Review for NeurIPS paper: Exchangeable Neural ODE for Set Modeling

Weaknesses: - Even though the work is theoretically sound, and the authors do mention that drift functions used in the neural ODE framework have to be Lipschitz continuous, there is no mention if and how Lipschitz continuity can be achieved in general, and if it is achieved for the specific parametrizations (deep set, set transformer) used in this work. More specifically, the vanilla self-attention of (set-)transformer is provably NOT Lipschitz continuous as shown in Kim et. Now, I do realize that the paper I am referencing is an arxiv submission that appeared after the NeurIPS application deadline, and I do not hold this against the authors. However, the authors do not even mention that there was no/authors were not aware of any results on Lipschitz continuity of the used modules. Moreover, deep set need not be Lipschitz continuous if no measures to ensure it are taken.