analiza aiqbench
The paper proposes a parameter-free message function that lets temporal knowledge graph models transfer to graphs with unseen entities, relations, and timestamps, a known weak point of existing temporal models. The evidence is limited to a workshop paper with new benchmarks the authors built themselves, and gains over the static base model hold only "in most settings." Readers following graph foundation models should watch whether this inductive transfer holds on independent benchmarks.
GRATE: Temporal Extensions for Inductive KG Foundation Models via Gated Rotary Attention
arXiv:2607.10197v2 Announce Type: replace Abstract: Knowledge graph foundation models such as Ultra and Trix achieve strong inductive transfer by learning relation-graph representations that generalise to unseen entities and relations. Extending this transferability to temporal knowledge graphs (TKGs) remains challenging: existing temporal models tie their parameters to dataset-specific entities, relations, or timestamps and are not designed to transfer to TKGs with disjoint vocabularies. We propose GRATE (Gated Rotary Attention for Temporal Encoding), an entity-side message function that adds no learnable parameters and encodes time through relative time differences by rotating each edge message according to its time gap to the query and applying a query-conditioned gate to select temporally relevant signals. GRATE integrates into NBFNet-style KG foundation models while preserving structural transferability. Existing TKG benchmarks evaluate within shared train/test vocabularies and cannot directly test cross-dataset temporal transfer; we therefore construct GDELTIndT and WIKIIndT, inductive transfer benchmark suites with disjoint entities, relations, and timestamps spanning both interpolation and extrapolation. Across these benchmarks and held-out forecasting datasets, a single jointly pretrained GRATE checkpoint improves over the static base model in most settings.
Abstract:Knowledge graph foundation models such as Ultra and Trix achieve strong inductive transfer by learning relation-graph representations that generalise to unseen entities and relations. Extending this transferability to temporal knowledge graphs (TKGs) remains challenging: existing temporal models tie their parameters to dataset-specific entities, relations, or timestamps and are not designed to transfer to TKGs with disjoint vocabularies. We propose GRATE (Gated Rotary Attention for Temporal Encoding), an entity-side message function that adds no learnable parameters and encodes time through relative time differences by rotating each edge message according to its time gap to the query and applying a query-conditioned gate to select temporally relevant signals. GRATE integrates into NBFNet-style KG foundation models while preserving structural transferability. Existing TKG benchmarks evaluate within shared train/test vocabularies and cannot directly test cross-dataset temporal transfer; we therefore construct GDELTIndT and WIKIIndT, inductive transfer benchmark suites with disjoint entities, relations, and timestamps spanning both interpolation and extrapolation. Across these benchmarks and held-out forecasting datasets, a single jointly pretrained GRATE checkpoint improves over the static base model in most settings.
Accepted at the ICML 2026 Workshop on Graph Foundation Models: A New Era for Graph Machine Learning. 17 pages, 4 figures
Artificial Intelligence (cs.AI)
arXiv:2607.10197 [cs.AI]
(or arXiv:2607.10197v2 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2607.10197 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 287 cuvinte
Subiecte conexe
A Statistical Perspective on Knowledge Distillation: Foundations, Classical Methods, and Large Language Model Extensions
arXiv · 30 sept. 2026
FMplex: Model Virtualization for Serving Extensible Foundation Models
arXiv · 9 iun. 2026
ReLEAP: Upcycling Vision Datasets through Cross-Task Unification and Semi-Supervised Annotation
arXiv · 27 oct. 2026
A Unified Policy Architecture (UPA): The Governance Kernel for Enterprise AI Operating Systems
arXiv · 9 sept. 2026