SoulTraceSoulTrace
Adaptive Testing

Information Gain

Also called: expected information gain, uncertainty reduction

Information gain measures how much an observation is expected to reduce uncertainty. In adaptive testing, the system can predict the possible answers to each eligible question, estimate the posterior uncertainty after each answer, and choose the question with the greatest average reduction. The actual gain is known only after a response is observed.

Reviewed July 14, 2026 · 2 min read

Table of Contents

When faced with a complex decision, I prioritize a methodical approach over intuitive leaps.

Neutral
Strongly DisagreeStrongly Agree
/en/new-test?utm_source=glossary&utm_medium=internal&utm_campaign=glossary-assessment&utm_content=information-gain-after-definition-cta-assessment-test

Expected reduction in uncertainty

Before asking a question, an adaptive system does not know which response will occur. It therefore considers every plausible response, calculates the uncertainty that would remain, and weights those outcomes by their predicted probability.

Expected information gain is the current uncertainty minus that expected remaining uncertainty. A question scores highly when its possible answers would meaningfully rearrange or concentrate the current probability distribution.

A simple example

If the model is split evenly between two profiles, a question that those profiles usually answer differently can be valuable. Either response would favor one possibility and reduce uncertainty. A question both profiles answer similarly provides little separation.

When one profile already has overwhelming probability, many questions have low expected gain. That can trigger a stopping rule.

Information gain vs item information

The terms are related but come from different formulations. IRT item information describes measurement precision at positions on a latent-trait scale. Expected information gain often compares entire probability distributions before and after a possible observation. Both can drive adaptive selection.

Why gain is not the only objective

Always choosing the largest mathematical gain can overuse certain content or neglect important dimensions. Real systems combine it with content coverage, exposure, fairness, user experience, and minimum evidence requirements. Statistical efficiency serves the construct definition; it does not replace it.

Go deeper: Information gain in Soultrace

Sources

Related terms

EntropyEntropy is a mathematical summary of uncertainty in a probability distribution. In an adaptive personality assessment, entropy is higher when probability is spread across several plausible results and lower when it is concentrated on fewer possibilities. A question can be valuable when its expected answers would reduce entropy.Adaptive Question SelectionAdaptive question selection is the rule an adaptive assessment uses to choose the next item from its bank. The algorithm evaluates what is currently uncertain and selects a question expected to improve the estimate while respecting constraints such as content coverage, item exposure, fairness, and maximum test length.Bayesian InferenceBayesian inference is a method for updating probabilities when new evidence arrives. An assessment begins with prior probabilities over possible trait values or profiles, evaluates how likely each answer would be under those possibilities, and produces updated posterior probabilities. Repeating this process makes uncertainty explicit as the test progresses.