We introduce the first dataset, MUSIC-AVQA-R, to evaluate the robustness of AVQA models. The construction of this dataset involves two key processes: rephrasing and splitting. The former involves the rephrasing of questions in the test split of MUSIC-AVQA, and the latter is dedicated to the categorization of questions into frequent (head) and rare (tail) subset. We followed the previous work in partitioning the dataset into "head" and "tail" categories. Based on the number of answers in the dataset, answers with a count greater than $1.2$ times the mean, denoted as $\mu(a)$, were categorized as "head" while those with counts less than $1.2\mu(a)$ were categorized as "tail"
We introduce FortisAVQA, a dataset designed to assess the robustness of AVQA models. Its construction involves two key processes: rephrasing and splitting. Rephrasing modifies questions from the test set of MUSIC-AVQA to enhance linguistic diversity, thereby mitigating the reliance of models on spurious correlations between key question terms and answers. Splitting entails the automatic and reasonable categorization of questions into frequent (head) and rare (tail) subsets, enabling a more comprehensive evaluation of model performance in both in-distribution and out-of-distribution scenarios.