CLEAR cuts Llama-3-8B HarmBench attack rate to 0.5%
TL;DR
- CLEAR uses a hidden-state gate to conditionally activate a safety LoRA on Llama-3-8B-Instruct rather than tune the whole model.
- HarmBench attack success rate on Llama-3-8B-Instruct falls from 32.3% to 0.5% under CLEAR.
- On GSM8K, CLEAR scores 7.1 percentage points higher than globally applied supervised fine-tuning or standard LoRA.
On Llama-3-8B-Instruct, a new gating scheme called CLEAR drops the HarmBench attack success rate from 32.3% to 0.5% while beating both supervised fine-tuning and standard LoRA on GSM8K math accuracy by 7.1 percentage points, according to a preprint on arXiv from Chengxiao Wang, Enyi Jiang, Xiaojing Liao and Sanmi Koyejo.
Instead of baking safety into every weight, the authors use "a lightweight hidden-state gate to continuously control the activation strength of a safety low-rank adapter," so the safety LoRA only fires as much as the input warrants and the backbone stays frozen. That framing is meant to spare the model's math and reasoning capabilities, which are typically the first things to erode under global safety SFT.
Two of the researchers we track flagged the paper the day it went up. The headline numbers cover one model and two benchmarks; the abstract does not address coding, tool use, or adaptive attacks that target the gate itself.
Shared on Bluesky by 2 AI experts
Originally reported by arxiv.org
Read the original article →Original headline: CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment