BLEU Score
BLEU score is an automated metric that scores how closely a machine-generated text output matches one or more reference texts by comparing overlapping word sequences (n-grams), originally built for machine translation and still used as a cheap, if imprecise, proxy for output quality in AI testing.
BLEU rewards word-for-word overlap with a reference answer, which makes it fast and deterministic to compute but a poor fit for anything where a correct answer can be phrased many different ways — two equally correct paraphrases of the same fact can score very differently, and a fluent but wrong answer can score higher than a correct one phrased unusually.
In modern LLM testing, BLEU is mostly used as one cheap, quick signal alongside — never instead of — semantic or human-graded evaluation, useful for catching gross regressions (a model suddenly producing much shorter or wildly different output) rather than judging genuine correctness.
Its close relative, ROUGE, is more common for summarization tasks specifically since it weights recall of the reference's content more heavily — the two are often confused, but they optimize for slightly different things.
