文章背景与核心概要
在大语言模型（LLM）的安全对齐过程中，提升模型安全性往往会引入一个不受欢迎的副作用：模型效用退化。传统的全局安全微调方法（如监督微调 SFT 或标准 LoRA）会进行统一的参数修改，从而对有害输入和良性输入的响应都产生影响。
为了打破这种安全性与效用之间的权衡困境，本文作者提出了 CLEAR（Continuous LatEnt Adapter Routing，连续潜在适配器路由）这一条件安全适配框架。CLEAR 利用一个轻量级的隐状态门控机制，动态且连续地控制安全低秩适配器的激活强度。实验表明，CLEAR 能够在保持冻结的主干模型不变的前提下实现精准安全干预，在 Llama-3-8B-Instruct 上将 HarmBench 攻击成功率（ASR）从 32.3% 大幅降至 0.5%，同时在 GSM8K 等基准测试中比传统全局 SFT 或标准 LoRA 保持了高出达 7.1 个百分点的基准效用准确率。
CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment
arXiv: arXiv:2608.21278 [cs.AI]
Submitted: August 21, 2026
Authors: Chengxiao Wang, Enyi Jiang, Xiaojing Liao, Sanmi Koyejo
CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment
arXiv: arXiv:2608.21278 [cs.AI]
Submitted: August 21, 2026
Authors: Chengxiao Wang, Enyi Jiang, Xiaojing Liao, Sanmi Koyejo
📌 Summary
Improving the safety of Large Language Models (LLMs) often introduces an unwanted side effect: utility degradation. Traditional global safety tuning methods (such as Supervised Fine-Tuning or standard LoRA) apply modifications uniformly, impacting responses to both harmful and benign inputs alike.
To solve this trade-off, the authors propose CLEAR (Continuous LatEnt Adapter Routing), a conditional safety adaptation framework. CLEAR utilizes a lightweight hidden-state gate to dynamically and continuously control the activation strength of a safety low-rank adapter.
Key Highlights:
- Targeted Safety Intervention: Reduces harmful completions while leaving the frozen model backbone intact for benign prompts.
- Superior Performance on Benchmarks: Tested on Llama-3-8B-Instruct, CLEAR successfully reduces the HarmBench Attack Success Rate (ASR) from 32.3% down to 0.5%.
- Preserved Model Utility: Retains baseline utility capabilities, achieving up to 7.1 percentage points higher GSM8K accuracy compared to traditional global SFT or standard LoRA alignment.
📌 Summary
Improving the safety of Large Language Models (LLMs) often introduces an unwanted side effect: utility degradation. Traditional global safety tuning methods (such as Supervised Fine-Tuning or standard LoRA) apply modifications uniformly, impacting responses to both harmful and benign inputs alike.
To solve this trade-off, the authors propose CLEAR (Continuous LatEnt Adapter Routing), a conditional safety adaptation framework. CLEAR utilizes a lightweight hidden-state gate to dynamically and continuously control the activation strength of a safety low-rank adapter.
Key Highlights:
- Targeted Safety Intervention: Reduces harmful completions while leaving the frozen model backbone intact for benign prompts.
- Superior Performance on Benchmarks: Tested on Llama-3-8B-Instruct, CLEAR successfully reduces the HarmBench Attack Success Rate (ASR) from 32.3% down to 0.5%.
- Preserved Model Utility: Retains baseline utility capabilities, achieving up to 7.1 percentage points higher GSM8K accuracy compared to traditional global SFT or standard LoRA alignment.
🔗 Quick Links
🔗 Quick Links
🗂️ Metadata
- Primary Subject: Artificial Intelligence (
cs.AI) - Cite As:
arXiv:2608.21278 [cs.AI]
🗂️ Metadata
- Primary Subject: Artificial Intelligence (
cs.AI)- Cite As:
arXiv:2608.21278 [cs.AI]