Title: Message Tuning Outshines Graph Prompt Tuning: A Prismatic Space Perspective

Abstract:

Graph Foundation Models (GFMs), grounded in the paradigms of pre-training and adaptation, have recently gained significant traction within the field of graph learning. For GFMs based on Graph Neural Networks (GNNs), graph prompt tuning has established itself as the dominant strategy for adapting to downstream tasks. While existing literature offers explanations for the efficacy of graph prompt tuning, a rigorous method for quantifying its adaptation capacity remains an unresolved challenge. Resolving this issue is essential for delineating the performance boundaries of graph prompt tuning and for facilitating the creation of more potent adaptation techniques.

To address this gap, we introduce Prismatic Space Theory (PS-Theory), a novel mathematical framework designed to quantify the adaptation capacity of various methods, with a specific focus on defining the upper limit for graph prompt tuning. Leveraging PS-Theory, we propose Message Tuning for GFMs (MTG), a lightweight adaptation strategy. MTG injects a minimal set of learnable message prototypes into each layer of the GNN backbone, enabling adaptive guidance of message fusion without necessitating updates to the pre-trained weights. Our theoretical analysis via PS-Theory demonstrates that the adaptation capacity of MTG surpasses the theoretical upper bound of graph prompt tuning. Comprehensive experiments confirm that MTG consistently outperforms graph prompt baselines across a wide range of benchmark datasets, offering robust empirical validation for our theoretical conclusions.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC