ResearchBenchmarksSelective Answering on HotpotQA RAG v3Follow0.1642Risk@30SURE-RAG0.1580360.1996430.241250.282857May 5, 2026Evaluation ResultsMethodMethodLinksRisk@30Risk@50Risk@70SURE-RAGCalibration=StandardCalibration=Standard2026.050.16420.360.5457SURE-RAGCalibration=CalibratedCalibration=Calibrated2026.050.1670.3510.5493DeBERTaPooling=mean-poolPooling=mean-pool2026.050.25880.37830.535DeBERTaPooling=max-poolPooling=max-pool2026.050.31830.40540.5344