Learn

How Many Questions Does an Adaptive Test Need?

Published October 31, 2026

It depends on how precise the score has to be, how many questions the pool holds and how good those questions are. In our simulations, a well-stocked pool reached a standard error of 0.3 in a median of 18 questions. Halving the pool nearly doubled that, and a pool of weak questions needed four times as many.

Adaptive tests stop on precision, not on a count

A fixed-form test ends when the questions run out. An adaptive test ends when its estimate of the test taker's ability is precise enough. Precision is measured by the standard error (SE) of the ability estimate, θ. Every answered question adds information and shrinks the SE, and the test stops once the SE falls below a target. Most tests also set a maximum number of questions, so nobody sits an endless test, and some set a minimum, so nobody is scored on three answers. If you are new to how questions are picked, start with what computerized adaptive testing is.

So the real question is which SE to aim for. On a scale where test takers' abilities have a standard deviation of 1, reliability is roughly 1 − SE². An SE of 0.3 is a reliability of about 0.91.

Questions needed for each precision target

500 simulated test takers, with abilities drawn from a normal distribution, took an adaptive test from a 200-question pool, capped at 80 questions.

Precision gets expensive fast. Going from SE 0.4 to 0.3 doubles the test, from 9 to 18 questions. Going from 0.25 to 0.2 more than doubles it again, to 65. Past that point the test starts using up the questions that suit each test taker, so each extra question helps less. Six test takers never reached 0.2 within 80 questions.

The pool matters as much as the target

The same 500 test takers with a target SE of 0.3, drawn from pools of different sizes with the same spread of difficulty and discrimination:

With 30 or 50 questions spread across the whole ability range, nobody reached SE 0.3, even after answering every question in the pool. Too few of them sit near any one test taker's level. The adaptive method does not create information; it only spends the pool's information well.

Question quality matters too. Discrimination (a) is how sharply a question separates people just below its difficulty from people just above it. Three 200-question pools, differing only in the range of discrimination:

A pool of weakly discriminating questions needed 76 questions for the same precision that a strong pool reached in 15. That is because a question's information grows with the square of its discrimination.

Rules of thumb

• For a low-stakes check, such as a quick placement or practice quiz, SE 0.4 to 0.5 is usually enough; in our simulation that took about 6 to 11 questions.
• For a score you will report, SE 0.3 is a common target. In our simulations that took about 15 to 30 questions, depending on the pool.
• Set a maximum of about twice the median you expect, so test takers whose answers are hard to pin down still finish.
• If tests run long, look at the pool before relaxing the target. Adding well-discriminating questions near the levels where tests run long shortens them without giving up precision.

How these numbers were made

All runs used the open-source adaptive-question-selector engine: the 2PL model, maximum-likelihood ability estimates and maximum-information question selection. Difficulties were drawn uniformly from −3 to +3. Discrimination was drawn from 0.5 to 2.0 unless stated otherwise. Abilities came from a normal distribution with mean 0 and standard deviation 1, clipped at ±2.5, and answers were simulated from the model. A test taker counts as reaching a target only after giving at least one right and one wrong answer.

Real tests need more questions than this. Question parameters are estimated from data rather than known exactly, and real tests add rules for content coverage and for how often each question is shown. Use these figures to compare choices, not as a promise for a particular test.

Hosted version

The AdaptiveTest API runs adaptive sessions over HTTP. You create a test, start a session and submit answers, and each response returns the next question and the current ability estimate. See the developer page for the endpoints and how to get access.