The original XMSS paper (eprint 2011/484) discusses an older version of XMSS, which does use rather larger public keys (and I suspect they got the signature size slightly wrong - they forget the idx_sig part of the signature which specifies which WOTS+ leaf they're using).
The XMSS RFC defines a newer version, with considerably smaller public keys (and slightly larger signatures). The more recent XMSS/LMS comparison paper which you cited (eprint 2017/349) uses the RFC version of XMSS for the comparison.
I believe that the version defined in the RFC is what people generally use, and so the values in the comparison paper are the ones which are accurate.
As for the differences between the original version of XMSS and the RFC, well,
here's the main difference: in the original paper used masks stored in the public keys to provide randomization; the problem with this is, because they use the same masks in different parts of the Merkle tree, this opens them up to multitarget preimage attacks. In the RFC version, they compute the masks dynamically (and hence they don't need to be in the public key). Because each mask is now different, multitarget preimage attacks don't apply (at the cost of the computation needed to generate the masks dynamically).
A more minor difference is that the RFC version includes a randomizer 'r' in with the signature (which is hashed along with the message); making this randomizer be unpredictable means that we don't depend on the collision resistance of the hash function (given that we rely on preimage or second preimage resistance everywhere else, this is a win).