The paper is interesting and promising but the three reviewers do agree that this is a borderline paper around the acceptance threshold. The main concerns they have expressed are the following: (i) "incremental" application of reinforcement/imitation learning, (ii) weak experimental work and unsufficient benchmarking to other methods. I personally find it has its merits as an end-to-end application of imitation learning in quite an original context and I think it should be confronted to other application papers in the same domain (if any). Now to answer the points made by the reviewers, I think that (a) modeling the learning agent in this context is far from incremental as it combines intimate knowledge of compilers and solid mastering of graph theory and machine learning methodology; as a matter of fact, it certainly brings a paradigm shift in that community, (b) their experiments cover the case of instruction packing which was very recently related to ILP and they discuss the performance of their method wrt to optimal methods and to existing compilers, which is the best they can do at this stage. Given these two points, I tend to bump up slightly the assessment made by Reviewers #1 an #2 and push for acceptance.