ROUGE-L
Training language models to follow instructions with human feedback — inherited
A standard n-gram/longest-common-subsequence overlap metric used to score generated summaries against reference summaries on CNN/DM and Reddit TL;DR.
How do you grade a summary automatically? ROUGE-L takes a rough but fast route: find the longest sequence of words the candidate shares, in order, with a reference summary, and score the overlap. No meaning is consulted, only shared word order.
Scoring the two summarization benchmarks is its role here, but the metric predates this paper by years and carries no connections of its own, a tool borrowed wholesale rather than built for the occasion.
What the metric measures
ROUGE-L is a recall-oriented automatic summarization metric based on the Longest Common Subsequence (LCS) between a generated summary and a reference summary, rewarding in-order shared word sequences without requiring them to be contiguous. Given candidate $X$ (length $n$) and reference $Y$ (length $m$), define recall and precision from the LCS length, $R_{lcs} = LCS(X,Y)/m$ and $P_{lcs} = LCS(X,Y)/n$, and combine them into an F-measure $$F_{lcs} = \frac{(1+\beta^2)\,R_{lcs}\,P_{lcs}}{R_{lcs} + \beta^2 P_{lcs}},$$ typically with $\beta$ set to weight recall heavily. ROUGE-L predates this paper by nearly two decades, part of the ROUGE family introduced for automatic summarization evaluation; InstructGPT inherits it unmodified as a standard summarization metric.
Its role here
ROUGE-L scores generated summaries against reference summaries on the CNN/Daily Mail and Reddit TL;DR summarization tasks in InstructGPT's automatic public NLP evaluation suite (instructgpt, §"CNN/DM Summarization", p. 49). Within the Automatic overlap scores versus human and preference judgments theme, it plays the same structural role for summarization that BLEU plays for translation and F1 score (span-overlap QA metric) plays for extractive question answering: a cheap, reference-overlap score, contrasted against the judgment-based Win rate (preference against baseline) and Likert overall-quality rating that the paper leans on for its central human-preference claims. Because overlap metrics like ROUGE-L cannot detect whether a summary is genuinely preferred, only whether it echoes a reference's wording, this theme's broader point is that InstructGPT needed the judgment-based metrics to make its main argument; ROUGE-L is retained mainly to keep results comparable with prior summarization-focused RLHF work such as Stiennon et al. (2020).