Review for NeurIPS paper: Learning Graph Structure With A Finite-State Automaton Layer

Additional Feedback: This is great work and should definitely be accepted. A few thoughts and comments on the follow-up: * The fact that observations depend on the initial node (L114) introduces a limited form of "variable capturing" in the regular language that the POMDP tries to approximate. It is still regular, but this can be broadened to depend on the whole agent's history and thus represent more complex languages. The POMDP would no longer be theoretically guaranteed to represent such a language, but it might still learn a useful one. I wonder what would happen if one used GFSA for self-supervised program learning like CodeBERT [4] rather than supervised program analysis tasks.