All nine were statistically significant and positive:


  • Algebra and algebraic reasoning — g = 0.55, 95% CI [0.29–0.81], 3 studies

  • Counting and cardinality — g = 0.27 [0.04–0.49], 2 studies

  • General mathematics achievement — g = 0.35 [0.24–0.47], 5 studies

  • Rational numbers computation — g = 1.55 [1.42–1.67], 9 studies

  • Rational numbers knowledge — g = 0.57 [0.46–0.68], 9 studies

  • Rational numbers magnitude understanding — g = 0.97 [0.86–1.09], 8 studies

  • Whole numbers computation — g = 0.64 [0.54–0.74], 16 studies

  • Whole numbers knowledge — g = 0.09, 1 study

  • Whole numbers magnitude understanding — g = 0.26 [0.11–0.41], 3 studies


Read that list for what is missing. There is no "basic fact retrieval" domain and no "automaticity" domain. The outcomes on which this recommendation earned its Strong rating are downstream mathematics outcomes — computation, magnitude, algebra, general achievement. Nobody rated the recommendation on whether students got faster at facts.


That is a stronger result than practitioners usually imagine, not a weaker one. But it is a result about an instructional package measured on mathematics, which is a different sentence from "automaticity transfers to problem solving."

What the 2026 meta-analysis found, and where the effect goes


The most useful single document published on this question is recent enough that it postdates most training material in circulation. Douglas, Myers, Mason, Powell and Lariviere (2026), "A Meta-Analysis of Mathematics Fact Fluency Interventions for Students With Mathematics Difficulties (MD)," in the Journal of Learning Disabilities, is open access. Thirty-five group-design studies, 178 effect sizes, 1975 to June 2024.


The headline is encouraging: a robust variance estimation model gives an average g = 0.76, 95% CI [0.46, 1.06]. The second number is the one to carry:


The 95% prediction interval was wide (PI = −0.60 to 2.12), indicating that the true effect of a new, similar study could plausibly range from moderately negative to highly positive.

A prediction interval is the range a new study's effect would plausibly fall in. This one includes zero and goes well below it. The average is good; the dispersion means the average is a poor guide to what will happen on your caseload. That is the same reading problem this site described for standard error of measurement — a point estimate reported without its spread invites a confidence nobody has earned.


Then the table that is, for a practitioner, the single most informative thing in the paper. Effect size by what was measured:


  • Fact fluency measures — g = 0.89 [0.52, 1.26], 29 studies, 92 effect sizes

  • Computational fluency measures — g = 0.70 [0.30, 1.10], 14 studies, 35 effect sizes

  • Word-problem solving measures — g = 0.25 [0.09, 0.41], 9 studies, 29 effect sizes

  • Other measures (place value, number line) — g = 0.32 [0.07, 0.56], 7 studies, 22 effect sizes


0.89, then 0.70, then 0.25. The effect is largest on the thing that was trained and smallest on the thing the training is justified by. The authors are measured about it, writing that "the significant average effect observed for word-problem outcomes (g = 0.25) suggests that fact fluency interventions may support transfer to more complex mathematical applications."


That is a fair reading and it is the one to use. There is a significant word-problem effect — and it is a quarter of a standard deviation, roughly a third of the fact-fluency effect, and it is the only one of the four whose prediction interval is narrow and entirely positive [0.11, 0.39]. Fact work appears to buy something downstream. It buys a good deal less downstream than it buys upstream.


Anyone who has read this site's piece on working memory training programs will recognise the shape exactly: near transfer is large, far transfer is small, and the argument in the field is about how far the improvement has to travel before it counts. The same gradient, in a different content area, with a better-behaved far-transfer estimate.

Dosage: the numbers do not do what you expect


The same meta-analysis ran all 20 moderators in one meta-regression. Two dosage indicators came out significant — grouping and total sessions — and the descriptive pattern inside them is not the pattern most practice assumes.


