Bayesian Knowledge Tracing, explained without the maths

If you have used a flashcard app, you know the feeling: you have seen a card thirty times, you can read it back without thinking, and the moment someone phrases it differently you blank. Repetition is not the same as understanding, and the easiest way to fool a study system is to confuse the two. Bayesian Knowledge Tracing was designed to keep them apart.

BKT keeps a single number for each concept you are studying: the probability that you have actually mastered it. Every answer you give is evidence. A correct answer pushes the probability up; an incorrect one pulls it down. Crucially, the model accounts for guessing and slipping, so a lucky correct answer does not retire a concept, and an unlucky wrong one does not bury you. The maths is unspectacular. The behaviour is not.

We retire a concept once its mastery probability clears 0.90 and bring it back into rotation only if a downstream question reveals a regression. The result is a study queue that shrinks as you actually learn, rather than one that just rolls forward in time. Spaced repetition still has a role, but it sits underneath the BKT layer, not on top of it.

The four numbers BKT tracks

Bayesian Knowledge Tracing represents your grasp of a single concept as one hidden probability, and it updates that probability using four parameters. None of them are mysterious once you see what each is for.

• Prior, p(L0), the starting belief. Before you answer anything, how likely is it you already know this? We set it to 0.25 for adult professionals, which says: assume most people arrive not knowing a given concept, and make them prove otherwise. A learner who answers correctly from the first question climbs out of that prior fast.
• Learn, p(T), the transition rate. Each time you engage with a concept, there is a probability you actually learn it in that moment, whether or not the answer was right. We use 0.30. This is what lets mastery rise steadily with practice rather than only on correct answers.
• Guess, p(G), the lucky-correct rate. On a four-option multiple choice, you can be right by luck about a quarter of the time, so we set guess to 0.25. This is why one correct answer does not retire a concept: the model knows a slice of that success is chance.
• Slip, p(S), the careless-wrong rate. Even when you know something, you sometimes misread the question or mis-click an option. We set slip low, at 0.08, because a wrong answer from someone who genuinely knows the material is rarer than a lucky guess from someone who does not.

Guess and slip are the two parameters that make BKT honest. Without them, every correct answer would be treated as proof and every wrong one as ignorance. With them, the model discounts luck in both directions.

How one answer changes the belief

Every question you answer is evidence, and BKT folds it in using Bayes' rule. You do not need the algebra to understand the behaviour, but the shape is worth seeing.

When you answer correctly, the model weighs how likely a correct answer is if you truly know this (scaled by 1 minus slip) against how likely it is if you are guessing (scaled by guess), and it raises your mastery probability accordingly. When you answer incorrectly, it runs the mirror image: a wrong answer is far more likely from someone who does not know the concept than from someone who slipped, so your probability falls.

Then, after either outcome, a second step nudges the probability up by the learn rate, on the reasoning that the act of practising taught you something. This is the subtle part. Even a wrong answer can leave you slightly better off than before, because wrestling with a concept and getting corrected is itself a form of learning. The net effect is a number that moves fast when the evidence is strong and clear, and cautiously when it is mixed, which is exactly how a good tutor updates their read of a student.

Why we retire at 0.90, not 'you saw it three times'

Most study tools use a counter: three correct reviews and a card is done. The problem is that a counter cannot tell a confident, repeated correct answer apart from three lucky guesses, and it has no way to represent partial knowledge. BKT retires a concept only when its mastery probability crosses 0.90, a genuine 90% model-estimated chance that you know it, and that threshold behaves very differently from a count.

Because guess and slip are baked in, reaching 0.90 usually takes a short run of correct answers on progressively harder framings, not one fluke. A single lucky correct answer barely moves a low prior; it takes consistent success to clear the bar. And crucially, mastery is not permanent. If a later question that depends on this concept reveals a regression, the probability drops and the concept re-enters rotation. The study queue shrinks as you actually learn and grows back exactly where you slip, which is the whole point. You can watch this happen on your dashboard, where each domain's mastery is the average of its concept-level probabilities, and on the knowledge map, where concepts shift from red to green as they clear the threshold.

Where spaced repetition sits underneath

Reaching 0.90 is not the end of a concept's life; it is the moment it moves from active learning into maintenance. That maintenance layer is SM-2, the spaced-repetition algorithm that decides when a mastered concept comes back for a check.

The schedule expands as you keep succeeding: the first review lands a day after mastery, then after 3 days, 7, 14, and 30. Each successful review pushes the next one further out, so well-known concepts consume less and less of your time. An easiness factor, starting at 2.5 and adjusting per concept, stretches or compresses those gaps based on how hard you find each one.

The important design decision is the ordering: SM-2 sits underneath BKT, not on top of it. A miss on a scheduled review does not just delay the next card. It feeds back into the BKT layer, drops the mastery probability, resets the repetition counter, and drags the concept back into active learning. Spaced repetition schedules the maintenance; Bayesian Knowledge Tracing remains the source of truth for whether you actually know something. That is why a regression on a 30-day review is treated as seriously as a wrong answer on day one.

What BKT deliberately does not model

A forty-year-old algorithm survives because it is honest about its scope. Classic BKT tracks one concept at a time and assumes, in its simplest form, that you do not forget between encounters, which is precisely why we pair it with spaced repetition to handle decay, and why a regression on a downstream question can pull a mastered concept back. It also treats each concept as independent, so the knowledge map does the work BKT does not: it enforces prerequisites, so you are never drilled on a concept whose foundations you have not yet cleared.

We also tune the parameters to the audience. The defaults above are calibrated for adult professionals sitting a technical exam, and they shift per learner over time as real data accumulates. A guess rate of 0.25 is right for a four-option question; it would be wrong for a free-text answer. The model is simple on purpose, and that simplicity is what makes its one output, a single mastery probability per concept, trustworthy enough to act on.

What this feels like when you study

In practice, none of this maths is visible. You answer scenario questions, Archie the tutor works through the ones you miss, and the queue quietly reshapes itself around your weak spots. A concept you keep getting right disappears within a session or two; one you keep misjudging stays in front of you until the evidence says you have it.

The result is the opposite of a flashcard deck that rolls forward on a fixed calendar regardless of what you know. Your progress view shows a study load that shrinks as your true mastery rises, and a predicted exam score that is the weighted sum of your domain masteries rather than a count of cards seen. Repetition fools a counter. It does not fool a probability that already knows how often you get lucky. That, four decades on, is still why Bayesian Knowledge Tracing is the right way to track what you actually know.