List your company

Jailbreak Attack Defense on Adversarial Jailbreak Prompts

63ATTEN ScoreSpecreason
-2.5214.4931.548.51Jun 18, 2026

Evaluation results

MethodLinks
SpecreasonModel=Qwen3-32B (Draft: 1.7B)2026.066366556976598467
Specreason+RPOModel=DeepSeek-70B (Draft: 8B)2026.06613912838545438
SpecreasonModel=DeepSeek-70B (Draft: 8B)2026.065665176645777457
No DefenseModel=Qwen3-32B (Draft: 1.7B)2026.065564627062578665
Spec-DecodingModel=Qwen3-32B (Draft: 1.7B)2026.065563616963578865
SecDecodingModel=DeepSeek-70B (Draft: 8B)2026.065363155746746653
Spec-DecodingModel=DeepSeek-70B (Draft: 8B)2026.065056186341774850
No DefenseModel=DeepSeek-70B (Draft: 8B)2026.064857196242804651
SafeDecodingModel=Qwen3-32B (Draft: 1.7B)2026.0646544350138032
RPOModel=Qwen3-32B (Draft: 1.7B)2026.063517123231573831
RPOModel=DeepSeek-70B (Draft: 8B)2026.063546114337722639
Specreason+RPOModel=Qwen3-32B (Draft: 1.7B)2026.06322872341266231
Specreason+SRModel=DeepSeek-70B (Draft: 8B)2026.06325332421713835
SRModel=DeepSeek-70B (Draft: 8B)2026.0617340255661623
Specreason+SRModel=Qwen3-32B (Draft: 1.7B)2026.06144253729476234
SRModel=Qwen3-32B (Draft: 1.7B)2026.0693163226526831
SafeSpecModel=DeepSeek-70B (Draft: 8B)2026.06914029767
SecDecodingModel=Qwen3-32B (Draft: 1.7B)2026.06665715466222
SafeSpecModel=Qwen3-32B (Draft: 1.7B)2026.065104113987
SafeDecodingModel=DeepSeek-70B (Draft: 8B)2026.060280020309