A quick question about the principle of using order/sequence to test contamination

[HelloWorldLTY]

Hi, thanks for your great work. I have a simple question about the relationship between sentence order or token order and data contamination. For example, if I ask GPT-4 to correct an order of one sentence, I have the following dialog and reasons:

According to the explainability of GPT-4, the order of tokens seem to be irrelevant to the training dataset but related to preference (like how to align models with human value). Therefore, how to precldue the reason of human alignment or common word order in testing the existance of data contamination? Thanks a lot.

Maybe I misunderstand your work, but I am looking for further explainations.