1 paper
Rong Yu Xu, Prayag Tiwari, Shaolei Zhang
Vision-language models (VLMs) can answer simple visual questions, but often struggle when one question requires several visual judgments. We study this gap with controlled tasks fo…