Privacy Pass docs [0] cover this, but it is mostly referenced deeper in the paper. I believe the idea is that the tokens returned by the server are "unlinkable" to the (modified) tokens passed back by the client. So the server knows it passed back tokens A, B and C to some users, and later receives tokens X, Y and Z. It knows that X, Y and Z are valid, but not their correspondance to the tokens it issued. It uses elliptic curve cryptography for this.
[0] https://privacypass.github.io/