Beam search
At each generation step every one of the b kept sequences ("beams") is extended by each candidate token, and the b highest-probability results survive. It finds likelier outputs than greedy decoding, which suits translation and summarization, but tends to be bland on open-ended text.
Example
With beam width 2, if "The" → {"cat" 0.5, "dog" 0.4} and the best continuations are "dog barked" 0.4·0.9 = 0.36 vs "cat sat" 0.5·0.6 = 0.30, beam search keeps "The dog barked", which greedy (committed to "cat") would miss.