Adaptive Learning AI and the UK Curriculum: A Guide

Every edtech product page now says "adaptive". For a multi-academy trust education lead, that word covers three very different machines: a fixed branching path written in advance, a tool that lets each pupil move at their own pace, and a system that keeps a running estimate of what each pupil actually knows. They behave differently in a classroom, they produce different evidence, and they age differently when the curriculum changes. The curriculum is about to change: the Department for Education aims to publish the final revised national curriculum by spring 2027 for first teaching from September 2028, with updated GCSEs from 2029 onwards.

This guide explains how adaptive learning AI works underneath, starting with knowledge tracing, and how an adaptive engine sits beside the national curriculum rather than replacing it. There is no single "UK curriculum": England, Wales, Scotland and Northern Ireland each set their own, so the curriculum facts here are England's, with the Curriculum for Wales and Scotland's Curriculum for Excellence linked for readers in those nations. The scope is deliberate. How a tutor questions a pupil is covered in our guide to the Socratic AI tutor. Safeguarding sits in the AI tutor safeguarding review for schools and trusts. Whether Dilr Academy claims curriculum alignment is answered in the Khanmigo alternative comparison, and this post stays consistent with that answer.

This guide is shipped by the team behind Dilr Academy, an AI tutor that builds interactive, multilingual courses on demand, with mastery tracking tuned by knowledge tracing. Or see DATS, the five-stage AI consulting system, delivered by senior practitioners who ship code, not decks.

What is adaptive learning AI?

Adaptive learning AI is software that changes what a learner sees next based on evidence about what that learner currently knows. The defining feature is a learner model: a running estimate of mastery, skill by skill, updated after every answer. A tool that only offers pre-written branches, or only lets pupils move at their own speed, is personalised in a weaker sense, because nothing in it is inferring understanding from the pattern of responses.

The distinction matters because, as a design matter, the three fail in different ways. A branching tool is predictable and comparatively easy to audit, but it only knows the routes its authors wrote, so a pupil whose misconception was not anticipated falls through. A self-paced tool hands the pacing decision to the pupil, and the EEF finds mastery approaches much less effective when pupils work at their own pace. A tool with a learner model can choose the next step for reasons the author never scripted, which is its strength and also why a school needs to see how it decides.

When a supplier says "adaptive", ask which row they mean. The answer shapes everything else in this guide, from the evidence to the 2028 re-mapping question. Our AI teacher platform buyer's guide covers the wider shortlist; this post goes one level down into the mechanism.

How does knowledge tracing work?

Knowledge tracing is the method an adaptive tutor uses to estimate, after each answer, the probability that a learner has mastered each skill in a course. Correct answers raise the estimate, errors lower it, and the tutor picks the next exercise from those estimates until each skill crosses a mastery threshold. The approach was described by Corbett and Anderson in 1994, and later work has added neural versions of the same idea.

The original paper, Corbett and Anderson's "Knowledge tracing: Modeling the acquisition of procedural knowledge", published in December 1994, describes a programming tutor that maintained an estimate of the probability that the student had learned each rule in its model, and presented an individualised sequence of exercises until each rule was mastered. That loop, answer, update, choose, check, is the core of the learner-model design in the table above.

Two practical complications shape how well it works. First, a single answer is weak evidence: a pupil can guess right without understanding, or slip on something they do know, so a good model treats each response as a signal rather than a verdict. Second, the model can only trace skills it has been told exist. Someone has to break a subject into skills and say which ones depend on which, and that skill map is where the curriculum enters.

In 2015, Piech and colleagues' Deep Knowledge Tracing paper at NeurIPS showed that recurrent neural networks could model student learning without explicit encoding of human domain knowledge, and outperformed the then state of the art in prediction on real student data. That result carries a trade-off a school should understand, which is our reading rather than a finding of the paper: the more flexible the model, the harder it tends to be to explain why it chose a given exercise. A head of department can read a hand-built skill map; a neural model does not offer its reasons in the same readable form. Neither family is automatically better for a classroom. The question to ask is whether the supplier can show you, in plain terms, why the tool moved a pupil on.

