Data contamination
When benchmark questions or answers end up in a model's training data, inflating its scores.
Also: benchmark contamination, test-set leakage
Benchmarks are published online, and the web is scraped for training, so test questions often leak into training sets. A model that has seen the answers can score well without the ability the test is meant to measure. Labs try to filter benchmark text out and report overlap checks, and researchers detect contamination by asking models to complete test questions word for word or by comparing results on fresh questions. Contamination is one reason the field keeps building new benchmarks and holding some test sets privately.