Saved in:
Bibliographic Details
Main Authors: Amarel, James, Hengartner, Nicolas, Miller, Robyn, Singh, Kamaljeet, Mansingh, Siddharth, Mohan, Arvind, Migliori, Benjamin, Casleton, Emily, Skurikhin, Alexei, Lawrence, Earl, Kunde, Gerd J.
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2509.00024
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909992875982848
author Amarel, James
Hengartner, Nicolas
Miller, Robyn
Singh, Kamaljeet
Mansingh, Siddharth
Mohan, Arvind
Migliori, Benjamin
Casleton, Emily
Skurikhin, Alexei
Lawrence, Earl
Kunde, Gerd J.
author_facet Amarel, James
Hengartner, Nicolas
Miller, Robyn
Singh, Kamaljeet
Mansingh, Siddharth
Mohan, Arvind
Migliori, Benjamin
Casleton, Emily
Skurikhin, Alexei
Lawrence, Earl
Kunde, Gerd J.
contents Foundation models trained as autoregressive PDE surrogates hold significant promise for accelerating scientific discovery through their capacity to both extrapolate beyond training regimes and efficiently adapt to downstream tasks despite a paucity of examples for fine-tuning. However, reliably achieving genuine generalization - a necessary capability for producing novel scientific insights and robustly performing during deployment - remains a critical challenge. Establishing whether or not these requirements are met demands evaluation metrics capable of clearly distinguishing genuine model generalization from mere memorization. We apply the influence function formalism to systematically characterize how autoregressive PDE surrogates assimilate and propagate information derived from diverse physical scenarios, revealing fundamental limitations of standard models and training routines in addition to providing actionable insights regarding the design of improved surrogates.
format Preprint
id arxiv_https___arxiv_org_abs_2509_00024
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Generalization vs. Memorization in Autoregressive Deep Learning: Or, Examining Temporal Decay of Gradient Coherence
Amarel, James
Hengartner, Nicolas
Miller, Robyn
Singh, Kamaljeet
Mansingh, Siddharth
Mohan, Arvind
Migliori, Benjamin
Casleton, Emily
Skurikhin, Alexei
Lawrence, Earl
Kunde, Gerd J.
Computational Physics
Machine Learning
Foundation models trained as autoregressive PDE surrogates hold significant promise for accelerating scientific discovery through their capacity to both extrapolate beyond training regimes and efficiently adapt to downstream tasks despite a paucity of examples for fine-tuning. However, reliably achieving genuine generalization - a necessary capability for producing novel scientific insights and robustly performing during deployment - remains a critical challenge. Establishing whether or not these requirements are met demands evaluation metrics capable of clearly distinguishing genuine model generalization from mere memorization. We apply the influence function formalism to systematically characterize how autoregressive PDE surrogates assimilate and propagate information derived from diverse physical scenarios, revealing fundamental limitations of standard models and training routines in addition to providing actionable insights regarding the design of improved surrogates.
title Generalization vs. Memorization in Autoregressive Deep Learning: Or, Examining Temporal Decay of Gradient Coherence
topic Computational Physics
Machine Learning
url https://arxiv.org/abs/2509.00024