How does an adaptive engine fit the national curriculum in England?

An adaptive engine fits inside the national curriculum rather than replacing it. The curriculum in England sets what must be taught, organised as programmes of study across four key stages; an adaptive engine decides the order, pace and difficulty of practice within a skill map a person has built from that content. Coverage, sequencing across the year and the school's wider curriculum remain decisions for teachers and leaders.

The legal shape is worth getting exactly right. The DfE's framework for key stages 1 to 4 says maintained schools in England are legally required to follow the statutory national curriculum, which sets out programmes of study on the basis of key stages, and that it is organised on the basis of 4 key stages and 12 subjects. GOV.UK's national curriculum guide adds that academies and private schools do not have to follow it today, although academies must teach a broad and balanced curriculum including English, maths and science. In its November 2025 response to the curriculum review, the government said it was legislating through the Children's Wellbeing and Schools Bill so that academies would be required to teach the refreshed national curriculum alongside maintained schools. A buyer working across maintained schools and academies should check where that legislation stands when it signs, a point worth settling before an AI placement diagnostic ranks where AI belongs.

The framework itself is clear about its limits:

"The national curriculum is just one element in the education of every child." DfE, The national curriculum in England: framework for key stages 1 to 4, paragraph 3.2

The same paragraph describes the curriculum as an outline of core knowledge around which teachers develop lessons. That is the right mental model for an adaptive tool. The programme of study says what a pupil should know by the end of a key stage. The skill map breaks that into teachable steps and says which come first. The knowledge tracing model works inside the map, deciding which step this pupil needs next. Each layer has an owner, and only the last is automated.

For any school or trust, a useful test is to ask a supplier for its skill map in one subject and hold it against the relevant programme of study. If the supplier cannot produce one, the tool is either branching or it is tracing skills nobody at the school can inspect. Our AI teacher platform landscape for UK schools sets out where teacher-facing planners and pupil-facing tutors divide this work.

What changes for adaptive tools when the curriculum is revised in 2028?

The revised national curriculum changes the skill maps adaptive tools depend on. The government plans to publish the final revised curriculum by spring 2027, giving schools four terms to prepare for first teaching from September 2028, with updated GCSEs from 2029 onwards. Any adaptive tool sequenced against the current programmes of study will need its skill map re-mapped, so a school should ask each supplier for that plan now.

The Curriculum and Assessment Review final report was published on 5 November 2025 and applies to England. The government's response agrees that programmes of study should prioritise mastery of core concepts and be constructed with careful sequencing within and between subjects. For an adaptive engine that is close to a design brief: a curriculum that states its own sequencing more explicitly gives a skill map firmer ground, and a supplier whose map contradicts the new sequence will be visible quickly.

The response also says the refreshed curriculum will teach pupils AI literacy, how to use data to complete tasks and solve problems, and about bias in technology. That may have a second-order effect on procurement. Pupils who are taught how AI systems can be biased may reasonably ask how the tool in front of them makes its choices, which is one more reason to prefer a supplier that can explain its learner model in plain language, and a point our AI governance guide for education institutions makes about any AI an institution adopts.

Practical questions for the re-mapping plan:

1. When will the supplier publish its revised skill map, relative to the spring 2027 curriculum?
2. Will mastery estimates built on the old map carry across, or reset?
3. Who at the supplier signs off the mapping, and can a head of department review it?
4. Does the tool support the Welsh and Scottish curricula, or England only?

A tool that cannot answer the first question is asking a trust to buy into a sequence that has a known end date.

What does the evidence say about adaptive and individualised learning?

The Education Endowment Foundation's Teaching and Learning Toolkit is the clearest UK summary. It rates individualised instruction at four months' additional progress on average and mastery learning at five months, both on limited evidence, and feedback at six months on extensive evidence. It also finds digital approaches to individualised instruction about as effective as those without technology, so the software is a way of delivering the method rather than a source of extra impact.

The detail behind those averages is more useful than the headlines. The EEF's individualised instruction strand notes that approaches using digital technology to individualise instruction are as effective as those without technology, and that individualised instruction may be better used as a supplement to usual class teaching than as a replacement. Its effects are higher in science (+4 months) than in mathematics or reading (+3 months).

