← AI Terminology
Benchmark
A benchmark is a standardised test or dataset used to measure and compare the performance of AI models on a defined task or set of tasks.
Benchmarks serve as the shared yardstick that makes progress in AI visible and comparable across research groups.
Benchmarks serve as the shared yardstick that makes progress in AI visible and comparable across research groups.
Why It Matters in AI
Without benchmarks, claims about model capability are untestable. Benchmarks like ImageNet, MMLU, and HumanEval have driven breakthroughs by giving the field a common goal to optimise against. They also carry risk: models increasingly train on benchmark data (contamination), and optimising for a benchmark can diverge from real-world usefulness — known as Goodhart's Law. Still, well-designed benchmarks remain the primary signal of genuine progress.
Key Points
| Aspect | Description |
|---|---|
| Types | Academic (ImageNet, MMLU), human-comparative (HumanEval, GPQA), chatbot arena (LMSYS), safety |
| Purpose | Measure capability on a specific task; enable apples-to-apples model comparisons |
| Saturation | Once top models score >95%, the benchmark no longer differentiates — new harder ones needed |
| Limitations | Benchmarks measure narrow proxies — high scores don't guarantee real-world usefulness |
| Key examples | MMLU, HumanEval, MATH, BIG-Bench, HELM, GLUE/SuperGLUE, ARC-AGI, Humanity's Last Exam |
| Contamination | When training data includes benchmark test examples — inflates scores without real capability |
Simple Analogy
A standardised exam lets you compare students from different schools — the exam is the benchmark. But if teachers start teaching to the test (training on the benchmark), the exam stops measuring underlying ability. The same dynamic plays out in AI: as models overfit to popular benchmarks, the field needs ever-harder exams.
Common Usage Examples
- MMLU: 57-subject multiple-choice — GPT-4 scores ~87%, human average ~89%
- HumanEval: 164 coding problems — GPT-4 scores ~67%, Claude 3.5 Sonnet ~92%
- LMSYS Chatbot Arena: crowd-sourced human preference ranking — treated as the gold-standard chatbot benchmark
- BIG-Bench: 204 diverse tasks designed to probe capabilities beyond pattern memorisation
- Humanity's Last Exam (2025): expert-level questions across 100+ disciplines — designed to be benchmark-contamination-resistant
Summary
In short: Benchmarks are the standardised tests that make AI progress measurable — invaluable for tracking capability, but worth scrutinising for contamination and real-world relevance.