Jailbreak Attack Success Rate on JailbreakBench (ASR Variants)
91.67ASR (all)HiDRA
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| HiDRABackbone=Qwen2.5-3B-Instruct, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 91.67 | 70 | — | |
| HiDRABackbone=Gemma2-9B-IT, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 90 | 57 | — | |
| HiDRABackbone=Qwen2.5-7B-Instruct, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 90 | 69.33 | — | |
| Mean-AcTBackbone=Qwen2.5-3B-Instruct2026.06 | 87 | 68 | — | |
| Mean-AcTBackbone=Qwen2.5-7B-Instruct2026.06 | 86 | 71 | — | |
| Mean-AcTBackbone=Gemma2-9B-IT2026.06 | 84 | 53 | — | |
| HiDRABackbone=Llama3.2-3B-Instruct, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 84 | 73.33 | — | |
| Mean-AcTBackbone=Llama3.2-3B-Instruct2026.06 | 83 | 73 | — | |
| HiDRABackbone=Llama3.2-1B-Instruct, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 80.67 | 76 | — | |
| Mean-AcTBackbone=Llama3.2-1B-Instruct2026.06 | 78 | 72 | — | |
| Base2025.10 | — | — | 2 | |
| RLVR2025.10 | — | — | 97 | |
| SFT2025.10 | — | — | 77 |