
TL;DR
- Adaptive learning is software that reads each student's signals and changes what comes next, from the next question to the route through the course.
- Most engines only read observed signals, such as right or wrong answers, response time and hints used.
- The signals that explain a struggle are often declared in conversation. A doubt sent on WhatsApp, a stated exam date or a reason for missing class rarely reaches the engine.
- Mastery learning works when the mastery rule is strict enough. One correct answer proves very little.
- Join observed and declared signals into one learner view, then map each signal to a specific curriculum move.
Two students miss the same fractions question on Tuesday night. One rushed it. The other messaged her tutor an hour earlier to say she has never understood fractions and her board exam is in six weeks.
Adaptive learning is an approach to teaching in which software reads each student's signals, such as answers, response time, hints used and stated confidence, and changes what comes next in real time. That can be the next item, the difficulty, the review schedule or the route through the curriculum. Mastery-based adaptive learning moves a student forward only once those signals show a skill is secure.
Most adaptive learning engines would treat those two students the same way, because the only signal they hold is the wrong answer. This guide covers which signals an engine should read, how strict a mastery rule needs to be, how signals become curriculum moves and how to build the learner view that feeds it all.
What is adaptive learning and how does it work?
Adaptive learning works by running three components in a loop. Each one can be simple, and most of the quality comes from how well they connect.
- A learner model. The engine's current estimate of what a student knows, skill by skill.
- A content model. Lessons, items and practice sets tagged to the skills they teach or test.
- Decision rules. The logic that reads the learner model and picks the next piece of content.
Every time a student answers, the learner model updates and the rules run again. In the simplest systems the rules are written by hand. In more advanced ones a statistical model, such as knowledge tracing, estimates the probability that a skill is mastered.
The three kinds of adaptivity
Adaptive learning technology changes one of three things.
- Content adaptivity changes the explanation or hint a student sees.
- Sequence adaptivity changes the order of topics, skipping what a student already knows.
- Assessment adaptivity changes the next question based on the last answer.
A real-time curriculum uses all three. The engine can pick an easier item, swap the explanation and send a review topic back into the queue within the same session.
An adaptive engine can only be as smart as the signals that reach its learner model.
What signals does an adaptive learning engine actually read?
An adaptive learning engine usually reads observed signals, the traces a student leaves while working inside the platform. They are cheap to collect and arrive in real time.
- Correctness. Right or wrong, and which wrong answer was chosen.
- Response time. Very fast wrong answers often mean guessing. Very slow right answers can mean shaky recall.
- Hints and retries. How much help a student needed before getting there.
- Session patterns. Time on task, drop-off points and gaps between sessions.
These signals say what happened. They rarely say why it happened, and the why is what decides the right curriculum move.
The declared signals most engines never see
Declared signals are the things a student or parent tells you directly. They usually live in conversations outside the learning platform, in WhatsApp threads, web chat, support tickets and tutor calls.
- Doubts. "I still do not get how to balance this equation" names the exact skill gap.
- Stated confidence. A student who says she is lost needs a different move from one who made a careless slip.
- Goals and dates. An exam in six weeks changes how much review fits before new material.
- Reasons for absence. Illness, travel or a family event explain a gap that the engine would read as disengagement.
- Language and channel preference. A student who asks questions in Hindi on WhatsApp may need hints in that language, on that channel.
A doubt on WhatsApp at 11 pm is often the earliest and clearest signal that a skill is not secure. We cover how those same signals predict dropout in our piece on student retention analytics.
Observed signals tell the engine that a student struggled. Declared signals tell it why.
What counts as mastery in mastery learning?
Mastery learning is a teaching model, proposed by Benjamin Bloom in 1968, in which students move to the next unit only after showing they have mastered the current one. Adaptive learning platforms automate the check, and the strictness of that check matters more than most teams expect.
A meta-analysis of 49 studies by Kulik and Kulik, published in 1987, found that mastery testing helps students learn, and that the size of the benefit depends on how demanding the mastery criterion is. A broader review of 108 controlled evaluations by Kulik, Kulik and Bangert-Drowns, published in 1990, found that mastery learning programmes improved exam performance from upper elementary school through college.
Why a single correct answer proves little
A single correct answer can be a lucky guess or a memorised pattern. A workable mastery rule combines several conditions, for example these.
- Consistency. Several correct answers on different items for the same skill.
- Speed. Answers at a normal pace, without heavy hint use.
- Retention. A correct answer on a spaced check a few days later.
- No open doubt. No unresolved question about the skill in any channel.
The last condition is the one almost nobody writes. A student can pass three items in a row and still message a tutor that night to say she guessed.
How do signals turn into real-time curriculum moves?
