Adaptive learning paths: personalisation for learners
What should you know about an adaptive learning platform?
Adaptive paths sequence content and difficulty per learner from progress data, with mastery measured, not assumed. An adaptive learning platform is a knowledge map, a mastery model and a sequencing engine, and it only pays off when the mastery signal is honest and the lift is proven against a control group.
An adaptive learning platform promises that every learner gets the next right thing: a harder problem when they are ready, a prerequisite when they are stuck, a review when they are about to forget. Most platforms that claim this are doing something simpler, such as branching on a single quiz score. The difference between the two is whether mastery is measured or assumed, and it decides whether the personalisation changes outcomes or just changes the order of the same content.
This article explains what an adaptive learning platform actually contains, how the learning path engine decides what comes next, where the data has to come from, how to measure whether it works, and what a build costs. It is for edtech product leaders, school networks and corporate learning teams who own their own content and want edtech personalisation that survives contact with real learners.
What adaptive learning is, and what it is not
Adaptive learning is the continuous adjustment of what a learner sees, in what order and at what difficulty, based on evidence of what they know. It is not a recommendation carousel of "popular courses", not a fixed set of three tracks chosen by a placement test, and not a chatbot. It needs three parts: a map of the knowledge (skills and their prerequisites), a mastery model that estimates each learner's state on each skill, and a sequencing engine that chooses the next activity given the map, the state and the learning goal. Remove any one of them and you have static content with a personalised label.
Three levels of personalisation
| Level | How it decides | What it needs | Where it breaks |
|---|---|---|---|
| Rule-based branching | If score below threshold, show remedial module | Quizzes with thresholds set by hand | Cannot distinguish a careless slip from a missing prerequisite |
| Mastery-based sequencing | Estimates mastery per skill from every attempt; picks the next skill whose prerequisites are met | A skill map, tagged content, attempt-level events | Bad tags or a wrong prerequisite edge misroute learners; needs content coverage per skill |
| Model-driven adaptation | Learns from all learners which activity most improves mastery for a state like this one | Everything above, plus volume and a controlled experiment | Cold start for new content; overfits to engagement if the objective is wrong |
Most organisations should start at the second level. Rule-based branching is what they already have; model-driven adaptation needs a volume of learners and an experiment discipline that few have on day one. Mastery-based sequencing delivers most of the value and produces the data the third level needs.
The knowledge map
The map is a graph of skills with prerequisite edges: two-digit subtraction depends on place value; balancing equations depends on the mole concept. Every piece of content (a video, a worked example, a practice set) is tagged with the skills it teaches and the skills it assumes. This is the unglamorous part of the build and the one that decides everything else. We build it with your subject experts, validate it against attempt data (if learners who "mastered" A keep failing B, the edge or the tags are wrong), and version it, because curricula change.
Measuring mastery, not assuming it
A single quiz score is a weak signal: it mixes knowledge, guessing, carelessness and the difficulty of the specific items. A mastery model uses every attempt, the difficulty of each item, the time since the last practice and the pattern of errors to estimate the probability that the learner has the skill, and it decays that estimate over time so review is scheduled before forgetting. Bayesian knowledge tracing and its successors are well studied; the research literature is accessible on arXiv for teams that want the details. The practical rule is that mastery is a probability with a threshold, updated continuously, never a tick box set once.
Honest signals
- Attempt-level events with item identity, correctness, response time and hints used
- Item difficulty calibrated from all learners, not set by the author
- Distinguishing a slip (one wrong after many right) from a gap (consistent errors on a skill)
- Decay, so a skill mastered a month ago is queued for review
- Teacher overrides, recorded as evidence rather than silently replacing the model
The sequencing engine
Given the map and the mastery estimates, the learning path engine picks the next activity. The default policy is simple: among skills whose prerequisites are above threshold and which are on the path to the learner's goal, choose the one with the lowest mastery; within it, choose an activity at a difficulty slightly above the learner's current estimate; interleave scheduled reviews. Constraints from the institution sit on top: the syllabus order for a class, an exam date, a teacher's assignment for the week. The engine explains each choice in plain language ("because you had trouble with fractions last week") because learners and teachers trust what they can see, and a teacher can override any choice from the class view without the engine fighting back.
