List your company
ResearchFormerly WizWand

Knowledge Recovery on WMDP-Bio 100-sample subset

0.93ASRREBEL
0.18120.37560.570.7644Feb 5, 2026
Updated 8mo ago

Evaluation results

MethodLinks
REBELUnlearning technique=IDKNLL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.93
REBELUnlearning technique=UNDIAL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.85
REBELUnlearning technique=IDKDPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.69
Leak@1000Unlearning technique=IDKNLL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.66
Leak@500Unlearning technique=IDKNLL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.65
REBELUnlearning technique=NPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.6
Leak@100Unlearning technique=IDKNLL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.58
Leak@500Unlearning technique=IDKDPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.58
Leak@1000Unlearning technique=IDKDPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.58
Leak@100Unlearning technique=IDKDPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.51
Leak@1000Unlearning technique=UNDIAL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.51
Leak@10Unlearning technique=IDKNLL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.5
Leak@500Unlearning technique=UNDIAL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.48
Leak@1000Unlearning technique=NPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.44
Leak@500Unlearning technique=NPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.43
Leak@100Unlearning technique=UNDIAL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.42
Leak@100Unlearning technique=NPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.39
Leak@10Unlearning technique=UNDIAL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.33
REBELUnlearning technique=SimNPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.33
BaselineUnlearning technique=IDKNLL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.32
Leak@10Unlearning technique=NPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.32
Leak@10Unlearning technique=IDKDPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.31
REBELUnlearning technique=GradDiff, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.31
BaselineUnlearning technique=IDKDPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.28
BaselineUnlearning technique=UNDIAL, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.26
Leak@1000Unlearning technique=SimNPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.26
BaselineUnlearning technique=NPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.25
Leak@500Unlearning technique=SimNPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.25
Leak@500Unlearning technique=GradDiff, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.24
Leak@1000Unlearning technique=GradDiff, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.24
Leak@10Unlearning technique=GradDiff, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.23
Leak@100Unlearning technique=GradDiff, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.23
BaselineUnlearning technique=SimNPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.22
Leak@10Unlearning technique=SimNPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.22
Leak@100Unlearning technique=SimNPO, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.22
BaselineUnlearning technique=GradDiff, Target model=Llama-3-8B-Instruct, Hacker model=Qwen2.5-7B-Instruct, Judge model=Qwen2.5-7B-Instruct2026.020.21