An autoregressive language model has generated the phrase 'After the long hike, we were all very'. To determine the next word, the model evaluates several options from its vocabulary. Which of the following calculations best represents the core principle the model uses to decide which word (e.g., 'tired', 'hungry', 'happy') is most likely to come next?

Deconstructing Next-Token Prediction

An autoregressive model is in the process of generating a sentence. So far, it has produced the sequence of words: 'The cat sat on the'. The model is now trying to determine the most probable next word. Which of the following mathematical expressions correctly represents the probability the model is calculating for the specific word 'mat' to be the next word in the sequence?