Block Circulant Adapter for Large Language Models

Abstract

Fine-tuning large language models (LLMs) is difficult due to their huge model size. Recent Fourier domain-based methods show potential for reducing fine-tuning costs. We propose a block circulant matrix-based fine-tuning method with a stable training heuristic to leverage the properties of circulant matrices and one-dimensional Fourier transforms to reduce storage and computation costs. Experiments show that our method uses 14× less number of parameters than VeRA, 16× smaller than LoRA and 32× less FLOPs than FourierFT, while maintaining close or better task performance. Our approach presents a promising way in frequency domain to fine-tune large models on downstream tasks.

Cite

Text

Ding et al. "Block Circulant Adapter for Large Language Models." International Joint Conference on Artificial Intelligence, 2025. doi:10.24963/IJCAI.2025/560

Markdown

[Ding et al. "Block Circulant Adapter for Large Language Models." International Joint Conference on Artificial Intelligence, 2025.](https://mlanthology.org/ijcai/2025/ding2025ijcai-block/) doi:10.24963/IJCAI.2025/560

BibTeX

@inproceedings{ding2025ijcai-block,
  title     = {{Block Circulant Adapter for Large Language Models}},
  author    = {Ding, Xinyu and Wang, Meiqi and Liao, Siyu and Wang, Zhongfeng},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year      = {2025},
  pages     = {5030-5038},
  doi       = {10.24963/IJCAI.2025/560},
  url       = {https://mlanthology.org/ijcai/2025/ding2025ijcai-block/}
}