Shayan Kiyani
I am a fourth-year PhD student at the University of Pennsylvania, where I am fortunate to work with Hamed Hassani and George Pappas. Before starting my PhD, I was a scientific intern at the Institute of Science and Technology Austria (ISTA), where I worked with Marco Mondelli for a year. I received my bachelor’s degree in Electrical Engineering from Sharif University of Technology, with a major in Communication Systems and a minor in Mathematics.
Broadly speaking, my research interests span distribution-free uncertainty quantification, decision-making under uncertainty, human-AI collaboration, and the alignment of AI uncertainty estimates with human preferences and downstream decisions.
You can find my CV here. For a full list of publications, see my Google Scholar profile.
.jpg)
Research
As AI advances to new generations every few years, both the models we build and the data we feed them are becoming increasingly complex and harder to understand. This creates a central challenge: how can we reliably deploy AI systems we do not fully understand into the real world, where they may inform medical decisions, scientific discovery, automated reasoning, and other high-stakes human workflows? I work toward this question. At the center of my research is the development of distribution-free, black-box approaches to uncertainty quantification for machine learning models. My goal is to use these tools to design better interfaces between ML predictions and the real world, so that downstream users can act on model outputs more reliably. This agenda has taken shape across several connected directions:
Principles for connecting predictions to actions
I study the decision-theoretic foundations of uncertainty quantification: what form of uncertainty should be communicated to a downstream decision maker, and what actions should it induce? This line of work aims to make predictive uncertainty operational, rather than merely descriptive.
Human-AI collaboration
A central theme in my work is how uncertainty and interaction should be designed when humans and AI systems collaborate. I am interested in frameworks that preserve human strengths, avoid counterfactual harm, and enable AI systems to be genuinely helpful in a range of scenarios.
Conformal prediction
Conformal prediction is a recurring foundation throughout my research. I work on making conformal methods more informative, efficient, and robust, including questions around conditional guarantees, distribution shift, online settings, and new uncertainty quantification challenges arising in modern generative models.