Have you heard of "beam search"? The model keeps a number of optional lines of text without being greedy and picking only the most probable token every time. Yo

Have you heard of "beam search"? The model keeps a number of optional lines of text without being greedy and picking only the most probable token every time. You can also sample with T>0 multiple times to get an ensemble of answers that can be used as input for another AI operation. You just need to do multiple interactions with the model.