Capability Localization: Capabilities Can be Localized rather than Individual Knowledge

Published as a conference paper at ICLR 2025C APABILITYL OCALIZATION: C APABILITIES C AN BE L OCALIZED RATHER THAN I NDIVIDUALK NOWLEDGE Xiusheng Huang 1,2,3, Jiaxiang Liu 1,2, Y equan Wang 3, Jun Zhao 1,2 and Kang Liu 1,2 1 The Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences 2 School of Artificial Intelligence, University of Chinese Academy of Sciences 3 Beijing Academy of Artificial Intelligence, Beijing, China huangxiusheng2020@ia.ac.cn, liujiaxiang21@mails.ucas.ac.cn, tshwangyequan@gmail.com, { jzhao,kliu }@nlpr.ia.ac.cn We found through fidelity and reliability evaluation experiments that individual knowledge cannot be localized. Afterwards, we constructed a dataset for decou-pling experiments and discovered the potential for localizing data commonalities. More and more research is focusing on the security (Bonaldi et al., 2024; Sun et al., 2024), ethics (Y an et al., 2024; Specifically, KN (Dai et al., 2021) believes that individual knowledge ROME (Meng et al., 2022a) believes that individual knowledge is stored on the Previous knowledge localization methods have proposed corresponding validation methods, reliability experiments will evaluate the reliability of these methods. In addition, the entire parameters chain occupies 2.6% of the overall model To further reveal the form of knowledge storage, we designed 1000 comparative samples and conducted decoupling experiments.