Team Members: Eli Laird, Tim Lee
World Models provide neural networks with an understanding of the underlying dynamics encoded in the training data. While these world models have shown improved performance in computer vision and audio tasks, they struggle to learn complex dynamics between many systems. Graph-based world models, on the other hand, provide strong inductive biases for learning interactions between objects, making them ideal for simulating complex systems.
Our research focuses on developing graph-based World Models using self-supervised learning and graph neural networks. We build upon our recent proposal, the Message-Passing Joint Embedding Predictive Architecture (MP-JEPA), which has achieved state-of-the-art results in several graph-related tasks.
We are enhancing MP-JEPA by adapting it for heterogeneous graphs, incorporating E(3) equivariant graph neural networks, and developing multi-modal models. Furthermore, we are exploring MP-JEPA's potential to train World Models that can enhance quantum molecular property prediction, generate ligands for drug discovery, and improve knowledge-graph reasoning for AI agents.