The logarithm of the probability a model assigns to a sequence of tokens; used to compare how likely a model thinks different sentences are.