ARIADNE: Agnostic Routing for Inference-time Adapter DyNamic sElection
Researchers have developed ARIADNE, a novel framework for dynamically selecting the most appropriate adapter for inference-time queries without requiring task labels. This training-free and adapter-agnostic method represents each adapter using centroids derived from its training data embeddings, enabling selection based on proximity in latent space. ARIADNE is compatible with various parameter-efficient fine-tuning (PEFT) methods and does not need modifications to existing adapters or training procedures. Evaluations showed ARIADNE achieved 97.44% of upper-bound performance on 23 NLP tasks and 89.7% average selection accuracy on 44 tasks when tested with Llama 3.2 1B Instruct. AI
IMPACT This method could streamline the deployment and management of specialized AI models by automating adapter selection, improving efficiency and portability.