List your company

Utility Preservation on MMLU 5-shot (test)

67.2Utility ScoreBenign (No Attack)
65.84866.19966.5566.901May 10, 2026

Evaluation results

MethodLinks
Benign (No Attack)Method=Benign (No Attack), Attack Target=All, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.2
SFT-basedMethod=SFT-based, Attack Target=BadDLMConcept, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.1
SFT-basedMethod=SFT-based, Attack Target=BadDLMAttribute, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.1
SFT-basedMethod=SFT-based, Attack Target=BadDLMPayload, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.1
VPIMethod=VPI, Attack Target=BadDLMAttribute, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.1
VPIMethod=VPI, Attack Target=BadDLMPayload, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.1
BadDLMMethod=BadDLM (Ours), Attack Target=BadDLMConcept, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.1
BadDLMMethod=BadDLM (Ours), Attack Target=BadDLMAttribute, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.1
BadDLMMethod=BadDLM (Ours), Attack Target=BadDLMPayload, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567.1
SFT-basedMethod=SFT-based, Attack Target=BadDLMAlign, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567
VPIMethod=VPI, Attack Target=BadDLMConcept, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567
BadDLMMethod=BadDLM (Ours), Attack Target=BadDLMAlign, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0567
RL-basedMethod=RL-based, Attack Target=BadDLMConcept, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0566.3
RL-basedMethod=RL-based, Attack Target=BadDLMPayload, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0566.3
RL-basedMethod=RL-based, Attack Target=BadDLMAttribute, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0565.9
RL-basedMethod=RL-based, Attack Target=BadDLMAlign, Base Model=Dream-Instruct-7B, Evaluation Protocol=5-shot2026.0565.9