ARIADNE Enables Training-Free Dynamic Adapter Selection for LLMs

Researchers have developed ARIADNE, an innovative framework designed to dynamically select the most suitable task-specific adapter for large language models (LLMs) during inference, without requiring any additional training. In environments where a single LLM backbone is paired with numerous specialized adapters, ARIADNE addresses the challenge of routing unlabeled input queries to the correct adapter. ARIADNE operates by representing each adapter through a set of centroids derived from the embeddings of its training data, effectively capturing the adapter's associated data distribution. When an unlabeled input arrives, the system measures its proximity to these centroids in the latent space to make a selection. This approach is compatible with any parameter-efficient fine-tuning (PEFT) method and requires no modifications to the adapters or their training procedures, demonstrating high performance across diverse NLP tasks.