Sequencing: What Comes Next and When to Review

Imagine a learner finishes a problem. Behind the scenes, your platform has to make a quiet, constant decision: what should this person see next? A brand-new idea? A review of something they learned last week? Or a patient re-teaching of the exact thing they just got wrong? Getting that one decision right, thousands of times per learner, is the difference between a tutor that genuinely moves someone forward and a chatbot that just answers whatever is in front of it.

This chapter is about that decision: the adaptive loop at the heart of the platform. Let me define the term plainly. An adaptive loop is a repeating cycle where the tutor checks where the learner is, picks the best next step, teaches it, then checks again, forever adjusting. It never assumes; it keeps measuring.

The four-beat loop

Every good tutoring session, human or machine, runs the same rhythm. We can write it as four beats that repeat:

   +-----------+      +-------------+      +---------+
   |  ASSESS   | ---> | PICK NEXT   | ---> |  TEACH  |
   | where are |      | new / review|      | explain,|
   | they now? |      | / re-teach? |      | quiz, do|
   +-----------+      +-------------+      +---------+
        ^                                       |
        |            RE-ASSESS                  |
        +---------------------------------------+
              did it actually stick?

1. Assess — read the learner's current state from their answers (Chapter on learner models covered how we estimate this, skill by skill).
2. Pick next item — choose between three kinds of move: a new concept, a spaced review of something older, or remediation (re-teaching a weak spot).
3. Teach — deliver that item: an explanation, a worked example, a hint, or a question to answer.
4. Re-assess — use their response as fresh evidence, and the loop turns again.

The art is almost entirely in beat 2. So let's slow down and look at how the tutor chooses among the three moves.

The three kinds of "next"

1. A new concept (move forward)

You only introduce something new when the learner is ready for it. "Ready" has a precise meaning here: every prerequisite — the earlier skill a new skill is built on top of — is already mastered. Subjects are not flat lists; they are dependency maps. You can't teach solving fraction equations to someone who can't yet add fractions. Systems like ALEKS call the set of skills whose prerequisites are all satisfied the "outer fringe": the things the learner is genuinely ready to learn right now. The next new concept comes from that set, never from skills floating above an unfilled gap.

2. A spaced review (refresh something fading)

Memory leaks. Hermann Ebbinghaus showed in 1885 that we forget roughly half of new material within an hour and 70–80% within a day, unless we revisit it. So the tutor must resurface old items just before they're about to be forgotten — the moment when a successful recall strengthens memory most. This is the engine of spaced repetition (covered in depth in the spacing-algorithms chapter): each item gets its own schedule, expanding after each success (1 day, then 3, then a week, then a month).

3. Remediation (fix what just broke)

When a learner fails, the naive move is to drill the failing skill harder. Often that's the wrong move. If someone keeps missing equation problems because their fractions are shaky, more equation practice won't help — you have to trace the failure back to the root prerequisite and teach that. Remediation means diagnosing the real cause, not just repeating the surface symptom.

Keeping the learner in the "just-right" zone

There's a single idea that governs how hard the next item should be: the Zone of Proximal Development (ZPD), from psychologist Lev Vygotsky. In plain words, it's the sweet spot between what a learner can already do alone and what they can't do even with help. Below the zone, tasks are boringly easy. Above it, they're hopelessly hard. Inside it, the learner struggles a little but succeeds with light support — and that's exactly where learning happens.

The same idea appears in motivation research as flow (Mihaly Csikszentmihalyi): the absorbed state you reach when challenge and skill are matched. Too much challenge breeds anxiety; too little breeds boredom. Your tutor's biggest advantage over a fixed textbook is that it can measure performance every turn and nudge difficulty to stay in this channel.

 hard  | ANXIETY (too hard) ....
       |                  ...
 chal- |              [ FLOW / ZPD ]   <- aim here
 lenge |          ...
       | ... BOREDOM (too easy)
 easy  +-----------------------------
        low                    high
                  skill

The decision policy, written out

Here is the actual rule the platform follows on each turn. It's the payoff of everything else: the learner model tells you what's mastered, the dependency map tells you what's reachable, and the ZPD bounds how big a step to take.

Notice the ordering instinct: reviews and remediation usually come before new material. There's no point pouring water into a bucket that's leaking from a hole you could patch first.

Balancing new material against review

This is the central tension. Push too much new content and the learner's earlier knowledge quietly rots; they "finish the course" and remember none of it. Schedule too much review and they crawl, never reaching new ground, and disengage. The honest target is not same-day quiz scores — those reward cramming — but retention measured weeks later. A tutor that spaces reviews will look slightly worse on today's quiz and dramatically better on a surprise test next month.

Knowing when to stop teaching an item

The loop also needs an exit rule: when is a skill "learned" so the tutor can move on? The classic convention from Bayesian Knowledge Tracing is to treat a skill as mastered when the estimated chance the learner knows it crosses about 95%. Set that bar too low and learners advance on shaky foundations that collapse on later, dependent skills. Set it too high and they're trapped in pointless over-practice and quit. One subtlety: a "60% mastered" reading can mean two very different things — "we're unsure, barely tested them" (gather more evidence) or "they reliably half-know it" (teach more). The right next move differs, so don't let one number blur the two.

A bonus from mixing: interleaving

Once basics are in place, deliberately interleave — mix different problem types together rather than drilling one to exhaustion before the next. Blocked practice (all of type A, then all of B) feels smoother, but Taylor and Rohrer found interleaving roughly doubled later test scores, because real tests force you to choose the right approach, not just execute a known one. As a bonus, mixing automatically spaces each type out. Your sequencing engine gets two wins from one decision.

The whole loop, honestly

Put together, the tutor is never just answering. On every turn it estimates mastery from the latest answer, checks whether anything is due for review or needs remediation, otherwise picks the next reachable new concept, sizes the step to the learner's zone, teaches it, and reads the response as fresh evidence. The decision is principled — graph (what's reachable) filtered by the learner model (what's needed) bounded by the ZPD (the right-sized step) — not "whatever the model felt like saying."