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

Hua Ye (nanjing university) · Siyuan Chen (University of Bristol) · Haoliang Zhang (The University of Oklahoma) · Weihao Luo (Donghua University, Shanghai) · Yanbin Li (The University of Tokyo) · Xuan Zhang (Texas A&M University)

domain discrepancydomain synergyempirical evaluations generalization abilities generalization boundsheterogeneous domainsinter-domain interferencelanguage understanding tasksmodel capacity constraintsmulti-stage fine-tuningnegative transferpartitioning strategystate-of-the-art baselines theoretical analysis

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.