Research
Dr. Ashia Wilson directs the Optimization, Safety, and Evaluation (OSE) Lab.
Our research develops the foundations needed to build machine learning systems that are efficient, safe, and meaningfully evaluated. We study how mathematical structure can be used to design better optimization and sampling algorithms, how AI systems can affect human knowledge, judgment, and agency, and how evaluation can support valid conclusions about their capabilities, limitations, and societal consequences.
Optimization
We develop the algorithmic foundations of optimization, sampling, and learning. A central theme of our work is that these areas are deeply connected, and that insights from one can lead to more efficient methods in another. We also study optimization under constraints imposed by modern machine learning systems, including privacy requirements, limited communication or compute, and reduced numerical precision. Across these directions, our goal is to identify the principles that make optimization and sampling algorithms efficient, adaptive, and reliable at modern machine learning scales.
Safety
We develop methods and foundations for understanding and mitigating harms from AI systems. One thrust focuses on child safety and image-based abuse, including methods for detecting and auditing models that have been adapted for harmful purposes without requiring evaluators to generate or inspect illegal or abusive content. A second thrust studies epistemic harms, which arise when AI systems distort what people know, believe, imagine, or regard as possible. This includes homogenization, where the widespread adoption of AI reduce diversity across creative, informational, and institutional outcomes, and sycophancy, where systems prioritize agreement or affirmation over truthfulness, critical reflection, and user agency.
Evaluation
Evaluations drive decisions about which AI systems are trusted, improved, and deployed. We develop methods and foundations for evaluation that better characterize system capabilities, limitations, and societal impacts. Our work asks how evaluations can support valid conclusions, reflect the contexts and consequences in which systems are used, and reveal meaningful differences that standard benchmarks and aggregate metrics may obscure. Across these directions, our goal is to design evaluations that are rigorous, decision-relevant, and aligned with what ultimately matters in practice.