A contextualized representation encodes a prompt after its tokens have been processed through transformer layers. The final prompt representation supports prediction, but relevant information can come from different places: token relationships within the prompt, or associations learned in the model’s parameters. For a factual question, attention helps represent the question and title; the title–author association is distributed across model parameters. The final representation is converted into vocabulary logits, then softmax assigns probabilities to possible continuations. section 12.2.2
A literary question
For “Who wrote The Count of Monte Cristo?”, the final prompt representation is used to score possible continuations such as “Dumas.” That association comes from patterns learned during training, not from the prompt directly attending to “Dumas” as a context token. section 12.2.2
An illustrative final representation used in the vocabulary-scoring example. section 12.2.2






