机器遗忘可用于从大语言模型(LLMs)中移除有害能力和记忆文本,但目前尚缺乏标准化的方法来严格评估其效果。在本文中,我们首先梳理了现有遗忘评估的技术与局限性。其次,我们对Eldan和Russinovich(2023)提出的“Who's Harry Potter”(WHP)模型应用了一套全面的测试,以检验其遗忘的鲁棒性和竞争力。虽然在使用Eldan和Russinovich的“Familiarity”指标评估时,WHP的遗忘表现出良好的泛化性,但我们发现:i)可以可靠地提取出高于基线水平的知识量;ii)WHP在哈利·波特问答任务上的表现与原始模型相当;iii)它对潜在知识的表征与原始模型具有可比性;iv)在相关领域存在附带遗忘。总体而言,我们的结果凸显了采用全面遗忘评估并避免临时性指标的重要性。
英文摘要:
Machine unlearning can be useful for removing harmful capabilities and memorized text from large language models (LLMs), but there are not yet standardized methods for rigorously evaluating it. In this paper, we first survey techniques and limitations of existing unlearning evaluations. Second, we apply a comprehensive set of tests for the robustness and competitiveness of unlearning in the "Who's Harry Potter" (WHP) model from Eldan and Russinovich (2023). While WHP's unlearning generalizes well when evaluated with the "Familiarity" metric from Eldan and Russinovich, we find i) higher-than-baseline amounts of knowledge can reliably be extracted, ii) WHP performs on par with the original model on Harry Potter Q&A tasks, iii) it represents latent knowledge comparably to the original model, and iv) there is collateral unlearning in related domains. Overall, our results highlight the importance of comprehensive unlearning evaluation that avoids ad-hoc metrics.
Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting
Faculty of Information Technology, Monash University(墨尔本大学信息科技学院);Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center, Qilu University of Technology (Shandong Academy of Sciences)(教育部计算能力网络与信息安全重点实验室,山东计算机科学中心,齐鲁大学(山东科学院));Shandong Provincial Key Laboratory of Computing Power Internet and Service Computing, Shandong Fundamental Research Center for Computer Science(山东省计算能力互联网与服务计算重点实验室,山东省计算机科学基础研究中心);School of Computer Science and Technology, Anhui University(安徽大学计算机科学与技术学院)
Knowledge Beyond Language: Bridging the Gap in Multilingual Machine Unlearning Evaluation
GSCST, Seoul National University(首尔国立大学GSCST);AIIS, Seoul National University(首尔国立大学AIIS);Department of Artificial Intelligence, Chung-Ang University(Chung-Ang大学人工智能系);Korean Surgical Researcher Foundation, Republic of Korea(韩国外科研究员基金会)
CRISP: Persistent Concept Unlearning via Sparse Autoencoders
Technion – Israel Institute of Technology(技术ion-以色列理工学院);Boston University(波士顿大学);University of Zagreb, FER(Zagreb大学,FER);Kempner Institute, Harvard University(哈佛大学 Kempner研究所)
Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMs
The Hong Kong Polytechnic University(香港理工大学);Carnegie Mellon University(卡内基梅隆大学);University of California, Santa Cruz(加州大学圣克ruz分校);Huawei Technologies(华为技术有限公司);Research Centre for Privacy and Security Technologies in Future Smart Systems, PolyU(未来智能系统中的隐私与安全技术研究中心,PolyU)
School of Advanced Interdisciplinary Sciences, UCAS(UCAS交叉学科研究院);Academy of Mathematics and Systems Science, CAS(中国科学院数学与系统科学研究院);Institute of Computing Technology, CAS(中国科学院计算技术研究所);University of Chinese Academy of Sciences(中国科学院大学)
Rethinking Machine Unlearning for Large Language Models
Michigan State University(密歇根州立大学);IBM Research(IBM研究院);University of North Carolina at Chapel Hill(北卡罗来纳大学教堂山分校);University of California, Santa Cruz(加州大学圣克鲁兹分校);ByteDance Research(字节跳动研究院);Stanford University(斯坦福大学)