What is adaptive learning? A plain-English guide for parents

By The Sophoz Curriculum Team

The short answer. Adaptive learning is software that changes what it asks your child next, based on how your child answered the last thing. A fixed worksheet gives every child the same twenty questions in the same order. An adaptive system gives a child who is struggling different questions from a child who is flying, and it decides which is which from the answers rather than from the year group. That is the whole idea. Everything else is detail about how well a particular product does it.

The term gets used loosely, so this guide separates what adaptive learning actually means from what the word is often used to sell.

The core idea, in one example

A child answers a question about equivalent fractions incorrectly.

A non-adaptive system serves the next question on the list. It is another equivalent fractions question, because that is what chapter 4 contains. If the child does not understand equivalent fractions, they now get nineteen more chances to not understand it.

An adaptive system treats the wrong answer as information. Equivalent fractions rest on something underneath: understanding multiplication as scaling, the idea that multiplying the top and bottom by the same number changes how the fraction is written without changing how much it is. If that idea is not secure, no amount of fraction practice will fix it, because the problem is not in fractions.

So the adaptive system goes down a level, checks whether multiplication as scaling is solid, and if it is not, practises that instead. Then it comes back.

That is the difference. Not more questions, or nicer questions, or questions with better animation. Different questions, chosen because of what the child just did.

The four things a system has to do to be genuinely adaptive

Any product can call itself adaptive. These are the mechanisms that make the word mean something.

1. It knows what depends on what

Mathematics is not a list of topics. It is a structure where some things cannot be learned before others. Long division needs subtraction and multiplication facts. Ratio needs division as sharing. Algebra needs the idea that a letter can stand for a number that has not been decided yet.

A system that knows this holds a map of skills and their prerequisites, sometimes called a skill graph. When a child fails on a skill, the map tells the system what sits underneath it and therefore what to check. Without that map, a system can only do one thing when a child gets something wrong, which is serve more of the same.

This is also why a genuinely adaptive system follows the curriculum sequence rather than working around it. Practice that ignores the order in which things have to be learned is not neutral; it is worse than no practice, because it builds confident wrong methods on top of gaps.

2. It looks at more than right or wrong

A tick and a cross carry very little information. Two children can both get a question wrong for completely different reasons: one misread it, one made an arithmetic slip, one used a method that does not apply, one has no idea. Those need four different responses.

Better systems look at what was actually written, how long it took, and whether the same error shape keeps recurring across different questions. A child who takes forty seconds and gets it right has not demonstrated the same thing as a child who takes four seconds and gets it right, and a system that records both identically has thrown away the useful part.

3. It changes difficulty while the child is working

This is the test that separates adaptive systems from systems that merely place a child at a level once. Placement at the start of term is not adaptation. Adaptation is the difficulty moving within a session, because the child is doing better or worse than the system expected.

4. It tells someone something useful

The output matters as much as the mechanism. "Your child scored 68 %" is a number a parent can do nothing with. "Your child is converting improper fractions correctly but is dividing instead of multiplying when scaling up a recipe, which suggests the scaling idea is not secure" is a sentence a parent can act on at the dinner table.

A system that can produce the second kind of sentence has understood something. A system that can only produce the first kind has counted something.

What adaptive learning is not

It is not personalisation of the front page. An app that remembers your child likes dinosaurs and puts dinosaurs in the word problems has personalised the decoration. The mathematics is identical.

It is not a recommendation engine. Suggesting the next video based on what similar children watched is the mechanism behind streaming services. It changes the order of content. It does not change the difficulty, and it does not diagnose anything. We have written a buyer's guide to telling these apart in personalised learning app or just a recommendation engine.

It is not a replacement for knowing your child. The system sees answers. It does not see that the last fortnight has been difficult, or that the child is quietly convinced they are bad at maths. Those are the things that most change what should happen next, and no software has access to them.

It is not artificial intelligence being clever. Most of what makes an adaptive system work is the map of prerequisites, which is curriculum design done carefully by people. The software's job is to walk the map quickly.

What the evidence actually shows

This is where parents are most often oversold, so here is the honest version.

Good tutoring systems work roughly as well as good human tutoring. In the most careful review of this question, human tutoring came out at d = 0.79 against no tutoring, and step-based intelligent tutoring systems at d = 0.76. The difference between them was d = 0.21, which was not reliably different (VanLehn, 2011, Educational Psychologist).

That is a striking result and it is the strongest thing that can honestly be said for this category.

But the famous two-sigma number is wrong. Bloom's 1984 claim that one-to-one tutoring is worth two standard deviations is widely quoted and does not survive scrutiny. VanLehn's conclusion was that "it seems that human tutoring is not usually 2 sigmas more effective as classroom instruction, as the six studies presented by Bloom (1984) invited us to believe", with the true figure closer to 0.79. Bloom's tutors were minimally trained undergraduates, and the tutored group was held to a 90 % mastery threshold against 80 % for the comparison, which inflated the gap.

So: still one of the largest effects in education. Less than half the number that gets repeated in edtech marketing.

