List your company

Jailbreak Defense on StrongReject

1.5Attack Success RateR2D
-1.14416.70334.5552.397Mar 18, 2026

Evaluation results

MethodLinks
R2DBase Model=Qwen-8B2026.031.5————
PreSafeBase Model=Qwen-34B2026.032.9————
PreSafeBase Model=Qwen-8B2026.034.5————
PreSafeBase Model=DS-R1-14B2026.034.8————
PreSafeBase Model=DS-R1-7B2026.035.4————
R2DBase Model=DS-R1-14B2026.035.8————
PreSafeBase Model=Skywork-7B2026.037————
PreSafeBase Model=DS-R1-8B2026.037.1————
R2DBase Model=DS-R1-8B2026.038.1————
Perplexity FilterBase Model=Qwen-8B2026.038.4————
R2DBase Model=Qwen-34B2026.039.4————
R2DBase Model=Skywork-7B2026.0316————
R2DBase Model=DS-R1-7B2026.0316.3————
Erase And CheckBase Model=DS-R1-7B2026.0318.8————
Erase And CheckBase Model=DS-R1-14B2026.0319.9————
Erase And CheckBase Model=DS-R1-8B2026.0320.4————
Erase And CheckBase Model=Qwen-34B2026.0321.1————
Erase And CheckBase Model=Skywork-7B2026.0321.2————
Erase And CheckBase Model=Qwen-8B2026.0322.1————
Remove Non DictionaryBase Model=Qwen-8B2026.0322.4————
Perplexity FilterBase Model=Qwen-34B2026.0324.4————
Smooth LLMBase Model=Qwen-8B2026.0324.4————
NoneBase Model=Qwen-34B2026.0326————
NoneBase Model=Qwen-8B2026.0326.4————
Remove Non DictionaryBase Model=Qwen-34B2026.0326.9————
Smooth LLMBase Model=Qwen-34B2026.0331.4————
Synonym SubstitutionBase Model=Qwen-8B2026.0337.5————
SafeChainBase Model=DS-R1-8B2026.0338.8————
Perplexity FilterBase Model=DS-R1-14B2026.0339.1————
SafeChainBase Model=Qwen-8B2026.0340.7————
Smooth LLMBase Model=DS-R1-14B2026.0341.3————
Synonym SubstitutionBase Model=DS-R1-14B2026.0343.6————
Smooth LLMBase Model=DS-R1-8B2026.0343.9————
Synonym SubstitutionBase Model=DS-R1-8B2026.0344.9————
SafeChainBase Model=DS-R1-7B2026.0345.2————
Synonym SubstitutionBase Model=Qwen-34B2026.0346.6————
Remove Non DictionaryBase Model=DS-R1-14B2026.0346.9————
SafeChainBase Model=DS-R1-14B2026.0347.1————
SafeChainBase Model=Skywork-7B2026.0348.7————
Smooth LLMBase Model=DS-R1-7B2026.0348.9————
NoneBase Model=DS-R1-14B2026.0350.6————
Perplexity FilterBase Model=DS-R1-8B2026.0351.3————
SafeChainBase Model=Qwen-34B2026.0351.6————
Remove Non DictionaryBase Model=DS-R1-8B2026.0352.2————
Synonym SubstitutionBase Model=DS-R1-7B2026.0352.9————
NoneBase Model=DS-R1-8B2026.0352.9————
Smooth LLMBase Model=Skywork-7B2026.0355.8————
Perplexity FilterBase Model=Skywork-7B2026.0359.9————
Remove Non DictionaryBase Model=DS-R1-7B2026.0362.2————
Synonym SubstitutionBase Model=Skywork-7B2026.0363.2————
Perplexity FilterBase Model=DS-R1-7B2026.0364.7————
NoneBase Model=DS-R1-7B2026.0366————
NoneBase Model=Skywork-7B2026.0367————
Remove Non DictionaryBase Model=Skywork-7B2026.0367.6————
DefaultSystemLLM=zephyr-7b-beta2025.06—6.6731.9125.24—
DefaultSystemLLM=Llama-2-7b-chat2025.06—66.6798.9432.27—
DefaultSystemLLM=Llama-3.1-8b2025.06—3010070—
DefaultSystemLLM=Phi-3.5-mini2025.06—3.3321.2817.95—
DefaultSystemLLM=Mistral-7B-v0.22025.06—13.3393.6280.29—
KLAlignment Strategy=Deliberative Alignment (GRPO)2026.04————24.8
Latent Similarity- Best LayerAlignment Strategy=Deliberative Alignment (GRPO)2026.04————75
Latent Similarity- Layer 12Alignment Strategy=Deliberative Alignment (GRPO)2026.04————48
LoRA*LLM=zephyr-7b-beta2025.06—2092.5572.55—
LoRA*LLM=Llama-2-7b-chat2025.06—1010090—
LoRA*LLM=Llama-3.1-8b2025.06—0100100—
LoRA*LLM=Phi-3.5-mini2025.06—6.6748.9442.27—
LoRA*LLM=Mistral-7B-v0.22025.06—1010090—
PerplexityAlignment Strategy=Deliberative Alignment (GRPO)2026.04————-40.1
RandomAlignment Strategy=Deliberative Alignment (GRPO)2026.04————3.5
Self CertaintyAlignment Strategy=Deliberative Alignment (GRPO)2026.04————-146.4
SysformerLLM=zephyr-7b-beta2025.06—13.3375.5362.2—
SysformerLLM=Llama-2-7b-chat2025.06—3.3380.8577.52—
SysformerLLM=Llama-3.1-8b2025.06—3.3310096.67—
SysformerLLM=Phi-3.5-mini2025.06—6.6758.5151.84—
SysformerLLM=Mistral-7B-v0.22025.06—1010090—
SystemEmbedderLLM=zephyr-7b-beta2025.06—6.6734.0427.37—
SystemEmbedderLLM=Llama-2-7b-chat2025.06—6.674033.33—
SystemEmbedderLLM=Llama-3.1-8b2025.06—3010070—
SystemEmbedderLLM=Phi-3.5-mini2025.06—6.6726.619.93—
SystemEmbedderLLM=Mistral-7B-v0.22025.06—13.3393.6280.29—