GVQA (from VQA-CP) builds on stacked attention networks (SAN).13 However,SAN and,byextension, GVQA architectures donotevaluate for,andgeneralize poorly on,17 unseen object attributes (CLEVR-CoGenT) and linguistic structural pattern (CLOSURE) combinations. The language parser is not trained, constructs text (s) object graphs (Gs) using rules-based entity38 recognizer [L126].(W3)CLOSURE ClarityMinor clarifications -5a: corrected inthecamera-ready version.

We thank all the reviewers for their insightful questions, comments and commendations (novelty, clarity, performance). T ask 1 motivation, complexity @R3, R4: Motivation behind Task 1 is efficacy, not complexity [L46-49]. GVQA (from VQA-CP) builds on stacked attention networks (SAN). Thus, they are orthogonal to MGN in problem setting and architecture. We show results using MAC (from the GQA authors, L291) with both CLEVR and GQA results.