ViSR-KGC: Visual Subgraph Reasoning with Vision-Language Models for Multimodal Knowledge Graph Completion

Hongan Wang; Jiafan Li; Jiaqi Zhu; Liang Chang; Mengxue Yang; Ying Li

REVIEW 4 major objections 6 minor 47 references

ViSR-KGC: Visual Subgraph Reasoning with Vision-Language Models for Multimodal Knowledge Graph Completion

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Visualizing query-relevant subgraphs as images lets a vision-language model outperform embedding- and text-based baselines for multimodal knowledge graph completion, including by retrieving correct answers outside the embedding candidate…

desk verdict A plausible new pipeline for VLM-based MMKGC, but the headline comparison rests on an unverified test-set mismatch and the visual mechanism itself is only weakly supported by its own ablation. read the letter →

arxiv 2608.05833 v1 pith:XVXHZEE6 submitted 2026-08-06 cs.AI

Jiafan Li , Mengxue Yang , Jiaqi Zhu , Liang Chang , Ying Li , Hongan Wang This is my paper · ORCID

classification cs.AI

keywords multimodalknowledgegraphcompletionlinkpredictionvision-languagemodelsubgraphreasoningvisualizationquery-aware

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that multimodal knowledge graph completion can be improved by turning the graph itself into something a vision-language model can see. Instead of scoring triples with embeddings alone or linearizing the graph into text, ViSR-KGC extracts a compact subgraph around the query, renders it as an image, and prompts a vision-language model to name the missing entity using the subgraph image, entity images, a textual serialization, and a candidate list. The claim is that this visual-subgraph mode of reasoning combines global topology from representation learning, local multimodal evidence, and the model's internal commonsense, and that the combination beats embedding-based, LLM-based, and VLM-based baselines on two real-world datasets. The authors report that about 10–11% of correct predictions fall outside the embedding candidate list, which they take as evidence of active reasoning rather than reranking.

What carries the argument

The machinery that carries the argument is the visualized query-aware subgraph: a small image showing the query entity's one-hop and two-hop neighbors together with edges that share the query relation, laid out hierarchically to minimize crossings and pasted with entity images at nodes. It is the only input that exposes graph topology to the VLM in a directly perceivable form; the same triples are also serialized as text, but the paper's ablation shows the image and text carry complementary signals, with text contributing more. Around this core, the method relies on a multimodal encoder whose four-view embeddings select which edges enter the subgraph, and on a candidate list that guides the VLM without hard-constraining it.

What would settle it

Run the full ViSR-KGC prompt but replace the subgraph image with a scrambled or randomized image of the same size while keeping the textual triple serialization unchanged; if Hit@1 stays at the reported level (0.6715 on DB15K), the visual topology is not doing the claimed work. A complementary check is an occlusion or attention analysis over the rendered graph to see whether the VLM's output depends on the drawn nodes and edges.

Watch

Extended reading notes

Core claim

The central discovery, stated on the paper's own terms, is that explicit visual presentation of graph topology unlocks VLM reasoning for link prediction. Given a query (h,r,?), the method learns text, visual, structural, and fused embeddings for each entity; scores candidate edges by relation and entity similarity, with a bonus for one-hop and relation-consistent edges; and keeps the top relation-consistent, one-hop, and two-hop edges as a compact subgraph. That subgraph is rendered with a hierarchical layout, entity images are attached to nodes, and the VLM receives the subgraph image, entity images, a textual list of triples, the query, and a candidate list, and generates the missing entity. On FB15K-237 and DB15K, tail Hit@1 reaches 0.8027 and 0.6715, respectively, above every compared baseline, and the paper further shows gains on low-frequency relations and on cases where the correct entity is absent from the embedding candidate list. The authors interpret this as the VLM tracing topology-aware paths in the image and supplementing them with internal commonsense.

