AgentGFM: Graph Foundation Model with Adaptive Node-Agent Information Flow

Graph Foundation Models (GFMs) aim to learn generalizable knowledge from various graph structures, but their effectiveness is often limited by fixed information propagation schemes that fail to adapt to the diverse local patterns within and across graphs. This rigidity hinders their ability to transfer knowledge effectively to new, unseen scenarios. AgentGFM addresses this by conceptualizing each node within a graph as an intelligent agent capable of autonomously deciding how information should flow. These node agents operate under a shared, end-to-end trainable policy, engaging in a predict-act-observe-correct cycle. During the "act" stage, each node makes critical decisions regarding source reception, signal-channel selection, and gain-aware halting, allowing for highly adaptive information-flow control. This agent-based approach significantly enhances the model's transferability and performance across a wide range of node-level, graph-level, and large-scale transfer tasks, demonstrating its effectiveness across diverse graph topologies.