An artificial intelligence benchmark defines a repeatable evaluation target, such as coding tasks, reasoning questions, multimodal problems, or serving performance. A score is meaningful only with the exact dataset, method, model settings, and grading rules.
Benchmarks can be contaminated, optimized against, selectively reported, or poorly matched to real use. Strong conclusions compare several evaluations and disclose uncertainty rather than treating one exceptional result as definitive.
