1
Leak@$k$: Unlearning Does Not Make LLMs Forget Under Probabilistic Decoding
大语言模型的“遗忘”并不彻底，概率解码下仍会泄露被遗忘数据，Leak@k指标揭示这一漏洞。
arXiv:2511.04934v3 Announce Type: replace Abstract: Unlearning in large language models (LLMs) is critical for regulatory compliance and for building …