MORAIDICTIONNAIRE IA
Évaluation

Benchmark

The standardized test that measures and compares AI models' performance on specific tasks.

A benchmark is to artificial intelligence what a standardized exam is to students: an identical test for everyone, designed to objectively compare the performance of different models on a well-defined task. Without a benchmark, claiming one model is "better" than another is mere opinion; with one, you get a reproducible number.

What a benchmark is made of

A benchmark combines three ingredients:

The metric depends on the task: accuracy for classification, F1 score, BLEU for translation, or pass rate. Accuracy is computed simply:

$$\text{Accuracy} = \frac{\text{correct predictions}}{\text{total number of examples}}$$

Famous examples

Benchmark Domain evaluated
ImageNet Image recognition
GLUE / SuperGLUE Language understanding
MMLU General knowledge and reasoning
HumanEval Code generation

Strengths and limits

Benchmarks structure progress: they give research a shared target. But they have a downside. Data contamination (a model having already seen the test during training) artificially inflates scores. And optimizing for a benchmark can drift away from real-world use — this is Goodhart's law: "when a measure becomes a target, it ceases to be a good measure."

A benchmark measures what it knows how to measure, never intelligence as a whole: it is a compass, not a map of the territory.

Explore the full AI dictionary →