Two reviewers see value in the paper, while two are somewhat swayed by the rebuttal but find that the paper's overall contribution is not quite at the (high) bar that is set by NeurIPS. One point of widespread disagreement is about whether the method is "simple". The rebuttal clarifies that "simple" should mean fewer lines of code. However, there is no value in fewer lines of code per se -- the real value is that the new idea has theoretical interest or practical merit. On the former, there is some agreement "the paper seems like it has an interesting idea at its core", " explores an interesting and potential useful problem setting", but there is also general agreement that the improvement is relatively small, that artefacts still exist, and that the problem domain, although interesting, is relevant only to a subset of the NeurIPS audience. In that sense, it is reasonable for the reviewers to require more comprehensive explorations. On one point: the issue of linear function*al* vs linear function was understood by the reviewers, who also understand that a low-finite-dimensional representation of the functional might indeed not be as rich as an MLP. The rebuttal somewhat sidesteps this point, but it should be addressed much more openly in any final version of this paper.