Is it able to detect benchmark hunting ?

[Greatz08]

Many models just train their data on benchmark questions and so if you choose those benchmark questions for - performance testing then 7B and 14B thinking models can also compete in many questions against big models (200B+) , so have you thought about any better performance evaluation for llm's ?

Some did create private data sets for testing but again we cant trust those individuals testing (they can be biased or they can be corrupted with money) but if we open source questions/tests then people can and are already training their data on those questions/tests to score high points in each benchmark or evaluation tests, so i was wondering like is their any better way to really evaluate llms ?