What is the mainstream approach to sentence embedding?

[smartbarbarian]

Is it to use the [CLS] token, or just to take the mean of the output embeddings?