CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment
CLEAR conditionally activates a safety adapter instead of applying safety tuning to every prompt, reducing harmful completions while limiting benign-task degradation.
**CLEAR** uses a hidden-state gate to vary the strength of a safety LoRA while keeping the backbone frozen. On Llama-3-8B-Instruct, HarmBench attack success fell from **32.3% to 0.5%**, with up to **7.1 points** more GSM8K accuracy than global SFT or LoRA.
For builders serving specialized models, conditional adapter routing is worth evaluating when global safety tuning damages benign tasks. Test the gate, adapter, and untouched backbone as separate failure surfaces across both adversarial and ordinary workloads.
**CLEAR** uses a hidden-state gate to vary the strength of a safety LoRA while keeping the backbone frozen. On Llama-3-8B-Instruct, HarmBench attack success fell from **32.3% to 0.5%**, with up to **7.1 points** more GSM8K accuracy than global SFT or LoRA. For builders serving specialized models, conditional adapter routing is worth evaluating when global safety tuning damages benign tasks. Test the gate, adapter, and untouched backbone as separate failure surfaces across both adversarial and ordinary workloads. The results come from selected safety and utility benchmarks on **Llama-3-8B-Instruct**. They do not establish production robustness, routing reliability under distribution shift, or behavior on other model families.
CLEAR narrows safety alignment from a global model rewrite to a conditional internal intervention: benign utility can be preserved by routing a frozen backbone through a safety adapter only as strongly as needed. This complements frozen-model reliability mechanisms but adds a new routing surface whose behavior must be evaluated as part of the deployment configuration; the single-model benchmark does not resolve production robustness or false refusals.