After fairly in-depth discussion, the reviewer consensus for this paper has shifted towards acceptance. The reviewers agree that the paper presents a simple but general framework for implicitly learning to reason under some constraints. The paper makes an important theoretical contribution, of a type which is welcome but perhaps too rarely is included in NeurIPS and similar conferences, and should be of interest to the community. The reviewers, however, all flagged that the paper could use some fairly significant improvement when it comes to presentation and clarity, and that it would benefit from clear exposition of the formalization section, and running/motivating examples to serve as an intuition pump for those readers with a less formal background. Overall, acceptable, but the authors are *strongly* encouraged to take heed of the recommendations of all three reviewers when preparing their camera ready, should the paper be accepted by the PC.