The marked tokens appear at the beginning of the model's turn in conversational exchanges, often capturing the first substantive word(s) or phrase of the AI assistant's response. The pattern primarily highlights response initiators, key content words in answers (especially nouns, verbs, and important phrases), transitions between ideas, and occasionally formatting elements or punctuation that structure the response. The markers seem to identify semantically significant tokens that establish the direction or core meaning of the assistant's reply.
