research

IESL develops machine learning methods that help people extract knowledge, reason with uncertainty, and accelerate discovery from complex data. The lab’s work sits at the intersection of natural language processing, probabilistic modeling, information extraction, and large-scale generative AI.

We are motivated by a broad question: how can intelligent systems build, revise, and use knowledge in ways that are reliable, adaptable, and useful for science and society?

Learning to Build Knowledge

Modern AI systems should do more than predict the next token. They should organize evidence, uncover structure, and connect information across documents, datasets, and domains. We study models that can turn unstructured information into actionable knowledge, with an emphasis on systems that remain grounded in data and useful to human experts.

Reasoning Under Uncertainty

Many important problems require making decisions with incomplete information: searching for better solutions, proposing hypotheses, planning experiments, or deciding what evidence to gather next. We develop probabilistic and Bayesian methods that allow language models and other AI systems to reason about uncertainty, explore alternatives, and improve through feedback.

AI for Discovery

We are especially interested in AI systems that can assist scientific and technical discovery: formulating hypotheses, designing searches, interpreting results, and helping researchers navigate large spaces of possible explanations. This requires combining language understanding, statistical reasoning, and domain knowledge in systems that support human judgment rather than replace it.

New Foundations for Generative Models

Text and other symbolic data are not always best generated one step at a time. We investigate new paradigms for generative modeling, including iterative refinement, and diffusion-inspired approaches for discrete data. Our goal is to make generative systems more flexible, controllable, and aligned with the way complex artifacts are created and revised in real workflows. For recent work in this area, check out dIESL.