TS-PEFT: Unveiling Token-Level Redundancy in Parameter-Efficient Fine-Tuning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ma, Dabiao, Dai, Ziming, Xin, Zhimin, Wang, Shu, Yang, Jian, Fei, Haojun
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908795249098752
author Ma, Dabiao
Dai, Ziming
Xin, Zhimin
Wang, Shu
Yang, Jian
Fei, Haojun
author_facet Ma, Dabiao
Dai, Ziming
Xin, Zhimin
Wang, Shu
Yang, Jian
Fei, Haojun
contents Current Parameter-Efficient Fine-Tuning (PEFT) methods typically operate under an implicit assumption: Once a target module is selected, every token passing through it contributes equally to the downstream task and requires a parameter update. In this paper, we challenge this convention by revealing a pervasive token-level redundancy in the fine-tuning of large models (LMs). We propose TS-PEFT, a theoretical framework utilizing proximal optimization that acts as a dynamic probe to identify token-level redundancy during the fine-tuning process. Extensive experiments demonstrate that indiscriminately updating all tokens is not only computationally superfluous but often introduces optimization noise. Surprisingly, by discarding 30%-70% of token updates, TS-PEFT consistently matches or exceeds the performance of dense baselines such as LoRA, DoRA. Our in-depth analysis shows that the learned token-level sparsity is a superior indicator of module importance compared to traditional weight criteria, providing a novel data-driven perspective on the intrinsic adaptation mechanism of LMs.
format Preprint
id arxiv_https___arxiv_org_abs_2511_16147
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle TS-PEFT: Unveiling Token-Level Redundancy in Parameter-Efficient Fine-Tuning
Ma, Dabiao
Dai, Ziming
Xin, Zhimin
Wang, Shu
Yang, Jian
Fei, Haojun
Computation and Language
Artificial Intelligence
Current Parameter-Efficient Fine-Tuning (PEFT) methods typically operate under an implicit assumption: Once a target module is selected, every token passing through it contributes equally to the downstream task and requires a parameter update. In this paper, we challenge this convention by revealing a pervasive token-level redundancy in the fine-tuning of large models (LMs). We propose TS-PEFT, a theoretical framework utilizing proximal optimization that acts as a dynamic probe to identify token-level redundancy during the fine-tuning process. Extensive experiments demonstrate that indiscriminately updating all tokens is not only computationally superfluous but often introduces optimization noise. Surprisingly, by discarding 30%-70% of token updates, TS-PEFT consistently matches or exceeds the performance of dense baselines such as LoRA, DoRA. Our in-depth analysis shows that the learned token-level sparsity is a superior indicator of module importance compared to traditional weight criteria, providing a novel data-driven perspective on the intrinsic adaptation mechanism of LMs.
title TS-PEFT: Unveiling Token-Level Redundancy in Parameter-Efficient Fine-Tuning
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2511.16147