Three of the four reports are clearly positive. The only review that is somewhat critical explains its lower score by lack of direct applications. I agree, but do not see that as a problem: NeurIPS has a long history of publishing foundational work, the topic of the paper falls clearly within the remit of the conference, and foundational and theoretical work on embeddings is still rare; I would consider this work a welcome exception. The rebuttal promises a number of revisions; please make sure to add these in the final version.