ResearchBenchmarksRobustness to Jailbreak Attacks on Adaptive attackFollow20Harmful Reasoning RatioIPO17.8432.424761.58Sep 29, 2025Evaluation ResultsMethodMethodLinksHarmful Reasoning RatioHarmful Response RatioIPOBackbone=DS-8BBackbone=DS-8B2025.09200RealSafeBackbone=DS-8BBackbone=DS-8B2025.09410BaseBackbone=DS-8BBackbone=DS-8B2025.097449