Knowledge graph foundation models aim to predict missing links on graphs they have not been trained on. The earlier ULTRA model pioneered this by building a relation graph from co-occurrence patterns that match motif templates, then passing messages over that graph. A new arXiv preprint introduces MERIT, which keeps that relation-graph construction but changes how information is propagated.
Instead of standard message passing, MERIT uses relation transformers to process the relation graph. The authors argue this design scales better while preserving the zero-shot capability: the model can be applied to an unseen graph and still make link predictions without any task-specific fine-tuning.
The paper is a single source, so there are no conflicting findings to compare. It positions MERIT as a step toward larger and more capable knowledge graph foundation models, building directly on the ULTRA framework.