If you cannot access the pretraining corpus to check fo...

If the pretraining corpus is unavailable for inspection, you must rely on evaluation design to minimize risk. Strategies include using held-out private benchmarks that have not been released publicly, or generating new examples programmatically at the time of evaluation, such as in LiveCodeBench or FrontierMath. Furthermore, you can use behavioral probes that force the model to demonstrate reasoning steps rather than relying on answer recall. Finally, methods like Chatbot Arena sidestep contamination by focusing on open-ended human preferences on novel, unseen prompts.