Researchers developed CLEAR, a technique that uses a lightweight hidden-state gate to continuously control the activation strength of a safety low-rank adapter, rather than applying safety measures uniformly across a model. This conditional approach selectively activates safety protections based on input content instead of blanket-applying them. On Llama-3-8B-Instruct, CLEAR reduced harmful response rates from 32.3% to 0.5% while preserving utility scores and outperforming standard safety techniques by up to 7.1 percentage points on math benchmarks.
