The paper introduces a novel multi-LLM agent framework, α-UMi, designed to enhance the tool learning capabilities of LLMs. Traditional approaches often rely on a single large LLM to perform all tasks, which can be limiting, especially with smaller models. α-UMi decomposes the tool learning task into three distinct sub-tasks: planning, calling, and summarizing, each implemented by a specialized LLM. This modular approach allows for individual updates and the potential use of smaller LLMs, improving adaptability and performance. To effectively train this multi-LLM framework, the authors propose a two-stage training paradigm: first, fine-tuning a backbone LLM on the entire dataset without discriminating sub-tasks; second, fine-tuning the planner, caller, and summarizer on their respective sub-tasks. This method ensures that each component is optimized for its specific role, enhancing overall performance. The evaluation across various tool-use benchmarks demonstrates that α-UMi outperforms traditional single-LLM approaches, highlighting its efficacy and advantages in tool learning. The framework's modular structure and fine-tuning strategy not only improve performance but also reduce the need for large model capacities, making it more scalable and cost-effective. The paper also discusses the limitations of the approach, such as the potential for further integration with powerful closed-source LLMs and the need for more extensive testing on diverse agent tasks. Ethical considerations are addressed, ensuring that the framework is trained on public datasets and does not pose ethical risks.The paper introduces a novel multi-LLM agent framework, α-UMi, designed to enhance the tool learning capabilities of LLMs. Traditional approaches often rely on a single large LLM to perform all tasks, which can be limiting, especially with smaller models. α-UMi decomposes the tool learning task into three distinct sub-tasks: planning, calling, and summarizing, each implemented by a specialized LLM. This modular approach allows for individual updates and the potential use of smaller LLMs, improving adaptability and performance. To effectively train this multi-LLM framework, the authors propose a two-stage training paradigm: first, fine-tuning a backbone LLM on the entire dataset without discriminating sub-tasks; second, fine-tuning the planner, caller, and summarizer on their respective sub-tasks. This method ensures that each component is optimized for its specific role, enhancing overall performance. The evaluation across various tool-use benchmarks demonstrates that α-UMi outperforms traditional single-LLM approaches, highlighting its efficacy and advantages in tool learning. The framework's modular structure and fine-tuning strategy not only improve performance but also reduce the need for large model capacities, making it more scalable and cost-effective. The paper also discusses the limitations of the approach, such as the potential for further integration with powerful closed-source LLMs and the need for more extensive testing on diverse agent tasks. Ethical considerations are addressed, ensuring that the framework is trained on public datasets and does not pose ethical risks.