AdaMMS: Model Merging for Heterogeneous Multimodal Large Language Models with Unsupervised Coefficient Optimization
Yiyang Du, Xiaochen Wang, Chi Chen, Jiabo Ye, Yiru Wang, Peng Li, Ming Yan, Ji Zhang, Fei Huang, Zhifang Sui, Maosong Sun, Yang Liu
2025-04-02
Summary
This paper is about a new way to combine different AI models that understand both images and text, even if they're built differently.
What's the problem?
It's hard to combine AI models that handle images and text because they often have different designs and ways of processing information.
What's the solution?
The researchers created AdaMMS, a method that carefully maps the models to each other, merges them in a smart way, and then finds the best settings to make them work together effectively.
Why it matters?
This work matters because it can lead to more powerful AI systems that combine the strengths of different models, even if they're not originally designed to work together.
Abstract
Recently, model merging methods have demonstrated powerful strengths in combining abilities on various tasks from multiple Large Language Models (LLMs). While previous model merging methods mainly focus on merging homogeneous models with identical architecture, they meet challenges when dealing with Multimodal Large Language Models (MLLMs) with inherent heterogeneous property, including differences in model architecture and the asymmetry in the parameter space. In this work, we propose AdaMMS, a novel model merging method tailored for heterogeneous MLLMs. Our method tackles the challenges in three steps: mapping, merging and searching. Specifically, we first design mapping function between models to apply model merging on MLLMs with different architecture. Then we apply linear interpolation on model weights to actively adapt the asymmetry in the heterogeneous MLLMs. Finally in the hyper-parameter searching step, we propose an unsupervised hyper-parameter selection method for model merging. As the first model merging method capable of merging heterogeneous MLLMs without labeled data, extensive experiments on various model combinations demonstrated that AdaMMS outperforms previous model merging methods on various vision-language benchmarks.