Lenz Pracher
I am currently working on training dynamics in Reinforcement Learning. If you are thinking about something in that direction and want to collaborate, feel free to reach out.
I am a PhD student at Stanford, grateful to be advised by Surya Ganguli. I recently finished my M.Sc. in Theoretical and Mathematical Physics at TUM/LMU, working on Physics of (Machine) Learning in the Rulands Group. Before that I was a working student and intern at Quantco, spent time as a visiting researcher at Harvard Mathematics and at the Hokkaido University Department of Information Science and Technology, was a visiting student at Kyoto University, and did my B.Sc. in Physics at TUM.
Projects
A Spectral Theory of Grokking · arXiv
Weight Decay induces Grokking through Evolution of the Neural Tangent Kernel.
L Pracher, P de Jong, O Liesaus, A Jeffares, S Rulands — Research Note — 2026capsule
A terminal database for reproducible research question + implementation.
L Pracher — Open source software — 2026From simultaneous to leader–follower play in direct reciprocity
Foreknowledge can be very useful to finding good strategies.
P LaPorte, L Pracher, S Pal — PNAS Nexus 5 (2), pgag005 — 2026Strategies of cooperation and defection in five large language models
LLMs have certain strategic preferences.
S Pal, A Mallela, C Hilbe, L Pracher, C Wei, F Fu, S Schnell, MA Nowak — arXiv:2601.09849 — 2026LOCO: Abstracting Spectral Numerical Integrators for PDEs into Neural Operators
We can gain some performance by adjusting the Fourier Neural Operator.
L Pracher, T MatsubaraEvolutionary processes that resolve cooperative dilemmas
Simple update rules can yield surprising behaviour.
P LaPorte, S Wang, L Pracher, S Pal, M Nowak — 2025
Contact
Email: lenz.pracher"at"gmail.com