Graph-O-Planner: Injecting Graph Neural Tool Embeddings into LLMs for Efficient and Accurate Task Execution

ACL ARR 2025 May Submission7652 Authors

Abstract: Recent advancements in Large Language Models (LLMs) have enabled the development of AI agents capable of multi-step reasoning. However, deploying these agents in real-world applications requires planners that adapt to domain-specific tools and workflows, where traditional prompting frameworks often struggle to accurately represent available functional dependencies. To address this gap, we propose Graph-O-Planner, a novel graph-learning method that explicitly encodes tool relationships and execution sequences into LLM planning. Our approach constructs graph embeddings of available tools, enabling agents to dynamically map dependencies while minimizing context window overload. Evaluations across multiple benchmarks, including UltraTool and Task Bench, demonstrate that Graph-O-Planner achieves up to 68% higher and 60% higher performance with our approach, compared to state-of-the-art hybrid graph+LLM based planners and LLM-finetuned planners respectively, while significantly reducing any hallucinations in LLM generation. The method’s tool knowledge compression further reduces inference latency by 20%, validating its effectiveness in resource-constrained environments and making it more compatible for real-life practical deployment. We release our code https://anonymous.4open.science/r/Graph-O-Planner-16B3.

Paper Type: Long

Research Area: Information Retrieval and Text Mining

Research Area Keywords: Graph neural network, AI agents, LLM, Planning, Graphs

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 7652