For instance, an LLM might learn that a question like “Where is Paris located?” is structured as adverb/verb/proper noun/verb. If there are many examples of sentence construction in the model’s training data, the LLM may associate that syntactic template with questions about countries. So, if the model is given a new question with the same grammatical structure but nonsense words, like “Quickly sit Paris clouded?” it might answer “France” even though that answer makes no sense.

It sounds to me like model performance goes down when, in the training data, there's a lot of correlation between speech patterns and domains.