RTA: Replacing Message Passing with Label-Aware Retrieval

RTA reframes Graph Neural Networks as retrieval-augmented models, replacing expensive structural message passing with an MLP-based framework that uses permutation-invariant summaries of retrieved graph context. The method maintains performance while increasing robustness to imperfect graph structures and mis-retrieved outliers.