ARD-LoRA: Dynamic Rank Allocation for Parameter-Efficient Fine-Tuning of Foundation Models with Heterogeneous Adaptation Needs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Shinwari, Haseeb Ullah Khan, Usama, Muhammad
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866915682932752384
author Shinwari, Haseeb Ullah Khan
Usama, Muhammad
author_facet Shinwari, Haseeb Ullah Khan
Usama, Muhammad
contents Conventional Low-Rank Adaptation (LoRA) methods employ a fixed rank, imposing uniform adaptation across transformer layers and attention heads despite their heterogeneous learning dynamics. This paper introduces Adaptive Rank Dynamic LoRA (ARD-LoRA), a novel framework that automates rank allocation through learnable scaling factors. These factors are optimized via a meta-objective balancing task performance and parameter efficiency, incorporating $\ell_1$ sparsity for minimal rank and Total Variation regularization for stable rank transitions. ARD-LoRA enables continuous, differentiable, per-head rank adaptation. Experiments on LLAMA-3.1-70B and PaliGemma-2 demonstrate ARD-LoRA's efficacy, achieving up to 99.3% of full fine-tuning performance with only 0.32% trainable parameters, outperforming strong baselines like DoRA and AdaLoRA. Furthermore, it reduces multimodal adaptation memory by 41%. These results establish dynamic, fine-grained rank allocation as a critical paradigm for efficient foundation model adaptation.
format Preprint
id arxiv_https___arxiv_org_abs_2506_18267
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ARD-LoRA: Dynamic Rank Allocation for Parameter-Efficient Fine-Tuning of Foundation Models with Heterogeneous Adaptation Needs
Shinwari, Haseeb Ullah Khan
Usama, Muhammad
Machine Learning
Artificial Intelligence
Conventional Low-Rank Adaptation (LoRA) methods employ a fixed rank, imposing uniform adaptation across transformer layers and attention heads despite their heterogeneous learning dynamics. This paper introduces Adaptive Rank Dynamic LoRA (ARD-LoRA), a novel framework that automates rank allocation through learnable scaling factors. These factors are optimized via a meta-objective balancing task performance and parameter efficiency, incorporating $\ell_1$ sparsity for minimal rank and Total Variation regularization for stable rank transitions. ARD-LoRA enables continuous, differentiable, per-head rank adaptation. Experiments on LLAMA-3.1-70B and PaliGemma-2 demonstrate ARD-LoRA's efficacy, achieving up to 99.3% of full fine-tuning performance with only 0.32% trainable parameters, outperforming strong baselines like DoRA and AdaLoRA. Furthermore, it reduces multimodal adaptation memory by 41%. These results establish dynamic, fine-grained rank allocation as a critical paradigm for efficient foundation model adaptation.
title ARD-LoRA: Dynamic Rank Allocation for Parameter-Efficient Fine-Tuning of Foundation Models with Heterogeneous Adaptation Needs
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2506.18267