ResearchBenchmarksSafety Evaluation on AIR-Bench 64Follow3.12Harm ScoreSFT3.04643.54324.044.5368Oct 17, 2025Evaluation ResultsMethodMethodLinksHarm ScoreUtility ScoreAccuracySFTStrategy=Threshold (τ...Strategy=Threshold (τ = 0.5)2025.103.127.7572.47RLVRStrategy=Threshold (τ...Strategy=Threshold (τ = 0.5)2025.103.337.479.6SFTStrategy=Top-k (k = 100)Strategy=Top-k (k = 100)2025.104.497.7865.73RLVRStrategy=Top-k (k = 100)Strategy=Top-k (k = 100)2025.104.687.4579.27SFTStrategy=BaselineStrategy=Baseline2025.104.817.4466.13RLVRStrategy=BaselineStrategy=Baseline2025.104.967.7380.93