# When Modalities Remember: Continual Learning for Multimodal Knowledge Graphs

Source: [When Modalities Remember: Continual Learning for Multimodal Knowledge Graphs](<https://www.sota2.com/research/paper/69d33dda7d0172140ec734f4>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-08-07T05:00:01.339Z

Authors: Linyu Li, Zhi Jin, Yichi Zhang, Dongming Jin, Yuanpeng He, Haoran Duan, Gadeng Luosang, Nyima Tashi

Published: 2026-04-03T06:40:04.000Z

## Abstract

Real-world multimodal knowledge graphs (MMKGs) are dynamic, with new entities, relations, and multimodal knowledge emerging over time. Existing continual knowledge graph reasoning (CKGR) methods focus on structural triples and cannot fully exploit multimodal signals from new entities. Existing multimodal knowledge graph reasoning (MMKGR) methods, however, usually assume static graphs and suffer catastrophic forgetting as graphs evolve. To address this gap, we present a systematic study of continual multimodal knowledge graph reasoning (CMMKGR). We construct several continual multimodal knowledge graph benchmarks from existing MMKG datasets and propose MRCKG, a new CMMKGR model. Specifically, MRCKG employs a multimodal-structural collaborative curriculum to schedule progressive learning based on the structural connectivity of new triples to the historical graph and their multimodal compatibility. It also introduces a cross-modal knowledge preservation mechanism to mitigate forgetting through entity representation stability, relational semantic consistency, and modality anchoring. In addition, a multimodal contrastive replay scheme with a two-stage optimization strategy reinforces learned knowledge via multimodal importance sampling and representation alignment. Experiments on multiple datasets show that MRCKG preserves previously learned multimodal knowledge while substantially improving the learning of new knowledge.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Link Prediction on DB15K Entity split](<https://www.sota2.com/research/sota/link-prediction-on-db15k-entity-split.md>): Link Prediction; DB15K Entity split; MRR: 28.56.

- [Link Prediction on DB15K Higher split](<https://www.sota2.com/research/sota/link-prediction-on-db15k-higher-split.md>): Link Prediction; DB15K Higher split; MRR: 0.2614.

- [Link Prediction on DB15K Equal split](<https://www.sota2.com/research/sota/link-prediction-on-db15k-equal-split.md>): Link Prediction; DB15K Equal split; MRR: 23.05.

- [Link Prediction on MKG-W (Entity split)](<https://www.sota2.com/research/sota/link-prediction-on-mkg-w-entity-split.md>): Link Prediction; MKG-W (Entity split); MRR: 27.12.

- [Link Prediction on MKG-W (Higher split)](<https://www.sota2.com/research/sota/link-prediction-on-mkg-w-higher-split.md>): Link Prediction; MKG-W (Higher split); MRR: 29.53.

- [Link Prediction on MKG-W Equal split](<https://www.sota2.com/research/sota/link-prediction-on-mkg-w-equal-split.md>): Link Prediction; MKG-W Equal split; MRR: 0.2183.

- [Link Prediction on MKG-Y Entity split](<https://www.sota2.com/research/sota/link-prediction-on-mkg-y-entity-split.md>): Link Prediction; MKG-Y Entity split; MRR: 0.2518.

- [Link Prediction on MKG-Y (Higher split)](<https://www.sota2.com/research/sota/link-prediction-on-mkg-y-higher-split.md>): Link Prediction; MKG-Y (Higher split); MRR: 32.85.

- [Link Prediction on MKG-Y Equal split](<https://www.sota2.com/research/sota/link-prediction-on-mkg-y-equal-split.md>): Link Prediction; MKG-Y Equal split; MRR: 31.46.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.02778>)

[PDF](<https://arxiv.org/pdf/2604.02778.pdf>)