CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment
CLEAR turns safety tuning on selectively, aiming to block harmful prompts without dragging down benign performance.
The paper proposes a lightweight hidden-state gate that controls how strongly a safety LoRA adapter activates. In tests, the authors say CLEAR improved HarmBench robustness while preserving more utility than globally applied SFT or LoRA. On Llama-3-8B-Instruct, it cut HarmBench ASR from 32.3% to 0.5% and kept GSM8K accuracy up to 7.1 points higher than those global tuning baselines. HF Daily Papers' note
The paper proposes a lightweight hidden-state gate that controls how strongly a safety LoRA adapter activates. In tests, the authors say CLEAR improved HarmBench robustness while preserving more utility than globally applied SFT or LoRA. On Llama-3-8B-Instruct, it cut HarmBench ASR from 32.3% to 0.5% and kept GSM8K accuracy up to 7.1 points higher than those global tuning baselines. HF Daily Papers' note
score 5