The engine is a service, called by the app when a learner finishes an activity, and it is the same architectural shape as the personalisation engines we build for commerce: an event pipeline, a state store, a policy and an explanation. Where an LLM helps is in generating the explanation, choosing among equivalent activities by learner preference, and, paired with a grounded tutor, explaining the concept itself.
Proving it works
Engagement is the wrong objective; a platform can maximise time-on-task while teaching nothing. The objective is mastery gain per unit of time, measured on assessments the learner has not practised on, compared against a control group that gets the standard sequence. Run the comparison for at least a term with learners randomised at class or cohort level. Report the result by segment, because adaptive sequencing often helps struggling learners most and advanced learners least, and that pattern is the one schools care about. Our post on personalisation lift and controlled tests sets out the method.
A worked example
A K-12 edtech company with a large library of practice questions and videos had a "personalised" mode that branched on chapter test scores. Learners who failed a chapter test were sent back to the start of the chapter regardless of which concept they had missed, and many left. We built a skill map with the company's teachers, tagged the existing content against it, replaced the chapter score with a per-skill mastery estimate updated on every attempt, and introduced a sequencing service that targeted the weakest prerequisite rather than the whole chapter, with a one-line explanation shown to the learner. A cohort-level comparison over a term measured mastery on unpractised assessments; the company also gained a content-gap report showing skills with too few activities. The onboarding side of the same problem is covered in personalisation for education and SaaS onboarding.
Team and timeline
A mastery-based adaptive platform on existing content is typically a data engineer for the event pipeline, an ML engineer for the mastery model and sequencing policy, a full-stack engineer for the service and the learner-facing explanation, and your subject experts for the skill map, over ten to sixteen weeks. The first four weeks build the skill map and tag content; the middle weeks build the model and engine and validate them against historical attempt data; the last weeks integrate with the app, launch to a pilot cohort with a control group, and hand over. A Sprint Zero is the right first step if you are unsure whether your event data supports mastery modelling at all. Personalisation engine builds start from $21,000 / ₹13.6L and AI/ML development from $17,500 / ₹11.2L; the pricing page has current figures, and our education page describes related work.
Before you start: a checklist
- Confirm you capture attempt-level events, not just completions and scores
- Draft the skill map with subject experts and agree who maintains it
- Tag content by skills taught and skills assumed; find the skills with no content
- Decide the mastery threshold and how teacher overrides are recorded
- Write the sequencing constraints: syllabus order, exam dates, teacher assignments
- Choose the outcome measure and design the control group before launch
- Settle consent and retention for learner data, especially for minors
Glossary
- Skill map: a graph of skills and the prerequisites between them
- Mastery model: an estimate of the probability a learner has each skill, updated on every attempt
- Sequencing engine: the policy that picks the next activity from the map, the mastery state and the goal
- Interleaving: mixing review of older skills into new practice to reduce forgetting
- Cold start: the period when a new learner or new content has too little data for confident adaptation
- Control group: learners who get the standard sequence, against whom the adaptive group is compared
Related reading
Pair this with AI tutors grounded in your curriculum for the explanation layer, event pipelines as the foundation of personalisation for the data side, and learner data protection for consent and access.
Personalisation for learners is worth building when mastery is measured honestly and the lift is proven; anything less is a reordered playlist.
Frequently asked questions
Do we need machine learning for adaptive learning?
▾
Not at first. A skill map, a probabilistic mastery estimate and a simple sequencing policy deliver most of the value. Learned policies come later, once you have volume and an experiment discipline.
How much learner data does an adaptive platform need?
▾
Attempt-level events (item, correctness, time, hints) for each learner. A few thousand learners with a term of attempts is enough to calibrate item difficulty and validate the skill map.
How do we know the adaptive path is better than the fixed one?
▾
Run a cohort-level controlled comparison for a term and measure mastery gain on assessments learners have not practised on, reported by segment. Engagement alone proves nothing.