Stephan Günnemann

Achieving 3-9x speedups in molecular dynamics simulations without compromising accuracy could revolutionize how we approach atomic system modeling.

Challenging the conventional wisdom that strong certified robustness requires heavy partitioning, this work shows how white-box knowledge of base classifiers in partition-aggregation ensembles can yield significantly tighter robustness guarantees against label-flipping attacks with fewer partitions.

Neural networks can now find all distinct energy levels of the beryllium atom, thanks to a novel wave function architecture that models multiple electronic states simultaneously and trains 200x faster.

LLMs can still be easily fooled by simple prompt rewrites because current adversarial training doesn't adequately cover the data distribution, but a new method using diffusion models closes this gap.