About
I am a PhD student at MIT, where I am fortunate to be advised by Chara Podimata. I work on statistical foundations for the control and evaluation of AI systems, motivated by safety and trustworthiness considerations as well as future regulatory pressures. My research revolves around two questions:
How can LLMs model their uncertainty, and how can this uncertainty be used to design systems that satisfy rigorous safety guarantees?
How can humans understand, monitor and audit the capabilities of AI systems?
For the former, my work develops a framework of risk-constrained generation, combining modern LLM inference techniques with classical tools from statistical inference. To address the latter, I develop tools that help us evaluate, monitor, and perform audits of black-box models more efficiently. I enjoy applying my research to real-world problems (New York Times coverage).
Previously, I was an undergrad and master's student at ETH Zürich, where I was lucky to be part of the Learning and Adaptive Systems group of Andreas Krause. I worked on online learning, continual learning and meta-learning, as well as optimization algorithms for reinforcement learning. During my studies I was generously supported by the Zeno-Karl Schindler Foundation and the Swiss Study Foundation, facilitating visits at the University of Copenhagen and MIT.
Papers
-
Nicolas Emmenegger*; Theo X. Olausson*; Armando Solar-Lezama; Chara Podimata Conformal Language Modelling via Posterior Sampling. In Conference on Neural Information Processing Systems (NeurIPS 2026) & Spotlight at Agentic UQ workshop at ICML 2026.
-
Nicolas Emmenegger; Ellery Stahler; Chara Podimata Prediction-Powered Inference across many Tasks for AI Evaluation and Social Science Research. In Conference on Neural Information Processing Systems (NeurIPS 2026).
-
Nicolas Emmenegger*; Mojmír Mutný*; Andreas Krause Likelihood Ratio Confidence Sets for Sequential Decision Making. In Conference on Neural Information Processing Systems (NeurIPS 2023).
-
Parnian Kassraie; Nicolas Emmenegger; Andreas Krause Anytime Model Selection in Linear Bandits. In Conference on Neural Information Processing Systems (NeurIPS 2023).
-
Nicolas Emmenegger; Rasmus Kyng; Ahad. N. Zehmakan On the Oracle Complexity of Higher-Order Smooth Non-Convex Finite-Sum Optimization. In International Conference on Artificial Intelligence and Statistics (AISTATS 2022).
