That depends on the sampling strategy. Greedy sampling takes the max token at each step.