Publications

This Is Your Doge, If It Please You: Exploring Deception and Robustness in Mixture of LLMs (preprint)

Placing a single malicious agent in the Mixture of LLMs can nullify all gains achieved. We study the vulnerabilities in the multiple choice passage comprehension and question answering settings and propose unsupervised defense mechanisms that recover a large portion of the lost performance.

Lorenz Wolf, Sangwoong Yoon, Ilija Bogunovic

Private Selection with Heterogeneous Sensitivities (accepted at SaTML 2025)

Investigating differentially private selection mechanisms with heterogeneous sensitivities.

Daniela Antonova, Allegra Laro, Audra McMillan, Lorenz Wolf

Augmented Modular Reinforcement Learning based on Heterogeneous Knowledge (arxiv preprint)

Proposing a hierarchical command arbitration architecture to flexibly incorporate heterogeneous decision-making modules.

Lorenz Wolf, Mirco Musolesi

F-EBM: Energy Based Learning of Functional Data (AISTATS 2023)

We propose an energy based generative model to synthesise functional data.

Jenning Lim, Sebastian Vollmer, Lorenz Wolf, Andrew Duncan

F-EBM: Energy Based Learning of Functional Data (AISTATS 2023)