# LLM是判别式的,可却被叫做“生成式AI”?
在统计机器学习理论中,判别式模型指的是学习条件概率$P(Y|X)$的模型(所有必须且仅利用监督信号进行学习的分类/回归,即模式识别),而生成式模型指的是学习联合概率$P(X,Y)$(或仅$P(X)$)的模型(如LDA/RBM/自编码器/扩散/聚类)
LLM是典型的【序列输入】【多分类输出】结构,并利用序列下一个词作为监督信号,因此是典型的判别式模型,而非生成式模型。
这不符合LLM“会说话”给人的直觉。摘取以下观点作为参考:
> I think a big reason many seem to miss is that the common sense meaning of "generative" for the layperson is an appropriate one for generative models. It's not the first case of statistics taking a laypeople's word and giving a very specific and unintuitive new meaning to that word, and then lament the technically wrong usage of the word ([Reddit](https://www.reddit.com/r/MachineLearning/comments/1acsq74/d_why_do_we_keep_calling_generation_models/))
-----
截止2025年,LLM展现的“生成”是基于对文本完成度的判断,而不是对世界的深层理解。也许这与它本质上是一个判别式模型,只能处理“模式”任务不无关系。