Tag
CLEAR introduces a continuous latent adapter routing framework for LLM safety alignment, using a hidden-state gate to modulate safety adapters and improve robustness on HarmBench while preserving utility on benign inputs.
Proposes ARIADNE, a training-free, adapter-agnostic routing framework that selects the optimal PEFT adapter at inference time by measuring input proximity to adapter-specific centroids in embedding space, recovering 97.44% of upper-bound performance on 23 tasks.