Our work spans foundation models, AI for science, and trustworthy AI — developing methods that are capable, reliable, and broadly useful.
We study the understanding, training, and evaluation of large-scale generative models — spanning language, vision, and multimodal reasoning. Our work covers model capabilities, alignment with human intent, and the design of agents that can reason and act across complex tasks.
We apply generative and predictive models to accelerate scientific discovery in chemistry (molecular design, reaction prediction, material discovery), biology, and physics. We also develop model-based simulation for the social sciences, uncovering patterns in behavior, norms, and policy.
We develop methods for alignment, safety, fairness, and interpretability of AI systems. This includes evaluation frameworks for detecting failures and biases in foundation models, as well as guardrails and oversight mechanisms for reliable deployment.
