The paper explores the development and reuse of a library of adapters, specifically LoRA adapters, to enhance the performance of large language models (LLMs) on new tasks. The authors propose Model-Based Clustering (MBC) to build a library of adapters by clustering tasks based on the similarity of their LoRA parameters, which indirectly optimizes transfer across multi-task datasets. They introduce Arrow, a zero-shot routing mechanism that dynamically selects the most relevant adapters for new inputs without retraining. The experiments with models like Phi-2 and Mistral demonstrate that MBC-based adapters and Arrow routing lead to superior generalization to new tasks, showing promise for creating modular and adaptable LLMs. The contributions include a novel method for building and reusing adapter libraries, and a zero-shot routing mechanism that facilitates efficient and effective task generalization.The paper explores the development and reuse of a library of adapters, specifically LoRA adapters, to enhance the performance of large language models (LLMs) on new tasks. The authors propose Model-Based Clustering (MBC) to build a library of adapters by clustering tasks based on the similarity of their LoRA parameters, which indirectly optimizes transfer across multi-task datasets. They introduce Arrow, a zero-shot routing mechanism that dynamically selects the most relevant adapters for new inputs without retraining. The experiments with models like Phi-2 and Mistral demonstrate that MBC-based adapters and Arrow routing lead to superior generalization to new tasks, showing promise for creating modular and adaptable LLMs. The contributions include a novel method for building and reusing adapter libraries, and a zero-shot routing mechanism that facilitates efficient and effective task generalization.