HiLoRA: High-frequency-augmented Low-Rank Adaptation
Keywords: Large Language Models, LoRA, Frequency, Catastrophic Forgetting
Abstract: As large language models (LLMs) have demonstrated remarkable performance, parameter-efficient fine-tuning (PEFT) has emerged as an important paradigm. As a solution, low-rank adaptation (LoRA) freezes the pre-trained weights and introduces small learnable adapters instead of fine-tuning the full set of parameters. However, LoRA suffers from $\textit{catastrophic forgetting}$, where pre-trained knowledge is overwhlemed and forgotten as new information is learned. One cause of this issue is $\textit{implicit regularization}$, where deep learning models tend to favor more generalized solutions. This tendency leads to a significant increase in the largest singular values of the weights, which correspond to low-frequency components. To address this problem, we propose an advanced LoRA that balances the retention of pre-trained knowledge with the learning of new information. Since fine-tuning involves learning fine-grained details, which correspond to high-frequency information, we designed HiLoRA, a method that injects learnable high-frequency components into the pre-trained model. By leveraging the parameterized SVD and constraining singular values to appropriate levels, HiLoRA adapts to new tasks by focusing on the high-frequency domain with minimal change from the pre-trained weights. To evaluate the effectiveness of HiLoRA, we conduct extensive experiments on natural language understanding and question answering tasks. The results show that HiLoRA not only improves performance but also effectively retains pre-trained knowledge compared to baseline models.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3646
Loading