Every Little Helps: Building Knowledge Graph Foundation Model with Fine-grained Transferable Multi-modal Tokens

Yichi Zhang; Zhuo Chen; Lingbing Guo; Wen Zhang; Huajun Chen

cs.CLFeb 11, 2026

Every Little Helps: Building Knowledge Graph Foundation Model with Fine-grained Transferable Multi-modal Tokens

Authors: Yichi Zhang, Zhuo Chen, Lingbing Guo, Wen Zhang, Huajun Chen

Organizations: Zhejiang University, Zhejiang, China.

Abstract

Multi-modal knowledge graph reasoning (MMKGR) aims to predict the missing links by exploiting both graph structure information and multi-modal entity contents. Most existing works are designed for a transductive setting, which learns dataset-specific embeddings and struggles to generalize to new KGs. Recent knowledge graph foundation models (KGFMs) improve cross-KG transfer, but they mainly exploit structural patterns and ignore rich multi-modal signals. We address these gaps by proposing a token-based foundation model (TOFU) for MMKGR, which exhibits strong generalization across different MMKGs. TOFU discretizes structural, visual, and textual information into modality-specific tokens. TOFU then employs a hierarchical fusion architecture with mixture-of-message mechanisms, aiming to process these tokens and obtain transferable features for MMKGR. Experimental results on 17 transductive, inductive, and fully-inductive MMKGs show that TOFU generally outperforms strong KGFM and MMKGR baselines, delivering strong performance on unseen MMKGs.

Figures & tables

Figure 1: A comparison between existing MMKGR/KGFM and TOFU, along with insights from large language models.

Figure 2: An overview of our TOFU framework. TOFU first models each modality (structure/visual/textual) into discrete tokens and employs a hierarchical fusion architecture to obtain the transferable entity and relation features, which consists of a structural encoder, a multi-modal encoder, and a fusion gate. Finally, TOFU applies global aggregation with a mixture-of-message mechanism to obtain multi-source information from the MMKG to make the MMKGR prediction.

Transductive MMKGs
Dataset	Train			Valid			Test
Dataset	$\|\mathcal{E}\|$	$\|\mathcal{R}\|$	$\|\mathcal{T}\|$	$\|\mathcal{E}\|$	$\|\mathcal{R}\|$	$\|\mathcal{T}\|$	$\|\mathcal{E}\|$	$\|\mathcal{R}\|$	$\|\mathcal{T}\|$
DB15K	12842	279	79222	12842	279	9902	12842	279	9904
MKG-W	15000	169	34196	15000	169	4276	15000	169	4274
MKG-Y	15000	28	21310	15000	28	2665	15000	28	2663
YAGO15K	15283	32	86020	15283	32	12289	15283	32	24577

Table 1: Statistical informtion of the 17 MMKG datasets.

Method		Transductive		Inductive		Fully-Inductive		Overall
Method		MRR	Hit@10	MRR	Hit@10	MRR	Hit@10	MRR	Hit@10
Supervised SOTA		42.47	54.70	47.73	63.58	16.55	29.63	37.61	50.89
ULTRA	zero-shot	32.62	47.86	49.68	67.18	39.45	60.73	38.24	55.43
ULTRA	fine-tune	43.59	56.03	50.83	68.18	39.03	60.35	44.22	59.90
KG-ICL	zero-shot	41.70	52.09	53.33	70.53	41.05	63.55	44.28	59.13
KG-ICL	fine-tune	41.83	52.68	54.03	71.95	44.38	66.48	45.30	60.46

Table 2: The KGFM experiment results on 17 MMKGs.

