WMDP
A benchmark focused on unlearning dangerous knowledge (biological, chemical, cyberattack, and similar domains), measuring whether a model retains general capabilities after harmful capabilities are removed.
Timeline (last 90 days)
Experiments on the TOFU and WMDP benchmarks show that UUE and UUE-L significantly outperform existing methods in unlearning effectiveness.
WMDP (Weapons of Mass Destruction Proxy) focuses on unlearning dangerous knowledge (biological, chemical, cyberattacks, etc.) and measures whether a model retains its general capabilities after removing harmful abilities.
All Facts (2)
Experiments on the TOFU and WMDP benchmarks show that UUE and UUE-L significantly outperform existing methods in unlearning effectiveness.
50%UnverifiedWMDP (Weapons of Mass Destruction Proxy) focuses on unlearning dangerous knowledge (biological, chemical, cyberattacks, etc.) and measures whether a model retains its general capabilities after removing harmful abilities.
50%