Research

My research aims to reimagine the current generative AI landscape by making foundation models more glass-box, explainable, and efficient.

My core approach is to fuse the complementary strengths of deep generative models (like LLMs and neural networks) with the formal guarantees of tractable probabilistic graphical models. While deep models are masters of learning from data, they often lack robust inference capabilities. Tractable models, conversely, provide a powerful framework for efficient and exact reasoning.

By integrating these paradigms, I develop new algorithms that augment large pre-trained systems with enhanced reasoning capacity. This enables complex probabilistic questions to be answered precisely, leading to more reliable control over model behavior and a clearer window into their decision-making processes.

Key Research Focus

• Methodology for Tractable Models: Designing and applying tractable probabilistic graphical models to new challenges in AI.
• Generative + Probabilistic Models for Efficient Inference: Developing scalable algorithms that bring formal reasoning to deep learning systems.
• Applications in AI Trustworthiness: Focusing on explainability, controllability, and safety for foundation models.