引用AIorNot · 2025-12
“e.g see this paper on Universal Embeddings Harnessing the Universal Geometry of Embeddings”
举例推荐这篇关于通用嵌入的论文,文本嵌入有通用潜在结构,可跨模型转换
无需配对数据即可在不同文本嵌入空间之间互译
论文提出首个无需配对数据、编码器或预设匹配集的无监督方法,把文本嵌入翻译到另一个向量空间,做法是经由一个通用潜在表示中转(对应柏拉图表征假说)。翻译后在不同架构、参数量和训练数据的模型之间都保持较高的余弦相似度,同时意味着只拿到嵌入向量的攻击者也能推断出文档的敏感信息,向量数据库的安全因此受到影响。
HN 上的人引用它说明不同模型的潜在空间存在可对齐的共同结构,以及拿到别家嵌入就可能还原其所含信息。
“e.g see this paper on Universal Embeddings Harnessing the Universal Geometry of Embeddings”
“If you can access a models emebeddings then it is possible to retrieve what it knows using a model you have trained”
“These, other papers, and the lottery ticket phenomenon; what it boils down to is that any neural network like system which encodes some common mapping of a phenomenon in the context of the world”
“some work that shows you can find a bijective map between the latent spaces of different transformers.”