Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Shi, Kexuan, Li, Hanxuan, Qiu, Zeju, Wen, Yandong, Buchholz, Simon, Liu, Weiyang
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2605.12492
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913119199035392
author Shi, Kexuan
Li, Hanxuan
Qiu, Zeju
Wen, Yandong
Buchholz, Simon
Liu, Weiyang
author_facet Shi, Kexuan
Li, Hanxuan
Qiu, Zeju
Wen, Yandong
Buchholz, Simon
Liu, Weiyang
contents We introduce Pion, a spectrum-preserving optimizer for large language model (LLM) training based on orthogonal equivalence transformation. Unlike additive optimizers such as Adam and Muon, Pion updates each weight matrix through left and right orthogonal transformations, preserving its singular values throughout training. This yields an optimization mechanism that modulates the geometry of weight matrices while keeping their spectral norm fixed. We derive the Pion update rule, systematically examine its design choices, and analyze its convergence behavior along with several key properties. Empirical results show that Pion offers a stable and competitive alternative to standard optimizers for both LLM pretraining and finetuning.
format Preprint
id arxiv_https___arxiv_org_abs_2605_12492
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation
Shi, Kexuan
Li, Hanxuan
Qiu, Zeju
Wen, Yandong
Buchholz, Simon
Liu, Weiyang
Machine Learning
We introduce Pion, a spectrum-preserving optimizer for large language model (LLM) training based on orthogonal equivalence transformation. Unlike additive optimizers such as Adam and Muon, Pion updates each weight matrix through left and right orthogonal transformations, preserving its singular values throughout training. This yields an optimization mechanism that modulates the geometry of weight matrices while keeping their spectral norm fixed. We derive the Pion update rule, systematically examine its design choices, and analyze its convergence behavior along with several key properties. Empirical results show that Pion offers a stable and competitive alternative to standard optimizers for both LLM pretraining and finetuning.
title Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation
topic Machine Learning
url https://arxiv.org/abs/2605.12492