How AI describes Data Contamination Quiz (DCQ)

Data Contamination Quiz (DCQ) is a tool that detects and estimates data contamination in large language models by presenting multiple-choice questions where the model must identify an original dataset instance among its word-level perturbations. The method, proposed by Shahriar Golchin and Mihai Surdeanu, is the first to estimate contamination in fully black-box LLMs by analyzing the model's ability to recognize previously seen data.