Exact Match Evaluation
Exact match evaluation scores a model's output as correct only if it matches a reference answer character-for-character (or after light normalization like case-folding), making it the strictest and simplest automated metric — well suited to tasks with one unambiguous correct answer, and badly suited to almost everything else.
Exact match works well for narrow, closed-answer tasks — a math result, a classification label, a yes/no answer, a specific extracted value — where there genuinely is one correct string. Applied to open-ended generation, it's nearly useless: a fully correct answer phrased even slightly differently from the reference scores as a total failure.
Because it's binary and cheap to compute, exact match is often the fastest first filter in a larger evaluation pipeline — quickly separating "clearly correct" and "clearly wrong" cases before routing the ambiguous remainder to a more expensive semantic or human-graded check.
Normalization choices matter more than they seem: whether to strip whitespace, ignore punctuation, fold case, or accept numeric equivalents (e.g. "5" vs "5.0") can silently move real pass rates by a large margin depending on how strictly they're applied.
