The probability distribution over all possible next words a language model can generate at each step.