GEM: A Scale-Aware and Distribution-Sensitive Sparse Fine-Tuning Framework for Effective Downstream Adaptation

Sungmin Kang, Jisoo Kim, Salman Avestimehr, Sunwoo Lee

AAAI, 2026

arXiv
DOI

This study proposes a distribution-sensitive and entorpy-aware parameter budget allocation method for parameter-efficient fine-tuning (PEFT).

BibTeX

@inproceedings{kang2026gem,
  title={GEM: A Scale-Aware and Distribution-Sensitive Sparse Fine-Tuning Framework for Effective Downstream Adaptation},
  author={Kang, Sungmin and Kim, Jisoo and Avestimehr, Salman and Lee, Sunwoo},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={40},
  number={27},
  pages={22509--22517},
  year={2026}
}