A developer is deploying a generative AI model to write marketing copy. They want the model to consider only the smallest set of the most probable next words whose cumulative probability reaches a specified threshold, filtering out unlikely tokens while still allowing some variety. Which configuration parameter should they adjust?