Synergy over Discrepancy: A Partition-Based Approach to Multi-Domain LLM Fine-Tuning
This addresses the problem of inter-domain interference in multi-domain LLM fine-tuning for AI researchers, representing an incremental improvement with a novel partitioning strategy.
The paper tackles the challenge of adapting large language models across multiple heterogeneous domains by proposing a partition-based multi-stage fine-tuning framework that exploits inter-domain synergies while minimizing negative transfer, and it shows consistent outperformance over state-of-the-art baselines in empirical evaluations.
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.