Paper Detail
Kexuan Shi, Hanxuan Li, Zeju Qiu, Yandong Wen, Simon Buchholz, Weiyang Liu
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.
No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.
No ranking explanation is available yet.
No tags.
@article{shi2026pion,
title = {Pion: A Spectrum-Preserving Optimizer via Orthogonal Equivalence Transformation},
author = {Kexuan Shi and Hanxuan Li and Zeju Qiu and Yandong Wen and Simon Buchholz and Weiyang Liu},
year = {2026},
abstract = {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 exa},
url = {https://arxiv.org/abs/2605.12492},
keywords = {cs.LG, stat.ML},
eprint = {2605.12492},
archiveprefix = {arXiv},
}
{}