Tag
This paper introduces a two-stage neuro-symbolic framework that uses weak supervision (as little as 1% labels) with a slot-based VAE to learn interpretable symbols for object-centric visual reasoning, outperforming foundation models in domain generalization.