Total number of sessions:


  • Fewer than 10 sessions — g = 1.52 [0.48, 2.57], 8 studies

  • 10–20 sessions — g = 0.32 [−0.34, 0.98], not significant, 8 studies

  • 21–29 sessions — g = 0.43 [0.00, 0.86], 4 studies

  • 30 or more sessions — g = 0.76 [0.31, 1.22], 15 studies


Total instructional hours:


  • Fewer than 10 hours — g = 0.81 [0.33, 1.28], 19 studies

  • 10–16 hours — g = 0.37 [0.24, 0.49], 6 studies

  • Over 16 hours — g = 0.97 [0.22, 1.72], 10 studies


Frequency:


  • Daily — g = 0.71 [−0.14, 1.57], not statistically significant, 8 studies

  • Once or twice per week — g = 0.58 [0.06, 1.10], 10 studies

  • Three or four times per week — g = 0.88 [0.42, 1.34], 17 studies


Neither duration nor frequency is monotonic, and daily practice — the convention the 2009 guide's "every session" language installed — is the one frequency band that did not reach significance in this synthesis. Frequency was not a significant moderator in the regression at all.


Be careful with this. These are between-study subgroup comparisons across a small and heterogeneous set, not a randomised comparison of dosages within one study, and the authors flag Type I error risk explicitly.


You may not tell a family that the research shows less practice works better. What you can say is that the published record does not support a confident claim that more minutes buys more learning. That is the same conclusion the record reached about reading intervention dosage and about group size.


One moderator did behave cleanly and is worth your attention: operation focus. Interventions targeting multiplicative facts returned g = 1.27 [0.47, 2.07] against g = 0.49 [0.17, 0.81] for additive. The authors read this as content alignment mattering, and it is the most actionable number in the paper for a caseload that is mostly third grade and up.

The app question: technology versus a deck of cards


Practitioners spend money here, so it is worth separating two comparisons that vendors tend to merge.


Burns, Duesenberg-Marshall, Romero, Sussman-Dawson and Singell (2025), "Meta-Analysis of the Effect of Technology-Based Mathematical Fact Practice on Mathematics Outcomes," synthesised 17 effects from 12 studies. Verbatim from the abstract:


There was a moderate effect of technology-based mathematical fact practice on mathematics outcomes (weighted g = 0.43). There was a moderate effect in comparison to business-as-usual conditions (weighted g = 0.54), but a much smaller effect for practice that was compared to practice that was not delivered with technology (weighted g = 0.25).

Two different questions, two different answers. Against no structured fact practice, the app returns 0.54. Against the same practice on paper, it returns 0.25. The larger number is mostly an argument for practising; the smaller one is the argument for the subscription.


Students at risk for mathematics disabilities returned g = 0.55. The large effect on general mathematics achievement, g = 0.87, rests on two studies, which the authors say in the abstract.


They also found that the interaction of sessions and weeks predicted effect size and accounted for 23% of variance, while neither variable predicted it on its own — another dosage result that refuses to be simple.

NCTM says avoid timed tests. IES recommends timed activities. Both are right.


This is the contradiction practitioners trip over, and it dissolves once you notice which verb each document is using.


NCTM's fourth declaration, verbatim:


Assessing must attend to fluency components and the learner. Assessments often assess accuracy, neglecting efficiency and flexibility. Timed tests do not assess fluency and can negatively affect students, and thus should be avoided.

Its third declaration, verbatim, is that "Basic facts should be taught using number relationships and reasoning strategies, not memorization."


IES recommends a timed activity as practice, explicitly placed after a concept has been taught over many lessons, explicitly not used to introduce content, explicitly bundled with strategy instruction and immediate feedback. NCTM objects to a timed test as assessment, on the grounds that a speeded accuracy measure does not capture flexibility or strategy choice and carries an affective cost.


Those are compatible positions about two different activities that look identical from the doorway. And the 2021 guide, remarkably, is on NCTM's side of the substance even while recommending the timer. Its own "potential obstacles" section answers a teacher who says students are not improving with:


In the opinion of the panel, giving timed worksheets alone does not support fluency.

And on anxiety, its advice is to "make sure students know they are not expected to finish all of the problems and that there are more items on the worksheet than they are expected to complete."


It also suggests using "flashcards or other activities that do not present many problems at the onset" instead of a long list, and having students work as a group to "meet or beat" a collective score, which "can also decrease the pressure they may feel if asked to perform individually."


The two 2009-guide findings that survive into both positions are worth stating together, because they are the operational core of everything above: strategy-based instruction for fact fluency is superior to rote memorization (the 2009 guide says so in those words, citing three studies), and the timed piece works inside a structure, not instead of one.


