Speech input amplifies partial failures in multimodal models compared to text input

speech amplifying partial failures

http://arxiv.org/abs/2608.27135v1