What it measures
A hazardous-knowledge proxy supports research on evaluating and reducing specific kinds of model knowledge.
How to read the result
Interpret it as a constrained knowledge probe and inspect retained general capability alongside any reduction.
Limits and comparison traps
Lower accuracy on a proxy does not establish that a model cannot cause harm. Keep the tested knowledge domain and evaluation scope explicit.
Published evaluation
Accuracy on the named hazardous-knowledge subset. Lower is the unlearning objective in this study, not greater model capability. Base and unlearned variants remain distinct model configurations.
Where this fits in an evaluation suite
Format compliance, preference, truthfulness, and safety require separate measurements. Keep desirable writing style from masking factual errors.
Start with a capability hypothesis, then choose a second test with a different failure mode. Keep the checkpoint, reasoning effort, agent tools, sampling policy, and scoring procedure fixed when measuring a training intervention. Record failures and missing outputs in the denominator.
A reproducible comparison
Accuracy on the named hazardous-knowledge subset. Lower is the unlearning objective in this study, not greater model capability. Base and unlearned variants remain distinct model configurations.
Publish the exact run configuration, item coverage, score units, and available uncertainty. If a source supplies only an aggregate or a single run, preserve that limit. A new model name or a rounded score alone does not establish a capability improvement.
PRIMARY EVIDENCE & FURTHER READING
- The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning — published result table ↗
All 57 authors
Li, Nathaniel, Pan, Alexander, Gopal, Anjali, Yue, Summer, Berrios, Daniel, Gatti, Alice, Li, Justin D., Dombrowski, Ann-Kathrin, Goel, Shashwat, Phan, Long, Mukobi, Gabriel, Helm-Burger, Nathan, Lababidi, Rassin, Justen, Lennart, Liu, Andrew B., Chen, Michael, Barrass, Isabelle, Zhang, Oliver, Zhu, Xiaoyuan, Tamirisa, Rishub, Bharathi, Bhrugu, Khoja, Adam, Zhao, Zhenqi, Herbert-Voss, Ariel, Breuer, Cort B., Marks, Samuel, Patel, Oam, Zou, Andy, Mazeika, Mantas, Wang, Zifan, Oswal, Palash, Lin, Weiran, Hunt, Adam A., Tienken-Harder, Justin, Shih, Kevin Y., Talley, Kemper, Guan, John, Kaplan, Russell, Steneker, Ian, Campbell, David, Jokubaitis, Brad, Levinson, Alex, Wang, Jean, Qian, William, Karmakar, Kallol Krishna, Basart, Steven, Fitz, Stephen, Levine, Mindy, Kumaraguru, Ponnurangam, Tupakula, Uday, Varadharajan, Vijay, Wang, Ruoyu, Shoshitaishvili, Yan, Ba, Jimmy, Esvelt, Kevin M., Wang, Alexandr, Hendrycks, Dan
- Primary benchmark documentation or model evaluation report ↗
Reviewed 2026-10-26. Explanation and interpretation by RLTasks. Benchmark ownership remains with its original authors.