Synergy over Discrepancy: A Partition-Based Approach to Multi-Domain LLM Fine-Tuning

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ye, Hua, Chen, Siyuan, Zhang, Haoliang, Luo, Weihao, Li, Yanbin, Zhang, Xuan
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914250015899648
author Ye, Hua
Chen, Siyuan
Zhang, Haoliang
Luo, Weihao
Li, Yanbin
Zhang, Xuan
author_facet Ye, Hua
Chen, Siyuan
Zhang, Haoliang
Luo, Weihao
Li, Yanbin
Zhang, Xuan
contents Large language models (LLMs) demonstrate impressive generalization abilities, yet adapting them effectively across multiple heterogeneous domains remains challenging due to inter-domain interference. To overcome this challenge, we propose a partition-based multi-stage fine-tuning framework designed to exploit inter-domain synergies while minimizing negative transfer. Our approach strategically partitions domains into subsets (stages) by balancing domain discrepancy, synergy, and model capacity constraints. We theoretically analyze the proposed framework and derive novel generalization bounds that justify our partitioning strategy. Extensive empirical evaluations on various language understanding tasks show that our method consistently outperforms state-of-the-art baselines.
format Preprint
id arxiv_https___arxiv_org_abs_2511_07198
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Synergy over Discrepancy: A Partition-Based Approach to Multi-Domain LLM Fine-Tuning
Ye, Hua
Chen, Siyuan
Zhang, Haoliang
Luo, Weihao
Li, Yanbin
Zhang, Xuan
Machine Learning
I.2.7; I.2.6
Large language models (LLMs) demonstrate impressive generalization abilities, yet adapting them effectively across multiple heterogeneous domains remains challenging due to inter-domain interference. To overcome this challenge, we propose a partition-based multi-stage fine-tuning framework designed to exploit inter-domain synergies while minimizing negative transfer. Our approach strategically partitions domains into subsets (stages) by balancing domain discrepancy, synergy, and model capacity constraints. We theoretically analyze the proposed framework and derive novel generalization bounds that justify our partitioning strategy. Extensive empirical evaluations on various language understanding tasks show that our method consistently outperforms state-of-the-art baselines.
title Synergy over Discrepancy: A Partition-Based Approach to Multi-Domain LLM Fine-Tuning
topic Machine Learning
I.2.7; I.2.6
url https://arxiv.org/abs/2511.07198