LED-Merging: Mitigating Safety-Utility Conflicts in Model Merging with Location-Election-Disjoint
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866916897251917824 |
|---|---|
| author | Ma, Qianli Liu, Dongrui Chen, Qian Zhang, Linfeng Shao, Jing |
| author_facet | Ma, Qianli Liu, Dongrui Chen, Qian Zhang, Linfeng Shao, Jing |
| contents | 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 $\href{https://github.com/MqLeet/LED-Merging}{GitHub}$. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2502_16770 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | LED-Merging: Mitigating Safety-Utility Conflicts in Model Merging with Location-Election-Disjoint Ma, Qianli Liu, Dongrui Chen, Qian Zhang, Linfeng Shao, Jing Computation and Language Artificial Intelligence 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 $\href{https://github.com/MqLeet/LED-Merging}{GitHub}$. |
| title | LED-Merging: Mitigating Safety-Utility Conflicts in Model Merging with Location-Election-Disjoint |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2502.16770 |