You are getting derailed because of the name we've chosen to call these models but only the first few and last few layers of LLMs deal with tokens. The rest deal with abstract representations and models learnt during training. Language goes in and Language comes out but Language is not the in-between for either LLMs or Humans.