Postdoc Neurosymbolic Reasoning in Multimodal World Models

While the performance of multimodal world models has significantly improved, they still lack commonsense abstraction and reasoning capabilities. These models fail to capture the causal aspects of the physical and social world. To address the grand challenge of commonsense abstraction and reasoning with multimodal world models, we will explore methods to enhance their robustness, abstraction, and visual grounding abilities. We will focus on learning explicit, effective abstractions from limited human-labeled data, serving as dynamic, neurosymbolic world models. Following cognitive principles, we will evaluate these methods on tasks that require commonsense reasoning over space, time, and causality.