The mastery learning strand is the one an adaptive tool most resembles, and it carries a warning. Classic mastery learning asks pupils to show mastery on unit tests, typically 80%, before moving on, and the EEF finds a high bar of around 80% to 90% matters. It also finds the approach much less effective when pupils work at their own pace. That is the self-paced row of the table above. It is also our reason, an inference from the EEF findings rather than one of them, for preferring a tool with a stated mastery threshold over one that simply lets pupils choose when to advance. The EEF feedback strand sits behind the other half of the loop: the explanation a pupil receives after an error is feedback, which is where a tutor that questions before it explains earns its place, and its quality matters as much as the choice of the next item.

One distinction prevents a common muddle. The EEF separates mastery learning from "teaching for mastery", the East Asian approach to maths characterised by teacher-led, whole-class teaching and common lesson content, which in England is promoted through the NCETM's Teaching for Mastery programme. A school running teaching for mastery in maths keeps the class together on common content; an adaptive tool by design sends pupils down different paths. The two can coexist, with the tool used for practice and consolidation, but a supplier claiming to "support mastery" should say which meaning it intends.

The same diagnostic logic underpins our AI placement diagnostic, which produces a ranked roadmap of where AI belongs and where it does not.

Who regulates adaptive assessment in England?

Ofqual regulates qualifications and the awarding organisations that offer them, and its approach to AI is addressed to those organisations rather than to schools or tutoring suppliers. An adaptive practice tool used in class is not a regulated qualification. Ofqual's objectives are still a sound yardstick for any tool that tests pupils: fairness, validity, assessment security, public confidence and innovation.

Ofqual's approach to regulating AI in the qualifications sector has been in place since April 2024 and is underpinned by those five objectives. It expects awarding organisations to demonstrate that any use of AI delivers valid, reliable and fair outcomes, with appropriate expert human involvement and oversight. A school is not bound by that expectation for its own classroom practice, but borrowing the validity question is good discipline: does the tool's mastery estimate actually predict how a pupil performs on a teacher's own assessment?

Safeguarding and data protection questions sit with the school or trust as the buyer, and we cover them in the safeguarding review for AI tutors, and the institutional governance layer in our AI governance guide for education institutions.

Where does Dilr Academy's knowledge tracing fit?

Dilr Academy is an AI tutor that builds interactive, multilingual courses on demand, with mastery tracking that adapts to how the learner learns and adaptive difficulty tuned by knowledge tracing. As a learner answers, the tutor estimates what has genuinely been understood and tunes the next step, moving on only once a concept has landed. It sits in the learner-model row of the table, not the branching row.

On that description it belongs in the third design above, and it is worth being exact about what that does and does not mean. Each Dilr Academy course is a real syllabus, with modules, topics and stated outcomes, plus a one-to-one AI tutor with Deep Dives. The tutor asks before it tells, and live diagrams, animations and simulations carry the explanation. The Dilr Academy explainer walks through a course from the learner's side.

What it does not claim matters as much. Dilr Academy makes no claim to be aligned to the national curriculum: it builds courses on demand rather than offering a curriculum-mapped question bank. It publishes no attainment statistics, and it claims no accreditation. A school should run the five tests below on Dilr Academy exactly as on any other tool. A trust that wants examined, curriculum-sequenced maths practice for key stage 3 and 4 should weigh the established UK tools first. Dilr Academy fits a school that needs a course built on demand, or delivered in several languages including right-to-left scripts. For the wider shortlist, our AI teacher platform buyer's guide sets out how to compare tutors side by side.

How should a school test an adaptive tool before buying?

A school should test an adaptive tool on five things before buying: whether it models knowledge or only branches, whether it can show its skill map against the relevant programme of study, what mastery threshold it uses, whether a teacher can see and override its decisions, and whether its mastery estimates agree with the school's own assessments. A structured pilot with one department can run all five.

