The Representations of Deep Neural Networks Trained on Dihedral Group Multiplication

Download PDF

Gavin McCracken, Sihui Wei, Gabriela Moisescu-Pareja, Harley Wiltzer, Jonathan Love

Published: 23 Sept 2025, Last Modified: 29 Oct 2025NeurReps 2025 PosterEveryoneRevisionsBibTeXCC BY 4.0
Keywords: modular addition, cosets, mechanistic interpretability, manifold hypothesis, universality hypothesis
Abstract: We find coset and approximate coset circuits play a key role in how multilayer perceptrons learn dihedral group multiplication, consistent with recent findings on modular addition.
Submission Number: 115