A reference-based metric evaluates a system output by comparing it with one or more answers treated as ground truth or acceptable exemplars. References can come from public benchmarks, labeled datasets, gold-standard outputs, or custom evaluation sets. section 3.2
Medical advice chatbot
For factual questions, compare a chatbot’s answers with clinician-verified references using exact match, semantic similarity, or a learned evaluation metric. section 3.2
Reference-based evaluation depends on having suitable answers to compare against. Where a correct answer does not exist or is too expensive to produce, reference-free metrics offer a contrast: they assess outputs without a fixed ground truth. section 3.2
