Thanks for your submission to NeurIPS. This paper was very much a borderline paper, with two accept scores and one reject score. One of the concerns raised by the negative reviewer was that, while the algorithm can achieve an approximation to the best merge at each step, it is unclear how the final clustering results would compare to the standard algorithm. The authors addressed this in their rebuttal, which helped. Also, there were some issues raised about experiments as well as various minor suggestions (typos etc.). In general it seems that the concerns are mostly minor, and on the whole this paper seems to make an interesting and worthwhile contribution, so I am recommending that the paper is accepted.