Authors propose a bilevel optimization based method for selecting coresets for continual learning and online applications. All reviewers agree on the merits of the submission. The raised issues are largely minor except the criticism on the experimental setup. I think the empirical study can be improved but the weakness is a rather acceptable one as continual learning is a new and upcoming topic without established baselines and empirical settings. Hence, many (possibly most) continual learning uses different empirical setups. Hence, I believe this is an acceptable weakness. Issues to fix in the camera ready: Authors should extend the discussion on the limitations of the work for neural networks.