ReaGAN: Node-as-Agent-Reasoning Graph Agentic Network

TL;DR

ReaGAN reframes graph learning by treating each node as an autonomous agent capable of planning, memory-based reasoning, and action via a frozen Large Language Model (LLM). It integrates local structural aggregation with global semantic retrieval through Retrieval-Augmented Generation (RAG), enabling node-specific decision-making without gradient-based training. Key contributions include formalizing the agentic node workflow, a hybrid local-global aggregation mechanism, and comprehensive ablations showing the necessity of planning and global retrieval, as well as the detrimental effect of exposing label semantics. The approach demonstrates competitive performance on standard node-classification benchmarks under few-shot, non-fine-tuned settings, highlighting the potential of LLM-powered agents for flexible, plug-and-play graph learning with global semantic context.

Abstract

Graph Neural Networks (GNNs) have achieved remarkable success in graph-based learning by propagating information among neighbor nodes via predefined aggregation mechanisms. However, such fixed schemes often suffer from two key limitations. First, they cannot handle the imbalance in node informativeness -- some nodes are rich in information, while others remain sparse. Second, predefined message passing primarily leverages local structural similarity while ignoring global semantic relationships across the graph, limiting the model's ability to capture distant but relevant information. We propose Retrieval-augmented Graph Agentic Network (ReaGAN), an agent-based framework that empowers each node with autonomous, node-level decision-making. Each node acts as an agent that independently plans its next action based on its internal memory, enabling node-level planning and adaptive message propagation. Additionally, retrieval-augmented generation (RAG) allows nodes to access semantically relevant content and build global relationships in the graph. ReaGAN achieves competitive performance under few-shot in-context settings using a frozen LLM backbone without fine-tuning, showcasing the potential of agentic planning and local-global retrieval in graph learning.

Figures (3)

• Figure 1: Message passing in traditional pre-defined way vs. ReaGAN's method. (a) All nodes perform local aggregation only. (b) Nodes aggregate in different ways. A aggregates both locally and globally; B performs local-only aggregation; C performs global-only aggregation; D performs no operation.
• Figure 2: Overview of ReaGAN. Each node in ReaGAN is modeled as an agent equipped with four core modules: Memory, Planning, Tools, and Action. The node stores its original information and receives local and global information into its memory, constructs a natural language prompt, and queries a frozen LLM for an action plan. The planner outputs one action per layer (e.g., Local Aggregation, Global Aggregation), which are executed using available tools such as RAG. The resulting outputs are written back into memory and may produce a predicted label. This forms a closed loop of perception, planning, action, and memory refinement across layers.
• Figure 3: Information flow from memory to prompt. Each node’s memory includes its original text feature, aggregated text feature from local and global neighbors, and selected labeled neighbor shots. During planning or prediction, these components are selectively injected into a natural language prompt, providing the LLM with (i) the input node’s raw identity, (ii) context-enhanced descriptions, and (iii) label-text pairs for few-shot learning. This design enables each agent to reason over multi-scale context and take personalized actions.