Sampling
Choosing each next token from the model's probabilities, with some randomness, rather than always taking the top one.
Also: decoding strategy
After softmax, the model offers a probability for every token in its vocabulary. Sampling draws one at random according to those probabilities, so the same prompt can give different answers. Settings shape the draw: temperature sharpens or flattens the probabilities, top-k keeps only the k most likely tokens, and top-p keeps the smallest set whose probabilities add up to a threshold. Some randomness makes writing more varied and natural; too much produces nonsense. A fixed random seed can make results repeatable.