Model	DB15K				MKG-W				MKG-Y				Overall
Model	MRR	Hit@1	Hit@3	Hit@10	MRR	Hit@1	Hit@3	Hit@10	MRR	Hit@1	Hit@3	Hit@10	MRR	Hit@1	Hit@3	Hit@10
IKRL ( Xie et al. 2017 )	26.82	14.09	34.93	49.09	32.36	26.11	34.75	44.07	33.22	30.37	34.28	38.26	30.80	23.52	34.65	43.81
TBKGC ( Sergieh et al. 2018 )	28.40	15.61	37.03	49.86	31.48	25.31	33.98	43.24	33.99	30.47	35.27	40.07	31.29	23.80	35.43	44.39
TransAE ( Wang et al. 2019 )	28.09	21.25	31.17	41.17	30.00	21.23	34.91	44.72	28.10	25.31	29.10	33.03	28.73	22.60	31.73	39.64
MMKRL ( Lu et al. 2022 )	26.81	13.85	35.07	49.39	30.10	22.16	34.09	44.69	36.81	31.66	39.79	45.31	31.24	22.56	36.32	46.46
RSME ( Wang et al. 2021 )	29.76	24.15	32.12	40.29	29.23	23.36	31.97	40.43	34.44	31.78	36.07	39.09	31.14	26.43	33.39	39.94

Table 3: Single-dataset transductive experiment results on three standard MMKGR benchmarks.

Figure 3: Single-MMKG transfer experiments following the setting proposed by IndMKG ( Yang et al. 2025 ) . A

\rightarrow

B denotes training on dataset A and testing on dataset B. The abbreviation for MMKG is as follows: YAGO15K(A), MKG-Y(B), MKG-W(C), WN18RR++(D), FB15K-237(E). We compare TOFU with two baselines under both zero-shot and fine-tuning settings.

Figure 4: Ablation study on module design, modality contribution, and message function selections. The ablation study is conducted on 5 MMKGs separately, including DB15K, MKG-W, MKG-Y, YAGO15K, and WN18RR++. Overall results represent the average. We report the MRR results.

Figure 5: Case study on the mixture-of-message modules. We visualize the distribution of temperature

\gamma_{i}

for different relations and the weights for the three message functions under different query relations. Note that YAGO15K is an OOD dataset.

Method	#Params	Time Cost
MyGO	45.63M	57.76s
MoMoK	72.40M	33.38s
ULTRA	0.169M	31.12s
KG-ICL	0.054M	49.13s
MOTIF	0.181M	>500s
TOFU	0.248M	58.41s

Table 4: The efficiency analysis experiment on DB15K.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: The influence of text/visual token number.

Dataset		Metrics	Supervised SOTA	ULTRA		KG-ICL		MOTIF		TOFU
Dataset		Metrics	Supervised SOTA	pre-train	fine-tune	pre-train	fine-tune	pre-train	fine-tune	pre-train	fine-tune
Transductive	DB15K	MRR	37.23	32.98	40.72	32.19	31.95	29.48	38.59	38.38	38.62
	DB15K	Hit@10	54.71	48.36	56.59	45.11	45.23	46.34	55.76	52.73	53.02
	MKG-W	MRR	38.46	34.41	37.77	34.10	34.07	35.82	38.07	39.80	40.55
	MKG-W	Hit@10	51.46	45.68	49.13	39.23	43.37	46.99	49.60	50.71	51.70
	MKG-Y	MRR	40.03	35.62	39.35	40.76	40.61	33.94	39.44	43.72	43.65

Appendix

Table 5: The full results of the KGFM experiments on 17 MMKG datasets compared with supervised SOTA results and other KGFM methods. The supervised SOTA results are mainly provided by KG-ICL ( Cui et al. 2024 ) .

