Beam search generates multiple potential completions and scores multiple tokens by likelihood, the picks the most likely after some threshold or length, which is close to a "go back and try again".