ResearchBenchmarksHATE backdoor EvaluationFollow0HATE ScoreGD-102.336-75.768-49.2-22.632Apr 12, 2026Evaluation ResultsMethodMethodLinksHATE ScoreAccuracyMMLU Delta AccuracyGD2026.040-1.3—CB2026.04-31.6-3.5—LIRA*2026.04-98.4-1.9—