This paper is very much borderline and sparked an extensive discussion among reviewers. On the positive side, this work presents a simple closed form generation rule for rank-1 lattice in QMC, which previously required exhaustive search. The method is novel and solid, with promising empirical results. On the negative sides, all reviewers have concerns with 1) the lack of comparison to methods in ML community, and the fitness of the venue (however, the target problem of integration approximation is of high importance in Neurips as well); 2) some limited clarity/questions on empirical methodology; and 3) some writing quality /typo issues. The authors did an excellent job in rebuttal, including providing some initial results on a toy restricted Boltzmann model. We encourage the authors to add more comparisons to MC/MCMC methods that are more widely used in machine learning and discuss its applicability to the high dimensional and graph-structured integration problems in machine learning. We hope this work can provide an opportunity for Neurips audience to learn about QMC.