The five tests, in the order a pilot reaches them:

1. Model or branch. Ask the supplier to explain, for one pupil, why the tool chose the last three exercises. A learner-model tool can point to mastery estimates; a branching tool will point to a rule.
2. Skill map. Ask for the map in one subject and lay it beside the programme of study. A head of department should recognise the sequence and spot any gap.
3. Threshold. Ask what estimate counts as mastered. The EEF evidence on mastery learning favours a high bar, and a tool that advances a pupil on a single correct answer sets a low one by that standard.
4. Teacher control. Check that a teacher can see each pupil's estimates, hold a pupil back, or push a class on, because coverage and pace across the year belong to the school.
5. Validity. Compare the tool's mastery estimates with the department's own end-of-unit assessment for the same pupils. If they disagree often, the model is measuring something else.

Alongside the five tests, ask for the 2028 re-mapping plan described earlier, and settle two operating questions. Teacher time is finite: full-time primary teachers reported working 51.4 hours a week in 2025 in the DfE's Working Lives of Teachers and Leaders survey, so a tool that needs constant manual correction may cost more time than it saves. And someone has to own the pilot. The AI operating model we design covers governance, RACI and lifecycle, which is that ownership question written down, and our approach sets out how we place AI inside an organisation. For the attendance and admin side of a trust, the school attendance calls guide and the education sector overview cover where AI pays outside the classroom.

What is the best adaptive learning AI for UK schools in 2026?

The best adaptive learning AI for a UK school in 2026 depends on the job. For examined, curriculum-sequenced practice in one subject, a specialist UK tool built for that job is often the place to start. For courses built on demand or delivered in several languages, a course-building tutor such as Dilr Academy fits better. In every case the right tool passes the five tests above.

Named options, with the concession stated plainly. Sparx Maths describes itself as personalised homework for students aged 11 to 16, and in our view a maths department that wants homework sequenced across key stage 3 and 4 may be better served by a specialist like it than by a general course builder. CENTURY describes itself as online learning in English, maths and science with curriculum-aligned resources, which suits a school wanting cross-subject practice tied to its current scheme of work. Both use the language of personalisation; whether either runs a learner model in the sense used here is a question for the first test, put to the supplier directly. Khanmigo, Third Space Learning and others appear in our Khanmigo alternative comparison. Dilr Academy fits where the course needs building on demand or the learner needs it in another language.

Whichever you shortlist, run the threshold and validity tests on your own pupils, and agree how the tool will be re-mapped once the revised curriculum is published, aimed for spring 2027, before you sign a multi-year agreement. The execution office is embedded delivery, with production placements the client owns.

Adaptive learning versus personalised learning

Adaptive learning is one kind of personalised learning, not a synonym for it. Personalised learning covers any approach tailored to the individual, including self-paced work and teacher-chosen tasks. Adaptive learning AI adds a learner model that infers what a pupil knows from the pattern of answers and chooses the next step from that estimate. The distinction matters because the EEF's mastery learning strand finds the approach much less effective when pupils work at their own pace.

A supplier that uses the two words interchangeably is worth one follow-up question: what, exactly, decides the next exercise? The FAQ covers common questions about how DILR.AI products are built and supported.

Does adaptive learning work equally well in every subject?

Adaptive learning does not appear to work equally across subjects or phases, on the evidence available. The EEF Teaching and Learning Toolkit finds mastery learning effects higher in mathematics and science, at six months, than in reading, at three, and higher in primary, at eight, than secondary, at three. Individualised instruction effects are higher in science than in mathematics or reading.

That is a pattern in the evidence base rather than a rule, and the strands measure approaches, not knowledge tracing tools specifically. Our own reading, not an EEF finding, is that a skill map is easier to build where one idea plainly depends on another, and that a secondary trust should not assume the primary figures transfer. A department running a pilot in an essay-based subject should expect the skill map to be the hardest part, and judge the tool on it.

Choosing an adaptive tool for your trust? Compare the field in the best AI education tool guide, book an AI placement diagnostic, see the DATS consulting system, or browse more AI strategy guides.

Written by the Dilr.ai engineering team, practitioners who ship enterprise AI in production. Follow us on LinkedIn for shipping notes, or subscribe via the RSS feed.