1
Bits and Memories: Measuring Verbatim Extraction Across LLM Quantization
LLM量化压缩后仍会泄露训练数据，新研究揭示逐字提取风险与量化位数的关系
arXiv:2607.25451v1 Announce Type: new Abstract: Language models are almost always quantized before they are deployed, and a growing line of work asks …