Safety Detection on Harmbench (test)
100MCA AccuracyReGA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| ReGAModel=qwen2025.06 | 100 | 91 | — | |
| MLP-Cls.Model=mistral2025.06 | 89 | 27 | — | |
| ReGAModel=llama2025.06 | 89 | 89 | — | |
| ReGAModel=vicuna2025.06 | 88 | 84 | — | |
| ReGAModel=mistral2025.06 | 86 | 63 | — | |
| ReGAModel=Average2025.06 | 81 | 74 | — | |
| ReGAModel=baichuan2025.06 | 80 | 80 | — | |
| MLP-Cls.Model=vicuna2025.06 | 68 | 51 | — | |
| MLP-Cls.Model=llama2025.06 | 62 | 22 | — | |
| Uni.-DTMCModel=vicuna2025.06 | 57 | 10 | — | |
| MLP-Cls.Model=qwen2025.06 | 56 | 17 | — | |
| MLP-Cls.Model=Average2025.06 | 56 | 28 | — | |
| Uni.-DTMCModel=qwen2025.06 | 55 | 0 | — | |
| ReGAModel=koala2025.06 | 43 | 39 | — | |
| MLP-Cls.Model=baichuan2025.06 | 39 | 39 | — | |
| MLP-Cls.Model=koala2025.06 | 24 | 12 | — | |
| Uni.-DTMCModel=Average2025.06 | 21 | 2 | — | |
| Uni.-DTMCModel=mistral2025.06 | 14 | 0 | — | |
| Uni.-DTMCModel=baichuan2025.06 | 1 | 1 | — | |
| Uni.-DTMCModel=llama2025.06 | 0 | 0 | — | |
| Uni.-DTMCModel=koala2025.06 | 0 | 0 | — | |
| LLM-JudgeModel=vicuna2025.06 | — | — | 89 | |
| LLM-JudgeModel=llama2025.06 | — | — | 12 | |
| LLM-JudgeModel=qwen2025.06 | — | — | 4 | |
| LLM-JudgeModel=mistral2025.06 | — | — | 94 | |
| LLM-JudgeModel=koala2025.06 | — | — | 99 | |
| LLM-JudgeModel=baichuan2025.06 | — | — | 0 | |
| LLM-JudgeModel=Average2025.06 | — | — | 50 | |
| PPL-FilterModel=vicuna2025.06 | — | — | 7 | |
| PPL-FilterModel=llama2025.06 | — | — | 1 | |
| PPL-FilterModel=qwen2025.06 | — | — | 0 | |
| PPL-FilterModel=mistral2025.06 | — | — | 1 | |
| PPL-FilterModel=koala2025.06 | — | — | 3 | |
| PPL-FilterModel=baichuan2025.06 | — | — | 12 | |
| PPL-FilterModel=Average2025.06 | — | — | 4 |