List your company

Jailbreak Safety on JailbreakBench

0.3Reasoning Harmful RatioGRPO
-2.58416.88336.3555.817Sep 29, 2025

Evaluation results

MethodLinks
GRPOBackbone=DeepSeek-R1-Distill-Llama-8B2025.090.3——0
GRPOBackbone=Qwen3-8B2025.091.7——0
GRPOBackbone=DeepSeek-R1-Distill-Qwen-7B2025.093——0.3
IPOBackbone=Qwen3-8B2025.095.2——0
IPOBackbone=DeepSeek-R1-Distill-Llama-8B2025.095.7——0.3
SafeKeyBackbone=DeepSeek-R1-Distill-Qwen-7B2025.096——0.3
STARBackbone=DeepSeek-R1-Distill-Llama-8B2025.098——0.3
SafeKeyBackbone=DeepSeek-R1-Distill-Llama-8B2025.098.7——0
IPOBackbone=DeepSeek-R1-Distill-Qwen-7B2025.0911——0
STARBackbone=DeepSeek-R1-Distill-Qwen-7B2025.0911.4——2
RealSafeBackbone=DeepSeek-R1-Distill-Qwen-7B2025.0918——0.7
RealSafeBackbone=DeepSeek-R1-Distill-Llama-8B2025.0920.7——0
BaseBackbone=Qwen3-8B2025.0924.3——4
SafeChainBackbone=DeepSeek-R1-Distill-Llama-8B2025.0956.1——25.3
SafeChainBackbone=DeepSeek-R1-Distill-Qwen-7B2025.0956.1——25.3
BaseBackbone=DeepSeek-R1-Distill-Llama-8B2025.0969——45
BaseBackbone=DeepSeek-R1-Distill-Qwen-7B2025.0972.4——56
BackTrack2025.09—70—
Base2025.09—6945—
GRPO2025.09—0.30—
IPO2025.09—5.70.3—
STAR2025.09—80.3—
TARS2025.09—60—