Why are BLEU and ROUGE poor metrics for judging LLM answers?
They measure word overlap with a reference text. A correct answer worded differently scores low, while a fluent wrong answer sharing many words can score high. They suit narrow tasks like translation drafts, but not open-ended answer quality.
Explain it simply
Suppose the reference answer is "The exam starts at 10 am." A model says "Papers begin at ten in the morning." The meaning is identical but the overlapping words are almost none, so the score is poor. Another reply, "The exam does not start at 10 am", shares almost every word and scores high, yet it is wrong.
BLEU and ROUGE count matching n-grams against reference text. That makes them cheap and reproducible, and fine as rough signals for tasks with limited valid phrasings, such as comparing summaries of the same style.
For modern LLM output, many correct phrasings exist, so you need measures of meaning and correctness: embedding-based similarity, task-specific checks, rubric grading by a model, and human review.
Example
Reference: "The exam starts at 10 am."
Answer A: "Papers begin at ten in the morning." -> correct, low overlap score
Answer B: "The exam does not start at 10 am." -> wrong, high overlap scoreWhat to say in the interview
"BLEU and ROUGE reward lexical overlap, not meaning or truth, so paraphrases score low and subtle errors score high. I use them at most as a rough signal and rely on task-specific checks, rubrics and human review."