When the Tool Decides: LLM Agents Defer Blindly to Graph Neural Network Tools, and Stronger Backbones Defer More

The paper presents findings on the interaction between large language model (LLM) agents and graph neural networks (GNNs) when the LLMs are tasked with making decisions. It reveals that LLM agents, regardless of their backbone size (ranging from 0.5B to 7B parameters), tend to defer to the GNN's outputs without exercising independent judgment, achieving agreement rates of 97.6-99.2%. This behavior persists even as model capability increases, highlighting the need for intentional design of selective invocation mechanisms, as the default reliance on GNN tools limits the potential for improved decision-making in LLM applications.