TECHNOLOGY WING · ADAPTIVE LEARNING
Adaptive Learning and Intelligent Tutoring in Mathematics
Adaptive systems change the learning route in response to evidence. Intelligent tutoring systems go further by selecting prompts, hints, questions or explanations based on an estimated learner state.
Adaptation is only as good as the state estimate underneath it.
The adaptive loop
Observe → Estimate State → Select Next Task → Observe Response → Update State.
This sounds simple, but each arrow contains a design problem. What evidence counts? Does a wrong answer mean misconception, retrieval failure, careless execution or overload? How much confidence should the system place in one response? Is the next task intended to teach, diagnose or verify?
Singapore example: SLS Adaptive Learning System
Singapore’s SLS currently provides an Adaptive Learning System for Mathematics in Upper Primary and Lower Secondary. MOE describes it as recommending learning modes, resources and practice questions and providing immediate feedback based on student readiness. That is a useful example of route adaptation rather than one universal sequence for every learner.
What an intelligent tutor can adapt
- Problem difficulty and prerequisite depth.
- Worked example versus independent attempt.
- Hint size and timing.
- Representation: numerical, pictorial, graphical, symbolic or verbal.
- Practice spacing and topic mixture.
- Question type: reproduce, recognise, explain, transfer, verify.
- Feedback resolution.
- When to stop remediation and reconnect to current work.
Evidence
Intelligent Tutoring Systems have a long research history. Meta-analyses generally report positive effects on educational outcomes, though impact varies by system, learner population and outcome. A 2025 meta-analysis of 30 studies reported a positive overall effect while noting that effects on specific outcomes such as knowledge acquisition, motivation, performance and problem solving were less uniform. The BTT conclusion is therefore functional: use adaptation where the system has meaningful evidence about learner state, not because “personalisation” sounds desirable.
Three failure modes
| Failure | What happens | BTT safeguard |
|---|---|---|
| Wrong state estimate | The system routes around the real weak link. | Carry confidence and allow teacher/student evidence to override the model. |
| Local optimisation | The learner gets better at the platform’s task family but not broader Mathematics. | Use changed surfaces and external transfer tests. |
| Permanent scaffolding | The adaptive system keeps rescuing the learner before independence develops. | Deliberately reduce hints and verify unassisted performance. |
Fade rule: the adaptive system should reduce support when learner control becomes reliable; adaptation should not become invisible permanent assistance.
References: SLS AI-enabled Features · 2025 ITS meta-analysis
PHASE 4 · ADAPTIVE LEARNING READER GUIDE
Quick Read: when is adaptive learning genuinely useful in Mathematics?
Adaptive learning is useful when the system has enough evidence to change the next task intelligently—and when the learner is still allowed to contradict the model.
A personalised route can be powerful because two students who score 60% may need different next steps. One may need prerequisite repair, one may need harder transfer questions, and one may simply have made a few execution errors. The danger is assuming that a system’s estimate of the learner is automatically correct because it is personalised.
One-sentence answer: adapt the route, not the learner’s identity; keep the state estimate provisional and verify progress outside the adaptive loop.
What an adaptive system is really trying to estimate
- What the learner knows. Which concepts and methods appear stable?
- What is currently retrievable. Knowledge may exist but fail to appear without a cue.
- Which representation is accessible. A student may understand graphically but not symbolically, or vice versa.
- How much support is needed. Worked example, hint, partial completion or independent attempt?
- How difficult the next task should be. Too easy gives little information; too hard can produce noise.
- Whether the learner is ready to reconnect to current work. Remediation should not become a permanent side path.
Each estimate contains uncertainty. A wrong answer can mean misconception, retrieval failure, poor reading, overload or a single slip. One observation rarely justifies a permanent state label.
Three adaptive-routing cases
- Student A misses three fraction questions. The system routes to basic fraction representation. A quick human check shows the student understands fractions but misread a unit conversion. The model should update rather than trap the learner in unnecessary remediation.
- Student B answers routine algebra correctly but fails mixed questions. More routine algebra is not the right adaptation. The learner needs recognition and transfer tasks where the method is not announced.
- Student C needs one hint, then completes the whole route. The next adaptation should often reduce support and test retrieval later rather than keep supplying the same hint automatically.
