Florian Bordes

we release a cognitively-inspired benchmark for reasoning across scenes that reveals hallucination is an open challenge for multimodal models

object-centric inductive biases to a address the binding problem in CLIP-like pretraining