How LLMs Are Made / Glossary

AI benchmark

An AI benchmark is a fixed set of tasks with checkable answers, such as exam questions, maths problems or coding tasks, used to measure and compare how well models perform.

Well-known benchmarks include MMLU, with multiple-choice questions across 57 subjects, HumanEval, with 164 programming problems checked by running tests, and GSM8K, a set of about 8,500 grade-school maths word problems.

A score only means something under fair conditions. If benchmark questions leaked into the training data, a problem called contamination, the model may simply remember the answers. Every model also has to be tested with the same prompts and rules.

Benchmarks get used up: once top models score near the maximum, a test can no longer tell them apart, so new and harder ones appear. Labs add blind human comparisons and safety tests, because a high score doesn't prove a model is useful or safe.