Revisiting Catastrophic Forgetting in Continual Knowledge Graph Embedding

The deeper issue here is not just that models forget old facts, but that the standard benchmarks used to measure forgetting were never designed to catch interference from new entities in the first place. The field has been grading itself on a partial rubric.

This connects most directly to the work covered in 'How Embeddings Shape Graph Neural Networks' from mid-April, which highlighted how embedding strategy choices ripple through model behavior in ways that controlled benchmarks can obscure. That paper isolated embedding impact by standardizing everything else; this paper makes the inverse point, that standardized evaluation can hide embedding-level failures entirely. More broadly, the embedding quality and compression threads running through recent arXiv coverage (including the K-Token Merging piece from April 16) suggest the field is in an active period of re-examining what embeddings actually preserve and what they silently discard. This story fits that pattern, applied to the specific pressure of incremental knowledge updates rather than inference efficiency.

Watch whether any of the major CKGE benchmark maintainers (FB15k-237 or YAGO variants are the likely candidates) issue revised evaluation protocols within the next two conference cycles. If they do not, this critique risks staying theoretical rather than reshaping how the subfield measures progress.