Does anybody know of research that shows that large language models can actually do 'out of pretraining domain' zero-shot prediction?

If we pretrain on diverse texts, would we not expect to always be 'partially in domain'. So does 'out of domain ZS prediction' exist?