Multi-Label Graph Foundation Models: From Single-Vector to Multi-Semantic Basis Learning

A new arXiv paper (2608.06394) proposes a framework for multi-label node classification in graph learning, addressing the limitation of existing Graph Foundation Models (GFMs) that assume single-label nodes. The authors argue that current GFMs embed each node into a single vector, which approximates multiple semantics with a single point, causing semantic entanglement. They introduce a multi-semantic basis learning approach to capture multiple labels more effectively. The paper highlights the challenge of cross-domain generalization, as existing methods are trained and tested within the same graph domain. The proposed method aims to improve transferability across diverse graph domains and downstream tasks. The paper is announced as a new submission on arXiv, with the abstract outlining the problem and proposed solution.