Model Merging and Safety Alignment: One Bad Model Spoils the Bunch

Abstract

Merging Large Language Models (LLMs) is a cost-effective technique for combining multiple expert LLMs into a single versatile model, retaining the expertise of the original ones. However, current approaches often overlook the importance of safety alignment during merging, leading to highly misaligned models. This work investigates the effects of model merging on alignment. We evaluate several popular model merging techniques, demonstrating that existing methods do not only transfer domain expertise but also propagate misalignment. We propose a simple two-step approach to address this problem (i) generating synthetic safety and domain-specific data, and (ii) incorporating these generated data into the optimization process of existing data-aware model merging techniques. This allows us to treat alignment as a skill that can be maximized in the resulting merged LLM. Our experiments illustrate the effectiveness of integrating alignment-related data during merging, resulting in models that excel in both domain expertise and alignment.

Hasan Abed Al Kader Hammoud

PhD Student

Hasan is an Electrical and Computer Engineering Ph.D. student in Image and Video Understanding Lab (IVUL) Group in the Artificial Intelligence Initiative (AII) at King Abdullah University of Science and Technology (KAUST) under the supervision of Professor Bernard Ghanem.