What is Adaptive testing?
Also called Computerised adaptive testing, CAT
Adaptive testing is a test design in which each question is chosen from a calibrated bank on the basis of the answers already given: a correct answer leads to a harder question, an incorrect one to an easier question. It reaches a given level of precision in far fewer items than a fixed test.
What it is for
A fixed test spends most of its items on the wrong people. A very strong candidate answers the easy half correctly and learns nothing about themselves; a weak candidate guesses at the hard half. Adaptive testing removes that waste by aiming each item at the current estimate of the person's ability, where the answer carries the most information.
The practical consequences are shorter tests at equal precision, less scope for copying between candidates because no two people see the same sequence, and a score that is comparable across people who answered entirely different questions.
What it requires before it works
Adaptive delivery is not a setting you switch on. It requires an item bank whose items have been calibrated on a common scale — usually under item response theory — with enough responses per item for the parameters to be stable, plus content-balancing and exposure-control rules so that the algorithm does not keep serving the same well-behaved items to everybody.
A test described as adaptive without a calibrated bank behind it is usually branching: a rule that sends you to section B if you scored badly in section A. Branching is a legitimate design, but it is not adaptive testing and the scores it produces are not equatable across paths.
The bank is not used evenly, and that is the cost of the efficiency
Every description of adaptive testing, including the one above, says the algorithm must be stopped from serving the same well-behaved items to everybody. Almost none of them says how unevenly it would otherwise behave, and the number is the part a buyer can act on.
A published set of adaptive-testing simulations run in R, reported in the Journal of Educational Measurement literature and available in full text at ERIC, makes it concrete. In one run — a 300-item pool, a 45-item test, 400 simulated examinees, with an exposure cap already in force at 0.2 — 135 of the 300 items hit that cap, meaning each of them was administered to a fifth of everybody who sat, while three items in the pool were never administered to anyone. A second run reports 63 items at the maximum exposure rate and 34 items in the bank never used at all.
Two consequences follow from the same fact. The first is security: if a small subset of the bank carries most of the testing, that subset is what leaks first, and exposure control is a mitigation rather than a fix. The second is economics. The pool in that illustration is roughly seven times the test length, and the authors present it as an illustration rather than a recommendation — but it is a useful floor to argue from when a vendor proposes adaptive delivery over a bank of a hundred items.
The question to ask, and it is answerable: what is the maximum exposure rate in your pool, how many items are at it, and how many items have never been administered? A vendor who cannot produce those three numbers has not measured the thing their own exposure-control rule exists to manage.
What an exposure cap of 0.2 actually means for one bank
- Pool: 300 items. Test length: 45 items. Candidates: 400.
- Items at the 0.2 exposure cap: 135, or 45 per cent of the pool.
- Each capped item is therefore seen by up to 0.2 × 400 = 80 candidates.
- Items never administered to anyone: 3.
- Total item administrations: 400 × 45 = 18,000 — spread across 297 items, not 300.
Figures from the simulation cited in Sources, not from an AssessAll bank. The point is the shape: an adaptive algorithm concentrates delivery by design, and the cap sets how much concentration you have agreed to.
Not the same as branching
Branching routes a candidate between fixed blocks on a rule. Adaptive testing selects the next single item from a calibrated bank to maximise information about the person. Ask which one a vendor means, and ask how many responses per item the calibration used.
Sources
- Wainer et al., Computerized Adaptive Testing: A Primer (2nd ed., 2000) — the standard reference
- Computer Adaptive Testing Simulations in R — full text at ERIC (EJ1246477), read 24 September 2026. Source of the 300-item pool, 45-item test, 400-examinee run in which 135 items reached the 0.2 exposure cap and three were never administered, and of the second run with 63 items at maximum exposure and 34 never used.
Read next
Related terms
More terms beginning with A
Check a selection process against the four-fifths rule
Free, no signup, computed in your browser — with the remedy, not just the verdict.
Last reviewed