Greedy decoding is the simplest decoding method for text generation: at each step the language model selects the single token with the highest probability. It is deterministic, producing the same output for the same input, and is effectively equivalent to setting the temperature to zero.
Because the best choice at each step is not necessarily the best for the sentence as a whole, it tends to produce repetitive or monotonous text. Beam search and sampling techniques such as top-k and top-p are used to compensate.