After author feedback and reviewer discussion, this paper received three final ratings of 8 (clear accept). Reviewers supported the general aim of combining neural and symbolic approaches, and were convinced by the strong empirical results on both the GQA and VQA-CP datasets, and the additional experiments provided in author feedback. The AC agrees with reviewers that this paper will be of significant interest to the NeurIPS vision-and-language community. However, we still encourage the authors to pay close attention to the detailed feedback provided by reviewers, in order to further polish the manuscript.