Benchmark
벤치마크
A test that measures AI models on the same questions so you can compare them.
Overview
When a new model comes out, it is given a set of fixed questions in areas like math, coding and reading, and gets a score. That question set and its way of scoring is a benchmark. A table that ranks the scores is called a leaderboard.
Benchmarks are handy, but they have limits. If the test questions leak into the training data, scores get inflated. And a high score does not guarantee the model will meet expectations on real work.
An everyday example
It is like a car's official fuel economy rating. It is useful for comparing, but it may differ from the mileage you get when you actually drive.
Why it matters
News that "the new model took first place" is almost always about a benchmark. Look at which test it won, and at how much the model is actually used.
Also calledperformance test