Optimal Brain Decomposition for Accurate LLM Low-Rank Approximation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Li, Yuhang, Lee, Donghyun, Yin, Ruokai, Panda, Priyadarshini
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910094123335680
author Li, Yuhang
Lee, Donghyun
Yin, Ruokai
Panda, Priyadarshini
author_facet Li, Yuhang
Lee, Donghyun
Yin, Ruokai
Panda, Priyadarshini
contents Low-rank decomposition has emerged as an important problem in Large Language Model (LLM) fine-tuning and inference. Through Singular Value Decomposition (SVD), the weight matrix can be factorized into low-rank spaces optimally. Previously, a common practice was to decompose the weight in the activation-whitened space, and then achieve satisfying results. In this work, we propose Optimal Brain Decomposition LLM (OBD-LLM), which studies the decomposition problem in the model space by utilizing second-order Hessian information. Through a rigorous Kronecker-factorization of the Hessian, we show that the decomposition needs to consider both input and output information of the layer, and achieves much better decomposition results compared to input only method. Our loss-aware decomposition method involves a bi-directional whitening on the weight matrix. As a result, OBD-LLM is a closed-form solution for the optimal decomposition of weights in the language model. Remarkably, we achieve ~20-40\% better results than previous state-of-the-art decomposition methods, the SVD-LLM.
format Preprint
id arxiv_https___arxiv_org_abs_2604_00821
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Optimal Brain Decomposition for Accurate LLM Low-Rank Approximation
Li, Yuhang
Lee, Donghyun
Yin, Ruokai
Panda, Priyadarshini
Machine Learning
Low-rank decomposition has emerged as an important problem in Large Language Model (LLM) fine-tuning and inference. Through Singular Value Decomposition (SVD), the weight matrix can be factorized into low-rank spaces optimally. Previously, a common practice was to decompose the weight in the activation-whitened space, and then achieve satisfying results. In this work, we propose Optimal Brain Decomposition LLM (OBD-LLM), which studies the decomposition problem in the model space by utilizing second-order Hessian information. Through a rigorous Kronecker-factorization of the Hessian, we show that the decomposition needs to consider both input and output information of the layer, and achieves much better decomposition results compared to input only method. Our loss-aware decomposition method involves a bi-directional whitening on the weight matrix. As a result, OBD-LLM is a closed-form solution for the optimal decomposition of weights in the language model. Remarkably, we achieve ~20-40\% better results than previous state-of-the-art decomposition methods, the SVD-LLM.
title Optimal Brain Decomposition for Accurate LLM Low-Rank Approximation
topic Machine Learning
url https://arxiv.org/abs/2604.00821