The argument is not timed versus untimed. It is practice versus assessment, and a timer inside a taught strategy versus a timer instead of one.

What the publishers claim, and what they cite


The market here is large and the evidence pages are uneven. Four examples, reported as what each publisher publishes about itself.


ExploreLearning Reflex. Its research page names no individual studies and reports no effect sizes, referring instead to "several research studies conducted by university-affiliated researchers over the past 10+ years" and offering a research summary and white paper as downloads.


Its companion page on why fact fluency matters uses the exact phrase that this whole post is about — that fluency lets students "free up their working memory" and "devote more time to advanced problem-solving and learning new skills" — and attaches two citations.


One is a PsycNET record from 1994, which could not be retrieved for this piece (see below). The other is PubMed 10373315, cited for the claim that "fact retrieval speed is a significant predictor of performance on standardized tests."


That record is Royer and colleagues, “Math-Fact Retrieval as the Cognitive Mechanism Underlying Gender Differences in Math Test Performance” — a real paper on a real association, and a study about gender differences in test performance rather than about what a fact-fluency intervention produces.


Rocket Math. Its research page is unusually candid in one respect: it lists its evidence base in full, and most of that base is unpublished. Of the ten entries, the published ones are Rave and Golightly (2014) in Education 134(4), 537–547, described on the page itself as "a case study," and Smith, Marchand-Martella and Martella (2011) in Education and Treatment of Children 34(2), 247–258, a single-student case study.


The rest are master's capstones, unpublished dissertations and unpublished manuscripts by the program's author. The page says so plainly — "While I am waiting for others to conduct and publish more research on Rocket Math" — and offers free subscriptions to researchers who will run studies.


Separately, the page asserts that "students who have mastered math facts show marked improvement in higher order math algorithms including fractions, word problems, long division, multi-digit multiplication," with no citation attached.


XtraMath. Its logic-model post states: "Educators should be aware that Xtramath's program has strong evidence of effectiveness, as demonstrated by its ESSA Tier IV of Evidence rating." Both halves of that sentence are quoted exactly. See the next section.


The pattern worth naming. Not one of these pages reports a confidence interval, and not one distinguishes an effect on facts from an effect on mathematics. That is not unique to math — it is the same gap this site documented in assistive technology claims and in reading program evidence ratings.

Reading an evidence badge: what the ESSA tiers actually are


Because the badges are everywhere, here is the scale, as the IES regional lab's module on ESSA evidence requirements sets it out:


  • Tier 1 — "Strong evidence." At least one study meeting WWC group design standards without reservations, large multisite sample, at least one statistically significant positive outcome.

  • Tier 2 — "Moderate evidence." At least one study meeting WWC standards with reservations, large multisite sample, significant positive outcome.

  • Tier 3 — "Promising evidence." At least one correlational study with statistical controls for selection bias, significant positive effect.

  • Tier 4 — "Demonstrates a rationale." A well-specified logic model with an intended outcome aligned to a stakeholder need, supported by existing or ongoing research.


Tier 1 is the tier whose name is "Strong evidence." Tier 4 is the tier whose name is "Demonstrates a rationale," and its requirement is a logic model. Those are the two ends of the same four-point scale. A reader who sees "strong evidence of effectiveness, as demonstrated by its ESSA Tier IV of Evidence rating" now has what they need to interpret it.


This is a general-purpose skill and it is worth practising on every badge you meet. The same applies to what each reading evidence rating system is actually asking.

The clearinghouse that rates tools has no fluency domain


The National Center on Intensive Intervention runs three tools charts, and they are the closest thing the field has to an independent technical review of products. Checked directly this week:


  • The Academic Progress Monitoring chart filters by Subject: Reading, Mathematics, Spelling & Written Expression.

  • The Academic Screening chart filters by Subject: Reading, Mathematics.

  • The phrase "fact fluency" appears zero times on all three charts, as do "math fact" and "basic fact" as rated categories.


There is no fact-fluency domain, no automaticity domain, and no rated subcategory for either. The only place the word "automatic" appears in the progress-monitoring tables is the column reporting whether a tool calculates rate of improvement automatically — a software feature, not a construct.


Two consequences. First, if you want an independently reviewed progress-monitoring tool for fact fluency specifically, the clearinghouse does not have a shelf for it; you are reading computation measures and deciding for yourself whether they index what you are teaching.


