ResearchBenchmarksError Detection on TruthfulQA 200-question subsetFor AI assistantsMarkdownContent indexError Detection on TruthfulQA 200-question subsetFollow0.512AUROCSE-NLI0.500560.503530.50650.50947Mar 25, 2026Evaluation resultsMethodMethodLinksAUROC95% CICost (NLI Steps)SE-NLIBackbone=DeBERTa-base,...Backbone=DeBERTa-base, Sample size (N)=102026.030.5120.42111SE-NLIBackbone=DeBERTa-large...Backbone=DeBERTa-large, Sample size (N)=102026.030.5110.41911SE-NLIBackbone=DeBERTa-xsmal...Backbone=DeBERTa-xsmall, Sample size (N)=102026.030.5010.40411