Saved in:
Bibliographic Details
Main Authors: Porras-Valenzuela, Javier, Hadou, Samar, Ribeiro, Alejandro
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2601.17257
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909999822798848
author Porras-Valenzuela, Javier
Hadou, Samar
Ribeiro, Alejandro
author_facet Porras-Valenzuela, Javier
Hadou, Samar
Ribeiro, Alejandro
contents We introduce a constrained optimization framework for training transformers that behave like optimization descent algorithms. Specifically, we enforce layerwise descent constraints on the objective function and replace standard empirical risk minimization (ERM) with a primal-dual training scheme. This approach yields models whose intermediate representations decrease the loss monotonically in expectation across layers. We apply our method to both unrolled transformer architectures and conventional pretrained transformers on tasks of video denoising and text classification. Across these settings, we observe constrained transformers achieve stronger robustness to perturbations and maintain higher out-of-distribution generalization, while preserving in-distribution performance.
format Preprint
id arxiv_https___arxiv_org_abs_2601_17257
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle A Constrained Optimization Perspective of Unrolled Transformers
Porras-Valenzuela, Javier
Hadou, Samar
Ribeiro, Alejandro
Machine Learning
We introduce a constrained optimization framework for training transformers that behave like optimization descent algorithms. Specifically, we enforce layerwise descent constraints on the objective function and replace standard empirical risk minimization (ERM) with a primal-dual training scheme. This approach yields models whose intermediate representations decrease the loss monotonically in expectation across layers. We apply our method to both unrolled transformer architectures and conventional pretrained transformers on tasks of video denoising and text classification. Across these settings, we observe constrained transformers achieve stronger robustness to perturbations and maintain higher out-of-distribution generalization, while preserving in-distribution performance.
title A Constrained Optimization Perspective of Unrolled Transformers
topic Machine Learning
url https://arxiv.org/abs/2601.17257