Model Merging and Safety Alignment: One Bad Model Spoils the Bunch
This addresses a critical safety issue for users of merged LLMs, though it is incremental as it builds on existing data-aware merging techniques.
The paper tackles the problem of safety misalignment in merged Large Language Models (LLMs), showing that existing merging techniques propagate misalignment, and proposes a two-step approach using synthetic safety data to improve alignment while retaining domain expertise.
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.