Source KG: YAGO15K
Method	MKG-Y			MKG-W			WN18RR++			FB15K-237			Average
Method	MRR	Hit@1	Hit@10	MRR	Hit@1	Hit@10	MRR	Hit@1	Hit@10	MRR	Hit@1	Hit@10	MRR	Hit@1	Hit@10
InGram	8.30	5.10	12.10	0.60	0.20	1.30	3.10	1.90	4.30	1.30	0.40	2.60	3.33	1.90	5.08
IMF	0.46	0.21	0.71	9.59	8.87	10.35	2.15	1.88	2.58	2.10	1.65	2.97	3.58	3.15	4.15
MoCi	2.06	1.51	2.63	14.27	10.23	20.30	2.55	2.01	3.78	2.43	1.79	3.67	5.33	3.89	7.60
ULTRA	35.13	31.13	42.99	29.91	23.19	43.10	27.53	19.19	42.94	15.96	9.16	29.83	27.13	20.67	39.72

Appendix

Table 6: Single MMKG transfer experiments.

Multimodal KGs

Multimodal knowledge graphs connect entities and relations with text, images, video, time, and other evidence for structured reasoning.

See all 42 papers

Graphlets as Building Blocks for Structural Vocabulary in Knowledge Graph Foundation Models
May 7, 2026Kossi Amouzouvi, Robert Wardenga, Jens Lehmann +1Graph Representation Learning Graph Foundation Models
ScaDS.AI Dresden/Leipzig, Technische Universität Dresden, Dresden, Germany · alphaspeech - c/o alpha NT GmbH, Dresden, Germany · Amazon, Technische Universität Dresden, Dresden, Germany +1
MERIT: Scaling KG Foundation Models with Relation Transformers
Oct 11, 2026Cosimo Gregucci, Daniel Hernandez, Antonio Vergari +1Graph Foundation Models KG Completion
Institute for Artificial Intelligence, University of Stuttgart, Germany · School of Informatics, University of Edinburgh, UK · University of Southampton
MMKG-R1: Constructing Multimodal Knowledge Graphs via Reinforcement Learning
Oct 14, 2026Zongyu Wu, Yilong Wang, Xiaochen Wang +5KG Construction Multimodal KGs
The Pennsylvania State University

May 7, 2026J/K move · Enter open · S save

Graphlets as Building Blocks for Structural Vocabulary in Knowledge Graph Foundation Models

Kossi Amouzouvi, Robert Wardenga, Jens Lehmann, Sahar Vahdati

ScaDS.AI Dresden/Leipzig, Technische Universität Dresden, Dresden, Germany · alphaspeech - c/o alpha NT GmbH, Dresden, Germany · Amazon, Technische Universität Dresden, Dresden, Germany+1

Graph Representation Learning Graph Foundation Models KG Completion

Open paper AskPDF arXiv

Foundation models excel at language, where sentences become tokens, and vision, where images become pixels, because both reduce to discrete symbols on a shared, fixed grid. Knowledge Graphs share the discreteness, but not the geometry. Their entities and relations are discrete symbols, yet their arrangement is relational and lacks a common, fixed grid. Knowledge Graphs (KGs) share the discreteness, but not the geometry. They form irregular, non-Euclidean topologies whose local neighborhoods differ from graph to graph. Therefore, Knowledge Graph Foundation Models (KGFMs) rely on identifying structural invariances to produce transferable representations. Without a universal token set, KGFMs are limited in their ability to transfer representations across unseen KGs. We close this gap by treating graphlets, small connected graphs, as structural tokens that recur in heterogeneous KGs. In this paper, We introduce a model-agnostic framework based on a vocabulary of graphlets that mines a KG between relations via pattern matching. In particular, we considered closed and open 2- and 3-path, and star graphlets, to obtain robust invariances. The framework is evaluated on 51 KGs from a wide range of domains, for zero-shot inductive and transductive link prediction. Experiments show that adding simple graphlets to the vocabulary yields models that outperform prior KGFMs.

Abstract

Multimodal KGs

Explore similar work

Graphlets as Building Blocks for Structural Vocabulary in Knowledge Graph Foundation Models

MERIT: Scaling KG Foundation Models with Relation Transformers

MMKG-R1: Constructing Multimodal Knowledge Graphs via Reinforcement Learning