LED-Merging: Mitigating Safety-Utility Conflicts in Model Merging with Location-Election-Disjoint
It addresses a critical issue for developers of multi-task LLMs by providing a lightweight, training-free method to enhance safety without compromising utility, though it is incremental as it builds on existing model merging techniques.
The paper tackles the problem of safety-utility conflicts in model merging for large language models, proposing LED-Merging, which reduces harmful response rates by 31.4% on Llama-3-8B-Instruct while preserving 95% of utility performance, such as 52.39% accuracy on GSM8K.
Fine-tuning pre-trained Large Language Models (LLMs) for specialized tasks incurs substantial computational and data costs. While model merging offers a training-free solution to integrate multiple task-specific models, existing methods suffer from safety-utility conflicts where enhanced general capabilities degrade safety safeguards. We identify two root causes: $\textbf{neuron misidentification}$ due to simplistic parameter magnitude-based selection, and $\textbf{cross-task neuron interference}$ during merging. To address these challenges, we propose $\textbf{LED-Merging}$, a three-stage framework that $\textbf{L}$ocates task-specific neurons via gradient-based attribution, dynamically $\textbf{E}$lects critical neurons through multi-model importance fusion, and $\textbf{D}$isjoints conflicting updates through parameter isolation. Extensive experiments on Llama-3-8B, Mistral-7B, and Llama2-13B demonstrate that LED-Merging effectively reduces harmful response rates, showing a 31.4\% decrease on Llama-3-8B-Instruct on HarmBench, while simultaneously preserving 95\% of utility performance, such as achieving 52.39\% accuracy on GSM8K. LED-Merging resolves safety-utility conflicts and provides a lightweight, training-free paradigm for constructing reliable multi-task LLMs. Code is available at $\href{https://github.com/MqLeet/LED-Merging}{GitHub}$.