A language-learning app uses an LLM to rewrite text to be more readable. The training dataset includes original text paired with more readable versions, and the company wants outputs to match those examples. Which metric should they use to evaluate how closely the model’s outputs match the reference rewrites?

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) score is the most appropriate metric because it measures the overlap of n-grams (sequences of words) between a model's generated text and a set of reference texts. This directly assesses how closely the model's rewrites match the desired readable versions provided in the training data, which is crucial for tasks like text summarization or, in this case, text simplification where specific word choices and phrases are important. The value of the loss function is used during training to guide optimization but doesn't directly evaluate the quality of the generated text in human-understandable terms. Semantic robustness measures how well a model maintains its performance when inputs are slightly perturbed, which is not the primary goal here. Text-generation latency measures the time it takes for the model to produce output, focusing on efficiency rather than the quality or accuracy of the rewrite.