Stochastic Parrots or True Virtuosos? Digging Deeper Into the Audio-Video Understanding of AVQA Models

AbstractAudio-video question answering (AVQA) systems for music show signs of multimodal "understanding", but it is unclear which inputs they rely on or whether their behavior reflects genuine audio-video reasoning. Existing evaluations focus on overall accuracy and rarely examine modality dependence. We address this gap by suggesting a method of using counterfactual evaluations to analyse the audio-video understanding of the models, illustrated with a case study on the audio-video spatial-temporal (AVST) architecture. This includes interventions that zero out or swap audio, video, or both, where results are benchmarked against a baseline based on linguistic patterns alone. Results show stronger reliance on audio than video, yet performance persists when either modality is removed, indicating learned cross-modal representations. The AVQA system studied thus exhibits non-trivial multimodal integration, though its "understanding" remains uneven.