Second, and this is the standing caution from NCII's own tools chart user guide, ratings rest on data "submitted by developers," and "the presence of a tool on a chart does not constitute endorsement." Absence from a chart means nobody submitted. It does not mean a tool failed. The same caveat governs how to read the charts in progress monitoring tools compared.

What you may and may not say to a family


Translating all of this into a sentence a parent can hear is the part of the job no document does for you. A defensible version:


Practising facts to the point of quick, accurate recall has good evidence behind it, especially for the facts themselves and for computation. There is some evidence that it helps with word problems too, and that effect is real but considerably smaller. It is one piece of a program, not the program.

Four things to avoid, each tied to a specific thing above:


  • Do not say that automatic facts free up working memory. The strongest federal statement in that direction is "more mental energy," it is footnoted rather than demonstrated, and the 2009 version of the downstream claim is labelled panel opinion in print.

  • Do not quote the average effect size without the spread. g = 0.76 with a prediction interval of −0.60 to 2.12 is not a promise about one student.

  • Do not promise that fact work will raise a general mathematics score. The 2009 guide says the findings "cannot be extrapolated to areas of mathematics outside of fact fluency and operations," and the 2026 synthesis puts the word-problem effect at a quarter of a standard deviation.

  • Do not describe a timed worksheet as an assessment of fluency. NCTM's position statement says timed tests do not assess fluency, and the IES glossary's own definition of fluency has no speed in it.


This is the same discipline described in is educational therapy evidence-based? — state the finding with its boundary in the same breath, and never launder an adjacent result into an on-point one.

What AET's Code of Ethics says about any of this


Nothing, and the nothing is specific.


The AET Code of Ethics, Revision #9, 2025 is about 2,980 words. Searched this week, it contains:


  • "math" — 0 occurrences

  • "arithmetic" — 0

  • "fluency" — 0

  • "automatic" — 0

  • "timed" — 0

  • "computation" — 0

  • "memorization" or "memorize" — 0

  • "drill" — 0

  • The string "fact" appears once, inside the word "factors."


The clause that actually governs a decision about whether to spend ten minutes a session on facts is II.I.E, which asks educational therapists to "select and use appropriate instructional methods, curricula, materials, and other resources to meet the unique needs of each client."


That is a competence standard and a client-fit standard, and it is the whole of the profession's written guidance here. II.II.I, on conveying realistic limitations, is the clause that bears on what you tell a family about what fact practice will buy.


So the professional body leaves the decision to you and asks that you be able to justify it as appropriate for the individual client. This matches what the same Code was found to say — which is to say, nothing — about group size and about documentation standards.


The pattern is consistent enough now to be predictive: AET regulates conduct, not method. Where practitioners want a method rule, there is a competence clause and a gap.


Note also that "not named" is not "not permitted." If you want AET's view, ask in writing and keep the reply.

When several organisations share a vocabulary


The finding in this post is not unusual. It is the standard result whenever you read a field's shared words back to their sources, and this site has now found the same shape eight times:



The method is the same every time, and it is the only one that reliably works: find the document the claim is attributed to, read the sentence, and check whether it says what the field says it says.

What is out of scope here


This post is about what the published research and the published manuals say. It is not about what anyone is legally required to do, and several adjacent questions are governed by law, regulation or contract rather than by evidence:


  • State mathematics screening and intervention mandates, where these exist.

  • Special education eligibility in mathematics, and what a school district must provide.

  • IEP goal language and service minutes, including any goal written around fact fluency.

  • What a district contract or an insurer requires of a provider.


Nothing here is legal advice. Those questions belong with your state education agency and, for anything touching a contract or independent practice, with an attorney. Where an IEP is involved, the mechanics in this post are offered as instructional information only and the obligations are the school's.

Where the published record runs out


Seven gaps, named so nobody reads past them:


  • The working-memory mechanism is not tested in any of this. Both federal guides gesture at it and neither measures it. The 2026 meta-analysis discusses cognitive load in its literature review and does not test it as a mechanism. What has been measured is outcomes, not the pathway.

  • The timed component has not been isolated. Every study in both federal evidence bases bundles it. There is no body of work comparing a program with the timer to the identical program without it.

