AI for learning mathematical structure
Using neural networks to learn topology and geometry directly from data.
Research
Research directions
Mechanistic interpretability for discovery and safety
Opening up trained models to find features and circuits.
Publications
2026
Interpreting latent protein language model features with geometric annotations
arXiv preprint, Aug 2026
An automated method for annotating sparse-autoencoder features in ESM-2 using geometric descriptors of the protein backbone, revealing residue-level structure that database and sequence labels miss.
Shortcut learning in geometric knot classification
Machine Learning: Science and Technology, Jun 2026
Neural networks trained on molecular dynamics knot data can exploit hidden non-topological features instead of topology. We expose these shortcuts and release a dataset and generator that explore geometric state space at fixed knot type.
Writhe-based polymer link classification using machine learning
arXiv preprint, Jul 2026
A feedforward network trained on the writhe density matrix classifies thermally equilibrated configurations of the first six prime two-component links with 97% accuracy, robustly across temperatures and chain lengths.
† first or joint first author