Signals turn into real-time curriculum moves through explicit rules that pair a pattern with an action. Writing the rules as a table keeps them reviewable by teachers and easy to test.
| Signal | Rule | Curriculum move |
|---|---|---|
| Three fast wrong answers | Likely guessing | Slow the pace, add one worked example |
| Correct but slow, heavy hint use | Fragile skill | Add two practice items, schedule a spaced check |
| Doubt sent on WhatsApp about a skill | Declared gap | Reopen that skill, send a short explainer on the same channel |
| Exam date stated as six weeks away | Time pressure | Weight review of weak prior topics over new content |
| Three days with no session after a stated illness | Explained absence | Resume with a recap, no disengagement alert |
| Three days with no session and no reason | Possible disengagement | Send a check-in message, offer a tutor slot |
| Mastery rule met on a skill | Skill secure | Skip remaining practice, unlock the next unit |
Three of the seven rows depend on a declared signal, and a fourth depends on knowing that none was given. Without them, the engine would treat an ill student as a disengaged one and a confused student as a careless one.
Every curriculum move should trace back to a named signal and a written rule.
What does the evidence say about adaptive learning?
The evidence on adaptive learning and intelligent tutoring is positive, with effect sizes that vary by subject, test type and setting.
- A 2016 meta-analysis by Kulik and Fletcher in the Review of Educational Research pooled 50 controlled evaluations and found intelligent tutoring raised test scores by a median 0.66 standard deviations, roughly from the 50th to the 75th percentile. Gains were much smaller on standardised tests than on tests the researchers built themselves.
- A November 2025 meta-analysis by Leite and colleagues pooled 18 US K-12 studies, with 77 effect sizes across 11 tutoring systems, and found an average effect of g = 0.271. Effects were lower in studies that included rural schools.
- The Stanford Tutor CoPilot trial, published in October 2024, covered 900 tutors and 1,800 K-12 students in Title I schools. Students whose tutors had real-time AI guidance were 4 percentage points more likely to master topics, and up to 9 points more with lower-rated tutors.
- A Harvard randomised trial by Kestin and colleagues in Scientific Reports, June 2025, with 194 introductory physics students, found median learning gains with an AI tutor were more than double those from in-class active learning.
Two lessons follow. Adaptivity helps on average, and the size of the gain depends on design and context, so every programme should test against a non-adaptive control group.
What do adaptive learning examples look like for real students?
Adaptive learning examples make the signal layer concrete. These three students are illustrative composites.
Maya, preparing for a board exam
Maya passes most algebra items but slowly, with hints. On WhatsApp she tells her tutor the exam is in five weeks and she panics on word problems. The engine flags algebra word problems as fragile, adds spaced practice and moves new geometry content back a week.
Daniel, a working adult on an online course
Daniel stops logging in for nine days. Read alone, that looks like churn. In a chat he had mentioned a work trip, so the system pauses alerts and greets him with a ten-minute recap when he returns.
Chloe, who already knows the material
Chloe answers every item in the first two units quickly and correctly, and passes the spaced checks. Her mastery rule is met, so the engine skips the remaining practice and unlocks the next unit, which keeps her from getting bored.
How do you build a unified learner view for adaptive learning?
A unified learner view for adaptive learning is a single timeline per student that holds both observed platform events and declared conversation signals, keyed to the same skills. Building one is mostly a data job.
- Tag content to skills. Every item and lesson maps to one or more named skills.
- Capture platform events. Stream answers, timings and hints from the learning platform through its API or webhooks.
- Capture conversations. Pull WhatsApp, web chat, email and call notes into the same student record.
- Extract declared signals. Turn each message into structured fields, such as skill, doubt, confidence, date and reason.
- Stitch identity. Match the WhatsApp number, the email address and the platform login to one student.
- Run the rules and measure. Apply the signal-to-move table, then compare outcomes against a holdout group.
Zigment, a Conversational Revenue Orchestration Platform for education and edtech growth teams, covers steps 3 to 5 and the outreach that follows. The Conversation Graph keeps every WhatsApp, web chat and email exchange on one timeline per student, extracts intent and sentiment from each message, and triggers the check-in, the explainer or the tutor handoff. The learning platform keeps running the curriculum.
One online education platform that runs its WhatsApp and web chat journeys on Zigment saw a 3x lift in conversion rate across more than 100,000 conversations a month. That number is about enrolment, and the same persistent timeline is what makes declared learning signals usable.
How is adaptive learning different from personalized learning?
Adaptive learning is one way to deliver personalized learning. Personalized learning is the broad goal of tailoring pace, content and support to each student, and a teacher can do it without software. Adaptive learning is the software-driven version that changes the path automatically from live signals.
We explain the wider idea, and why so much of it still feels generic, in our guide to personalized learning.
Where should you start with adaptive learning?
Start with one course, one skill map and the seven rows in the signal table above. Add declared signals early, because they fix the misreads that make students lose trust in an adaptive path.
Your students are already telling you where they are stuck, on WhatsApp, in chat and on calls. Is your curriculum listening? See how the Conversation Graph keeps every one of those conversations on one timeline.