arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2406.11801cs.CL
收藏

安全算术:一个通过引导参数和激活实现语言模型测试时安全对齐的框架

Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations

Singapore University of Technology and Design(新加坡科技设计大学);Indian Institute of Technology Kharagpur(印度理工学院卡拉格普尔分校)

Rima Hazra, Sayan Layek, Somnath Banerjee, Soujanya Poria

AI总结:

针对大语言模型在动态意图与复杂目标下易生成有害内容的问题,提出无需训练的 Safety Arithmetic 框架,通过移除有害方向和安全对齐提升多场景安全性,并构建 NoIntentEdit 数据集。

摘要:

随着大型语言模型(LLM)成为翻译和问答等应用中不可或缺的组成部分,确保其与人类价值观安全对齐至关重要。当前对齐方法难以应对动态用户意图和复杂目标,使模型容易生成有害内容。我们提出 Safety Arithmetic,这是一种无需训练的框架,可在不同场景下增强 LLM 安全性:基础模型、监督微调模型(SFT)和编辑模型。Safety Arithmetic 包括用于避免有害内容的 Harm Direction Removal(有害方向移除),以及用于促进安全响应的 Safety Alignment(安全对齐)。此外,我们提出 NoIntentEdit,这是一个突出显示若被无意使用可能损害模型安全性的编辑实例的数据集。实验表明,Safety Arithmetic 显著提升安全指标、降低过度安全并保持模型效用,在确保安全内容生成方面优于现有方法。

英文摘要:

Ensuring the safe alignment of large language models (LLMs) with human values is critical as they become integral to applications like translation and question answering. Current alignment methods struggle with dynamic user intentions and complex objectives, making models vulnerable to generating harmful content. We propose Safety Arithmetic, a training-free framework enhancing LLM safety across different scenarios: Base models, Supervised fine-tuned models (SFT), and Edited models. Safety Arithmetic involves Harm Direction Removal to avoid harmful content and Safety Alignment to promote safe responses. Additionally, we present NoIntentEdit, a dataset highlighting edit instances that could compromise model safety if used unintentionally. Our experiments show that Safety Arithmetic significantly improves safety measures, reduces over-safety, and maintains model utility, outperforming existing methods in ensuring safe content generation.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