What is benchmark variance?

Definition

Variation can come from model sampling, task selection, scoring judgment, service changes and the small number of cases tested. A narrow subset may give very different rankings when only a few results change.

Repeated runs, confidence intervals and larger representative sets help quantify uncertainty. A score should not be treated as a precise capability measurement when its expected variation is large.

ELI5

Benchmark variance describes how much a model's test score might move if the test were repeated or used a different small sample. High variance means the reported number is less stable.

For example, on a ten-task set, one changed result moves the score by ten percentage points. Repeating the test and adding more representative tasks gives a clearer view of whether the apparent lead is dependable.

Frequently asked questions

Why do small AI benchmarks have high variance?

Each individual task has a large effect on the total score, and random model behavior can change a few outcomes.

How can benchmark variance be reduced or reported?

Use more representative tasks, repeat runs and report score distributions or confidence intervals.

Videos explaining benchmark variance