The research introduces the concept of a tool menu, a constrained subset of available tools presented to an agent before execution. This design addresses the challenge of agents operating within large tool libraries. The framework utilizes a state path, a pre-defined sequence of actions, to guide the agent's execution. The State-Path Tool Menu learns this path, treating the menu as an execution prior. An encoder within the system represents tool capabilities and dependencies, while a retriever identifies missing inputs and the final action. A reranker then orders producers before consumers. Experiments on ToolBench show a significant improvement in online success, raising the rate from 0.737 to 0.898. This performance gain is observed across different model capacities and executor families. The State-Path menu also handles more complete chains, covering 32 tools compared to the 128 in the original list.
The key innovation lies in reducing the complexity of the agent's decision space. By restricting tool selection, the system avoids the overhead of evaluating thousands of potential tools. This approach focuses the agent's attention on the most relevant actions, streamlining the execution process. The framework's design is modular, allowing for adaptation to different agent architectures and toolsets.
This research provides a practical solution for improving the efficiency and reliability of online agents. The State-Path Tool Menu offers a robust method for guiding agent behavior, particularly in multi-step tasks. The demonstrated performance improvement highlights the potential of this approach for real-world applications. The code is available at https://github.com/Met2348/State-Path.
Source: https://arxiv.org/abs/2609.09395