When is ROUGE-L more appropriate than ROUGE-2 for evaluating LLM outputs?

**ROUGE** is a family of overlap metrics for comparing a generated text against a reference, used heavily for summarization. ROUGE-2 counts matching bigrams (consecutive word pairs). ROUGE-L counts the longest common subsequence (LCS): the longest run of words that appear in both texts in the same order, but not necessarily next to each other. The correct answer is the first option. Summaries are a good rephrase of the source, not a copy. A summary might say 'revenue declined sharply' where the reference said 'sharp decline in revenue'. ROUGE-2 finds almost no shared bigram there, so it scores the summary low even though the content matches. ROUGE-L still picks up 'revenue' and 'declined' in order, so it rewards the in-order overlap that ROUGE-2 misses. The other options are wrong. ROUGE-L is not language-agnostic. It compares word tokens, so cross-language pairs share almost nothing. It also does not always beat ROUGE-2, and it does not subsume bigram matching. They reward different things. ROUGE-L favors order; ROUGE-2 favors exact local phrasing. Both are surface metrics that ignore meaning.

**ROUGE-L** is the right choice over ROUGE-2 when you care about in-order content overlap but exact phrasing varies. That describes summarization well. ROUGE-2 scores contiguous bigram matches. A good summary rephrases the source, so it rarely reuses exact word pairs. The reference says 'sharp decline in revenue'. The summary says 'revenue declined sharply'. ROUGE-2 finds little shared bigram and scores it low. ROUGE-L instead uses the longest common subsequence. It rewards words that appear in both texts in the same order, even with gaps. So 'revenue' then 'declined' still counts. That tolerance to gaps and reordering is exactly what summarization needs. The core formula is the LCS F-measure: $$F_\text{LCS} = \frac{(1 + \beta^2)\, R_\text{LCS}\, P_\text{LCS}}{R_\text{LCS} + \beta^2 P_\text{LCS}}$$ Recall divides LCS length by reference length. Precision divides it by candidate length. The original metric is recall-oriented, weighting recall heavily. The distractors are wrong. ROUGE-L is not language-agnostic; it matches tokens. It does not subsume ROUGE-2; the two reward different phenomena. Neither captures meaning. Both are surface overlap.