ToolNet is a framework that enables large language models (LLMs) to interact with a vast number of tools through a directed graph structure. Unlike traditional in-context learning methods that list tools as plain text, ToolNet organizes tools into a graph where each node represents a tool and edges denote transitions between tools. This allows LLMs to navigate the graph to select appropriate tools for solving tasks. The framework dynamically updates the graph based on tool performance, improving efficiency and resilience to tool failures. ToolNet outperforms existing methods in multi-hop tool learning tasks, achieving higher accuracy with fewer tokens. It also effectively handles noisy or irrelevant tools by adjusting transition weights. The framework is evaluated on various datasets, demonstrating its effectiveness in real-world scenarios. ToolNet's adaptive graph structure and dynamic tool selection make it a promising approach for integrating LLMs with massive tools.ToolNet is a framework that enables large language models (LLMs) to interact with a vast number of tools through a directed graph structure. Unlike traditional in-context learning methods that list tools as plain text, ToolNet organizes tools into a graph where each node represents a tool and edges denote transitions between tools. This allows LLMs to navigate the graph to select appropriate tools for solving tasks. The framework dynamically updates the graph based on tool performance, improving efficiency and resilience to tool failures. ToolNet outperforms existing methods in multi-hop tool learning tasks, achieving higher accuracy with fewer tokens. It also effectively handles noisy or irrelevant tools by adjusting transition weights. The framework is evaluated on various datasets, demonstrating its effectiveness in real-world scenarios. ToolNet's adaptive graph structure and dynamic tool selection make it a promising approach for integrating LLMs with massive tools.