Utility Preservation on MMLU 5-shot (test)
67.2Utility ScoreBenign (No Attack)
Evaluation Results
| Method | Links | |
|---|---|---|
| Benign (No Attack)Method=Benign (No Attack), Attack Target=All, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.2 | |
| SFT-basedMethod=SFT-based, Attack Target=BadDLMConcept, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.1 | |
| SFT-basedMethod=SFT-based, Attack Target=BadDLMAttribute, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.1 | |
| SFT-basedMethod=SFT-based, Attack Target=BadDLMPayload, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.1 | |
| VPIMethod=VPI, Attack Target=BadDLMAttribute, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.1 | |
| VPIMethod=VPI, Attack Target=BadDLMPayload, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.1 | |
| BadDLMMethod=BadDLM (Ours), Attack Target=BadDLMConcept, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.1 | |
| BadDLMMethod=BadDLM (Ours), Attack Target=BadDLMAttribute, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.1 | |
| BadDLMMethod=BadDLM (Ours), Attack Target=BadDLMPayload, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67.1 | |
| SFT-basedMethod=SFT-based, Attack Target=BadDLMAlign, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67 | |
| VPIMethod=VPI, Attack Target=BadDLMConcept, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67 | |
| BadDLMMethod=BadDLM (Ours), Attack Target=BadDLMAlign, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 67 | |
| RL-basedMethod=RL-based, Attack Target=BadDLMConcept, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 66.3 | |
| RL-basedMethod=RL-based, Attack Target=BadDLMPayload, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 66.3 | |
| RL-basedMethod=RL-based, Attack Target=BadDLMAttribute, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 65.9 | |
| RL-basedMethod=RL-based, Attack Target=BadDLMAlign, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.05 | 65.9 |