文献类型：期刊文献

英文题名：MV-MoE: A multi-view mixture-of-expert tuning method for medical LLMs

作者：Zhang, Weiyan[1];Yan, Yongyu[1];Wang, Jiacheng[1];Wu, Xueyan[1];Yuan, Cheng[1];Ruan, Tong[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China

年份：2027

卷号：332

外文期刊名：EXPERT SYSTEMS WITH APPLICATIONS

收录：;WOS:【SCI-EXPANDED(收录号:WOS:001823611200001)】；

基金：This work is supported by the Shanghai Natural Science Foundation Project under Grant 25ZR1402116.

语种：英文

外文关键词：Mixture-of-experts; Multi-view learning; Tuning method; Large language models

摘要：Recently, the performance of large language models (LLMs) has surged, thanks to larger parameters, massive data, and greater computational power. However, when LLMs are applied to specialized domains like healthcare, full-parameter fine-tuning faces several challenges: 1) the high cost of acquiring domain-specific annotated data, and 2) the specialized terminology, complex semantics, and multi-task requirements of medical texts, which demand enhanced fine-grained adaptability. Parameter-efficient fine-tuning (PEFT) offers a viable option due to its lower computational cost, yet existing methods still struggle with insufficient knowledge capture across multiple tasks. To address this issue, we propose a medical mixture-of-expert tuning method based on multi-view learning. Specifically, our method first enhances tuning through leveraging multi-view representations by introducing multiple expert modules, and then designs a multi-view mixed router for tuning. We conduct extensive experiments, and the results show that LLMs fine-tuned with our method can better learn knowledge from various tasks, leading to more effective task performance. Our codes are publicly available at https://github. com/yanyongyu/MV-MoE.