CORA: Per-Slice Coherent Orthogonal Rotation for SVD-based Low-Rank Adaptation

arXiv:2607.02576 · stat.ML, cs.LG · Submitted 2026-06-30 · Read on arXiv

Pengcheng Wang, Ziran Liu, Wei Wang, Wei Jiang

stat.ML, cs.LG

Submitted: 2026-06-30

Comments: 13 pages, 1 figure

Code: https://github.com/sahil280114/codealpaca

License: http://creativecommons.org/licenses/by/4.0/

The gist: Parameter-Efficient Fine-Tuning (PEFT) commonly adapts pretrained weights through low-rank updates, and recent methods further exploit the singular value decomposition (SVD) of the base weight for

Terminology

Abstract

Parameter-Efficient Fine-Tuning (PEFT) commonly adapts pretrained weights through low-rank updates, and recent methods further exploit the singular value decomposition (SVD) of the base weight for initialization or subspace selection. However, these methods do not explicitly preserve the coupled geometry between the pretrained left and right singular bases. Motivated by recent minimum-perturbation theory, which shows that stable finetuning follows a coherent SVD rotation in which a single orthogonal Q acts on both the left singular basis U 0 and the right singular basis V 0, we prove a per-slice analogue: each row slice of W 0 can be adapted by a shared orthogonal rotation Q i on its left basis U i and right basis V i together with a diagonal spectrum shift. We implement this form as CORA (Coherent Orthogonal Rotation Adaptation), which applies per-slice orthogonal rotations and a per-layer diagonal scale to the rank- r SVD truncation of W 0. CORA uses 2 m(r - 1) trainable parameters per linear layer, about 4 times fewer than LoRA at the same rank. CORA outperforms LoRA, DoRA, PiSSA, and MiLoRA on commonsense reasoning and code generation while using about 8 times fewer parameters.

Sources

Related papers