NaviAgent: Bilevel Planning for Scalable Tool Orchestration in LLMs
NaviAgent introduces an innovative architecture for Large Language Models (LLMs) that distinguishes between task planning and tool execution through a graph-based representation of tool relationships. This design overcomes the challenge faced by LLMs that typically engage tools individually without an overarching perspective, resulting in cumulative errors and limited scalability when dealing with hundreds or thousands of tools. At the planning stage, decisions are made to either respond directly, seek clarification, or assemble a toolchain. Meanwhile, the execution phase employs a Tool World Navigation Model (TWNM) to capture the structural and behavioral connections among tools, facilitating scalable and effective invocation sequences. This methodology is elaborated in a paper available on arXiv (2506.19500).
Key facts
- NaviAgent is a bilevel architecture for LLMs that decouples task planning from tool execution.
- It uses graph-based modeling of tool relations.
- The planning level decides to respond, clarify, or retrieve a toolchain.
- The execution level uses a Tool World Navigation Model (TWNM).
- TWNM encodes structural and behavioral relations among tools.
- It aims to reduce error accumulation and improve scalability.
- The paper is available on arXiv with ID 2506.19500.
- The approach targets scaling to hundreds or thousands of tools.
Entities
Institutions
- arXiv