These cases show why adaptation must be tied to the learning function. Personalisation is not simply changing difficulty; it can change representation, support, spacing, mixture, feedback or the need for a diagnostic probe.
The wrong-state problem
Adaptive systems can become very efficient at sending a learner down the wrong route. That happens when the state estimate is wrong but the system treats it as fact.
| Signal | Possible interpretation | Alternative interpretation |
|---|---|---|
| Slow response | Weak fluency | Careful checking or unfamiliar interface |
| Wrong answer | Concept gap | Execution slip or reading error |
| Needs hint | Does not know method | Retrieval temporarily blocked |
| Fast correct answers | Mastery | Repeated familiar item family |
| Repeated platform success | Transfer | Interface familiarity |
The safeguard is simple in principle: carry confidence, preserve contradictory evidence and allow the teacher or later learner performance to override the earlier state estimate.
A learner model should behave like a hypothesis that gets updated, not a label that gets defended.
Adaptation should eventually become less visible
The goal is not a student who always receives the perfect next task from a machine. The stronger destination is a learner who can increasingly select, monitor and correct their own work.
- High support: system chooses representation, gives bounded hints and keeps difficulty narrow.
- Reduced support: hints shrink, representations vary and the learner carries more method selection.
- Mixed transfer: the system removes topic labels and presents unfamiliar surfaces.
- External verification: the learner performs on paper, in school work or in a different environment.
- Self-regulation: the student begins to identify which weak area needs practice without waiting for the system to decide everything.
Adaptation is successful when it changes the learner, not merely when it changes the sequence.
Teacher override is not a system failure
A teacher may see information the platform cannot: working style, recent school demands, emotional state, a conversation revealing a misconception, or evidence that the student has already repaired the issue elsewhere. Overriding the system can therefore be exactly the right use of the system.
- Override when direct evidence is stronger than the inferred state.
- Override when the learner has become trapped in low-level remediation.
- Override when the recommended task does not match the current syllabus priority.
- Override when high platform performance is not transferring to independent work.
- Override when the support level is preventing productive struggle or independence.
The better design allows teacher correction to update the model rather than treating the override as an exception to be forgotten.
What parents should ask about adaptive learning
- What evidence is the system using to decide the next task?
- Can a teacher inspect or override the learner model?
- Does the system distinguish teaching, diagnosis and verification?
- Are hints and scaffolds reducing as the student improves?
- Can the student perform outside the platform?
- Does the system keep revisiting old capabilities or only move forward?
- Can one wrong response send the learner too far backward?
- Is personal data being collected because it has an educational function?
The strongest evidence is not that the platform says the learner is at a higher level. It is that the learner can use the Mathematics later, in another environment, with less support.
Frequently asked questions
Is adaptive learning better than a fixed sequence?
It can be when learner states genuinely differ and the system has reliable evidence about those differences. A fixed sequence can still be preferable when the instructional progression is already appropriate or when the adaptive model is weak.
Can a student become dependent on an adaptive tutor?
Yes, especially if hints and route selection remain permanently available. Support should reduce and independent transfer should be verified deliberately.
What if the system keeps repeating work my child already knows?
That may indicate an incorrect state estimate, insufficient confidence thresholds or a mismatch between platform evidence and real capability. Teacher review can be important.
Does personalised practice guarantee faster progress?
No. Personalisation can improve fit, but progress still depends on mathematical validity, learner engagement, quality of feedback, transfer and whether the state estimate is correct enough to justify the route.
What is the best independence test?
Ask the learner to solve a changed problem after some delay in a different environment without the adaptive prompts. If the capability survives, the system has stronger evidence that adaptation produced learning rather than platform-specific performance.
The larger idea: the best adaptive system eventually teaches the learner to adapt
At first, the system may decide what to show, when to hint and what to revisit. Over time, a strong learner should begin to internalise those decisions: recognising a weak prerequisite, choosing a useful practice type, spacing revision and deciding when a solution needs verification.
That is the developmental destination. Personalisation starts outside the learner and gradually becomes self-regulation inside the learner.
Adaptive learning is most successful when the student eventually needs less adaptation from the system because the student has become better at adapting their own learning.