On the Recoverability of Private Information Unlearning in Large Language Models
A synthetic dataset containing fake private information is constructed and a white-box auditing framework is proposed to systematically assess whether claimed-forgotten information is genuinely removed and finds that a simpleverse greedy decoding -- selecting the least likely token at each step -- can recover supposedl...
Shi-Cheng Hu, Runhe Tian, Ziqiao Wang et al.
· 0 citations