BriefGPT.xyz
Jul, 2024
PrimeGuard:通过无调优路由实现安全且有用的语言模型
PrimeGuard: Safe and Helpful LLMs through Tuning-Free Routing
HTML
PDF
Blazej Manczak, Eliott Zemour, Eric Lin, Vaikkunth Mugunthan
TL;DR
本研究解决了当前语言模型在安全性与有用性之间的平衡问题,提出了一种新颖的无调优路由方法PrimeGuard。该方法通过结构化控制流有效提升了模型的安全性和有用性,使得安全响应比例从61%提高到97%,并且在对抗性攻击中表现出更强的抵抗能力。
Abstract
Deploying
language models
(LMs) necessitates outputs to be both high-quality and compliant with
safety
guidelines. Although Inference-Time Guardrails (ITG) offer solutions that shift model output distributions to
→