  • Two sources could not be verified for this post and are therefore not relied on. The PsycNET record cited by ExploreLearning for the working-memory claim (1994-98944-000) returned an access error on direct fetch, so what that source says is not reported here. The Reflex white paper and the XtraMath logic model are both publisher-hosted PDFs whose contents are not summarised here because they were not retrieved.

  • Almost all of this is elementary. The 2021 guide covers grades K–6; the 2009 evidence base ran grades 1 to 4. The 2026 meta-analysis did include grade 3+, returning a larger effect than lower elementary (g = 0.94 against 0.59) from 19 studies spanning grades 3–12, with no secondary-only estimate.

  • For adolescents specifically, the record is thin, as it is for older students in reading.

  • The external source universe here is small — two federal practice guides, two meta-analyses, one professional position statement, three NCII charts, four publishers and one ethics code. That is not a shortfall in the research for this post; it is the complete set of primary documents that exist on the question. Padding it would mean citing other people's commentary about these documents instead of the documents.

  • There is no WWC intervention report on a math fact fluency product. The practice has a Strong rating; no individual program carries an equivalent federal rating for this purpose, which is the same asymmetry found in math intervention for dyscalculia and in math assessment tools.

Common questions


Is there strong evidence for timed math fact practice?


Yes, with a precise meaning. The 2021 IES practice guide rates its timed-activities recommendation Strong, on 27 studies covering 4,336 students in 403 schools.


The rating attaches to a multi-component package — a short timed activity on previously taught content, inside strategy instruction with immediate feedback and student progress charting — and not to a timer used on its own. Every study in the base tested the practices as a major component of a bundle, so the timed element has never been isolated.


Does automaticity with math facts free up working memory for problem solving?


That is not what the federal record says. The 2009 guide states that weak retrieval is "likely to impede" understanding because teachers and texts assume the facts, and its technical appendix then states that the broader implication — that general mathematics proficiency improves when fact fluency improves — "is the opinion of the panel."


The 2021 guide's version is "Automatic retrieval gives students more mental energy," and that guide uses the phrase "working memory" three times in 165 pages, all outside the argument. The mechanism is plausible and it is not what was tested.


How many minutes a session should fact practice take?


The two federal guides give different answers and the newer one does not acknowledge the change. The 2009 guide recommends "about 10 minutes in each session," with "a minimum of 5 minutes." The 2021 guide specifies timed activities "between 1 and 5 minutes," as one component rather than the focus. If you are quoting ten minutes, you are quoting the superseded document.


Do math fact apps work better than flashcards?


Burns and colleagues (2025) separate the two comparisons. Technology-based fact practice against business-as-usual returned weighted g = 0.54; against the same practice delivered without technology, g = 0.25. Both are positive. The second is the one that answers the purchasing question, and it is less than half the first.


Why does NCTM say to avoid timed tests when IES recommends timed activities?


Because they are describing different acts. NCTM's January 2023 position statement objects to timed tests as assessment: "Timed tests do not assess fluency and can negatively affect students, and thus should be avoided."


IES recommends a brief timed activity as practice, after a concept has been taught, inside a structure, and explicitly not as a way to teach new content. The 2021 guide itself says that giving timed worksheets alone "does not support fluency." The two positions are compatible.


Should I use strategy instruction or drill?


The 2009 guide answers this directly and it is one of the few places the record is unambiguous: "there is evidence that strategy-based instruction for fact fluency (such as teaching the counting-on procedure) is superior to rote memorization."


NCTM's third declaration says the same: "Basic facts should be taught using number relationships and reasoning strategies, not memorization." Both the research-synthesis route and the professional-organisation route land in the same place, which is rare enough to be worth noticing.

Related reading


This post sits inside a cluster. If you are building out the mathematics side of a practice, the companion pieces are math intervention for dyscalculia on what the programs are and what the label does, math assessment tools for dyscalculia on what the instruments measure, and working memory training programs for the near-versus-far-transfer argument in its original setting.


For the surrounding method questions: intervention session frequency and duration on what published dosage figures are and are not, reading program group size on the same problem in grouping, reading program evidence ratings on how five rating systems differ, and progress monitoring tools compared on reading a growth graph outside a district.


Get the practitioner list


We write one of these a day: primary sources, exact quotations, and the places where the published record runs out, written for people who do this work rather than for the families who hire them.


Join the educator list and the next one comes to you.

Comments

bottom of page