Hints are harder to get right than they look. Research on tutoring systems found students viewed 68 % of non-final hint levels for less than one second, too fast to have read them. And the opposite problem is just as common: even after three errors on a step, students requested a hint only 34 % of the time. Most sobering, a system explicitly designed to teach better help-seeking produced durable improvement in help-seeking behaviour months later, and no improvement in what students actually learned (Aleven, Roll, McLaren and Koedinger, 2016).

The right amount of support changes as the child improves. Detailed guidance that helps a beginner becomes redundant and then actively harmful as they gain expertise, because processing the unnecessary guidance costs attention. This is called the expertise reversal effect (Kalyuga, Ayres, Chandler and Sweller, 2003). It is the strongest argument for adaptation as a principle, because it means there is no single correct level of support. The correct level moves, which is exactly the thing software can do and a printed workbook cannot.

The strongest finding in this whole area is not about software at all. Mixing problem types rather than practising one topic at a time produced 61 % against 38 % on a delayed test in a randomised trial with 787 seventh-graders (Rohrer, Dedrick, Hartwig and Cheung, 2020). A workbook is organised by chapter, which is blocked practice by construction. Any parent can fix that with an index card and ten minutes a week. No purchase required.

Is it right for your child?

Adaptive practice suits a child who keeps making the same kind of error and nobody has worked out why, or who is ahead in some areas and behind in others so that any single worksheet is wrong for them in both directions, or who has a parent without the time to sit alongside four evenings a week.

It suits a child less well when the real issue is motivation rather than diagnosis, when the child is very young and needs a person, or when what is actually needed is somebody to notice they are discouraged. Software is poor at all three.

One age-specific caution. For children in roughly grades 2 to 5, the evidence on letting children work things out before being taught turns negative: in the meta-analysis of productive failure, effects were positive overall but negative for grades 2 to 5, where instruction-first outperformed problem-solving-first (Sinha and Kapur, 2021). A system that drops a young child into an unfamiliar problem and waits is applying an approach the evidence does not support at that age.

Four questions to ask before paying

1. When my child gets something wrong, what happens next? If the answer is "another similar question", it is a question bank. If the answer describes checking something underneath, it is adaptive.
2. Does the difficulty change during a session, or only at placement? Placement once a term is not adaptation.
3. What will I be told, and in what words? Ask to see a real parent report. If it is percentages and progress bars, the system has counted. If it names a misconception in a sentence, it has diagnosed.
4. Does practice mix topics or follow chapters? Chapter order is blocked practice, which the evidence says is the weaker way to practise. A system that only walks the syllabus in order has automated a workbook.

Those four questions take one conversation and they separate most of the market.

What Sophoz does, for transparency

Sophoz is an adaptive learning platform, so this guide is not neutral and you should read it accordingly.

What Sophoz does specifically: each board and exam sits on a skill graph mapping competencies and their prerequisites, so a wrong answer on equivalent fractions leads to a check on whether multiplication as scaling is secure rather than to more fraction questions. Practice follows the school's curriculum sequence rather than an independent one. Progress is tracked across six levels of mastery, from Recall through Explain, Apply, Connect and Generalise to Teach, because being able to state a rule and being able to explain it to someone else are different things. Assessment looks at reasoning and time taken rather than only correctness. Weekly parent reports quote what the child actually wrote and name the misconception in a sentence. And the assistant answers with a hint or a different way of seeing the problem rather than with the answer.

Sophoz reports 1,200-plus students and an average score improvement of 28 %-plus across practice journeys. Both figures are self-reported by us and unaudited, as is every comparable number published by anyone else in this category, and you should treat them accordingly.

Frequently asked questions

What is adaptive learning in simple terms? Software that changes what it asks your child next, based on how they answered the last thing. A fixed worksheet gives every child the same questions in the same order; an adaptive system chooses different questions for different children, using their answers to decide.

Is adaptive learning better than a normal learning app? It is better at one specific thing: working out which underlying skill a child is missing when they keep getting a topic wrong. It is not automatically better at teaching, motivating, or explaining. Many apps that call themselves adaptive only change the order of content, not the difficulty or the diagnosis.

Does adaptive learning actually work? The best evidence is that step-based intelligent tutoring systems produce roughly the same effect as human tutoring, d = 0.76 against d = 0.79, with no reliable difference between them (VanLehn, 2011). That is a genuinely large effect. The widely quoted claim that tutoring is worth two standard deviations does not hold up and should be treated as marketing.

What age is adaptive learning suitable for? It is used across primary and secondary. One caution for younger children: research on problem-solving before instruction finds negative effects for grades 2 to 5, so a system that drops a young child into unfamiliar problems and waits is using an approach the evidence does not support at that age.

Can an adaptive app replace a tutor? On the measured evidence they are closer than most people expect, but a tutor notices things software cannot: discouragement, distraction, a child who has quietly decided they are bad at maths. The honest comparison depends on what the child actually needs.

How can I tell if an app is genuinely adaptive? Ask what happens after a wrong answer. If it serves another similar question, it is a question bank. Ask whether difficulty changes mid-session or only at placement. And ask to see a real parent report: percentages mean it counted, a named misconception means it diagnosed.

---

Research citations are linked in full in what the research says about olympiad preparation. Sophoz builds adaptive practice, which is a reason to read this guide sceptically; the evidence above is cited so you can check it.