ResearchBenchmarksNatural Language Inference on MultiNLI controlled shortcut injectionFollow32.3AccuracySG3.711.12518.5525.975Apr 14, 2026Evaluation ResultsMethodMethodLinksAccuracyWorst-Group Accuracy (WGA)SG2026.0432.31DFR2026.0427.31ERM2026.04271JTT2026.0426.83NFL2026.044.83