GTool: Graph Enhanced Tool Planning with Large Language Model

2025-08-18

AI Summary

Existing LLM-based tool planning methods treat tools as isolated units, ignoring their intrinsic dependencies—particularly problematic when dependency information is incomplete, leading to suboptimal tool selection. Method: We propose a graph-enhanced tool planning framework that (i) models tool dependencies via an explicit graph structure; (ii) introduces *graph tokens* to encode dependency relations; (iii) formulates a *missing dependency prediction* task to mitigate dependency sparsity; and (iv) employs a graph neural network to construct request-specific tool graphs, integrating graph-aware contextual signals into LLMs via prompt learning—enabling plug-and-play deployment without architectural modification. Contribution/Results: Evaluated on a 7B-parameter LLM, our method achieves over 29.6% absolute improvement over state-of-the-art baselines. Crucially, it generalizes seamlessly across diverse large language model architectures without fine-tuning.

Technology Category

• Planning, Routing, and Scheduling: Planning with Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Human-Aware Planning and Behavior Prediction

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

Tool planning with large language models (LLMs), referring to selecting, organizing, and preparing the tools necessary to complete a user request, bridges the gap between natural language understanding and task execution. However, current works treat different tools as isolated components and fail to leverage the inherent dependencies of tools, leading to invalid planning results. Since tool dependencies are often incomplete, it becomes challenging for LLMs to accurately identify the appropriate tools required by a user request, especially when confronted with a large toolset. To solve this challenge, we propose exttt{GTool}, which is the first work aiming to enhance the tool planning ability of LLMs under incomplete dependencies. exttt{GTool} constructs a request-specific tool graph to select tools efficiently and generate the exttt{<graph token>} which provides sufficient dependency information understandable by LLMs. Moreover, a missing dependency prediction task is designed to improve the reliability of exttt{GTool} with incomplete dependencies. Without trimming LLMs, exttt{GTool} can be seamlessly integrated with various LLM backbones without extensive retraining. Extensive experiments show that exttt{GTool} achieves more than 29.6% performance improvements compared with the state-of-the-art (SOTA) baselines with a light-weight (7B) LLM backbone.

Problem

• Enhancing tool planning in LLMs with incomplete dependencies
• Constructing request-specific tool graphs for efficient selection
• Improving reliability via missing dependency prediction task

Innovation

• Constructs request-specific tool graph
• Generates graph token for LLMs
• Predicts missing dependencies for reliability