Tag
This paper identifies surface-level feature leakage in truthfulness benchmarks like TruthfulQA, where models can cheat by exploiting answer form differences, and introduces Audit-Prune to clean benchmarks, ensuring more reliable evaluations.