GRATE: Temporal Extensions for Inductive KG Foundation Models via Gated Rotary Attention

GRATE extends inductive knowledge graph foundation models to temporal knowledge graphs without adding learnable parameters. Its entity-side message function rotates each edge message according to the relative gap between the edge time and query time, then uses a query-conditioned gate to retain temporally relevant signals. The method integrates with NBFNet-style models while preserving transfer across disjoint vocabularies. The authors also introduce GDELTIndT and WIKIIndT, covering interpolation and extrapolation with disjoint entities, relations, and timestamps, and report that one jointly pretrained checkpoint beats the static base model in most evaluated settings.