Saved in:
Bibliographic Details
Main Authors: Kim, Joon Hyeok, Park, Yong-Hyun, Østby, Mattis Dalsætra, Gu, Jiatao
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.17673
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908978396528640
author Kim, Joon Hyeok
Park, Yong-Hyun
Østby, Mattis Dalsætra
Gu, Jiatao
author_facet Kim, Joon Hyeok
Park, Yong-Hyun
Østby, Mattis Dalsætra
Gu, Jiatao
contents Despite their empirical success, how diffusion models generalize remains poorly understood from a mechanistic perspective. We demonstrate that diffusion models trained with flow-matching objectives exhibit grokking--delayed generalization after overfitting--on modular addition, enabling controlled analysis of their internal computations. We study this phenomenon across two levels of data regime. In a single-image regime, mechanistic dissection reveals that the model implements modular addition by composing periodic representations of individual operands. In a diverse-image regime with high intraclass variability, we find that the model leverages its iterative sampling process to partition the task into an arithmetic computation phase followed by a visual denoising phase, separated by a critical timestep threshold. Our work provides the mechanistic decomposition of algorithmic learning in diffusion models, revealing how these models bridge continuous pixel-space generation and discrete symbolic reasoning.
format Preprint
id arxiv_https___arxiv_org_abs_2604_17673
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Grokking of Diffusion Models: Case Study on Modular Addition
Kim, Joon Hyeok
Park, Yong-Hyun
Østby, Mattis Dalsætra
Gu, Jiatao
Machine Learning
Despite their empirical success, how diffusion models generalize remains poorly understood from a mechanistic perspective. We demonstrate that diffusion models trained with flow-matching objectives exhibit grokking--delayed generalization after overfitting--on modular addition, enabling controlled analysis of their internal computations. We study this phenomenon across two levels of data regime. In a single-image regime, mechanistic dissection reveals that the model implements modular addition by composing periodic representations of individual operands. In a diverse-image regime with high intraclass variability, we find that the model leverages its iterative sampling process to partition the task into an arithmetic computation phase followed by a visual denoising phase, separated by a critical timestep threshold. Our work provides the mechanistic decomposition of algorithmic learning in diffusion models, revealing how these models bridge continuous pixel-space generation and discrete symbolic reasoning.
title Grokking of Diffusion Models: Case Study on Modular Addition
topic Machine Learning
url https://arxiv.org/abs/2604.17673