The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning
UC Berkeley, Scale AI · UC Berkeley · Massachusetts Institute of Technology · Scale AI · Center for AI Safety · University of Illinois, Urbana Champaign · Principles of Intelligent Behavior in Biological and Social Systems (PIBBSS) · International Institute of Information Technology Hyderabad, Dhirubhai Ambani Institute Of Information and Communication Technology · Stanford University · SecureBio · University of Oxford · Georgia Institute of Technology · University of Southern California · UIUC · University of California, Los Angeles · Harvard University · CMU, Carnegie Mellon University · Tiny Fish Inc. · Carnegie Mellon University · New College of Florida · Clemson University · University of California, Berkeley · University of Waterloo · Drake University · University of Chicago · Keio University · International Institute of Information Technology Hyderabad · University of Newcastle · Arizona State University · University of Toronto / xAI
PDF 由论文原始站点提供，PaperCompass 不保存论文文件。
摘要
The White House Executive Order on Artificial Intelligence highlights the risks of large language models (LLMs) empowering malicious actors in developing biological, cyber, and chemical weapons. To measure these risks, government institutions and major AI labs are developing evaluations for hazardous capabilities in LLMs. However, current evaluations are private and restricted to a narrow range of malicious use scenarios, which limits further research into reducing malicious use. To fill these gaps, we release the Weapons of Mass Destruction Proxy (WMDP) benchmark, a dataset of 3,668 multiple-choice questions that serve as a proxy measurement of hazardous knowledge in biosecurity, cybersecurity, and chemical security. To guide progress on unlearning, we develop RMU, a state-of-the-art unlearning method based on controlling model representations. RMU reduces model performance on WMDP while maintaining general capabilities in areas such as biology and computer science, suggesting that unlearning may be a concrete path towards reducing malicious use from LLMs. We release our benchmark and code publicly at https://wmdp.ai.