Tag
This paper formalizes benchmark evaluation as an empirical measure over question difficulty, identifying failure modes like measure mismatch and coverage gaps, and proposes diagnostic metrics plus BenchFill, a difficulty-targeted item rewriting algorithm, with analysis across 25 popular benchmarks in 7 domains.