Abstract
We introduce CAUSAL3D, a comprehensive benchmark spanning 19 datasets designed to evaluate causal reasoning capabilities from visual data. Our evaluation reveals that model performance drops sharply as causal complexity increases, highlighting the need for more causally aware vision systems.
Authors
Disheng Liu*, Yiran Qiao*, Wuche Liu, Yiren Lu, Yunlai Zhou, Tuo Liang, Yu Yin, Jing Ma (*Equal contribution)