Graph foundation models embed each node as a single vector, which entangles semantics for nodes that legitimately carry multiple labels and cripples cross-domain transfer; this work replaces the single point with an adaptive composition of learned "semantic bases" plus a dual-channel semantic/structure architecture with domain-adversarial training. If you're building GNN-based classifiers or graph retrieval where entities are inherently multi-topic, this is the design pattern to watch: multi-vector/basis representations over single embeddings to avoid the semantic collapse that hurts multi-label accuracy, though it's an early research result without production benchmarks yet.