Load-bearing premise

The load-bearing premise is that the vision-language model actually reasons from the rendered subgraph image—that seeing the topology adds signal beyond the same triples written out as text—because if the image is ignored, the method reduces to a text-prompt model and its core mechanism is not being exercised.

Editorial extensions

If this is right

Making graph topology visually explicit is a workable alternative to textual graph linearization for large-model KG reasoning; the reported gains over LLM-based baselines support this.
Pretrained commonsense in the VLM can compensate for sparse relation evidence, improving completion for low-frequency relations without additional training.
Candidate lists from embeddings can serve as soft guidance: the VLM can and does return correct entities outside the list, so the method is not merely reranking.
Subgraph quality is a bottleneck: replacing the multimodal encoder with weaker backbones substantially lowers accuracy, so representation learning and VLM reasoning are complementary.
Rendering choices matter: switching from the hierarchical dot layout to twopi or sfdp degrades performance, making layout part of the method rather than a cosmetic detail.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The same render-and-prompt recipe could be carried to other structured reasoning tasks, such as graph question answering or node classification, whenever a VLM is the reasoning engine and the graph is small enough to render.
The ablation shows the subgraph image contributes a smaller gain than the textual serialization; a natural extension is to test whether higher-resolution rendering, colored edge labels, or layout optimization would widen the image's contribution.
Because the gains rely on the VLM's internal knowledge, the method's margin over baselines may shrink with smaller or less knowledge-rich VLMs; measuring performance across model scales would map where the visual-topology mechanism helps.
The evaluation filters test queries to those with a unique correct answer; applying the approach to open-world or ambiguous queries would require handling multiple valid answers rather than a single ranked entity.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about ViSR-KGC before reading further. First, it is a real attempt to do something new: render query-aware subgraphs as images and let a VLM reason over them for multimodal knowledge graph completion. The combination of IMF-style multimodal embeddings, similarity-based subgraph extraction, layout-selected graph rendering, and Qwen3VL prompting is, as far as the text shows, not previously applied to MMKGC. That is worth acknowledging. The framework is clearly explained, the algorithm pseudocode is concrete, and the ablations cover the main design choices: subgraph edge types, layout strategy, candidate list, and encoder backbone. The authors also report that roughly 10-11% of correct predictions fall outside the embedding-based candidate list, which is an honest attempt to show the VLM is not just reranking the encoder's output. Those parts are decent engineering work. The soft spot is the comparison in Table 2. The paper filters the test set to unique-answer triples, leaving 6,102 FB15K-237 and 8,355 DB15K test triples, and then copies baseline numbers from the original papers under a caption claiming same settings. No code, data, or supplementary materials are released to verify that the original baseline numbers were computed on this filtered subset. The supplementary link is literally the placeholder arXiv id. If the baselines were evaluated on the unfiltered sets, every ViSR-KGC advantage in Table 2 could be an artifact of the easier filtered evaluation. This is load-bearing, not a minor presentation issue, because the abstract's central claim is consistently outperforms SOTA baselines. The stress-test note has this right. The second weakness is one the authors themselves expose. Their ablation on DB15K shows that removing the subgraph image drops tail Hit@1 from 0.6715 to 0.6667, while removing the subgraph text drops it to 0.5075. The paper interprets this as complementarity, but honestly it reads as the VLM mostly ignoring the image. Given the title and framing are about visual subgraph reasoning, that is a credibility problem for the core mechanism, even if the pipeline as a whole performs well. There are also no error bars or multiple seeds for stochastic VLM decoding, which matters when the reported improvements over MLaGA are a few points. On balance, the paper deserves referee time, not because the claims are verified but because the idea is timely and the engineering is substantial enough to be worth a careful revision. A serious referee should insist on code/data release, rerunning baselines on the exact filtered protocol, and a sharper analysis of what the image actually contributes. I would not cite the current numbers in my own work until those are resolved, but I would bring the paper to a reading group to discuss the evaluation pitfalls. My recommendation: send it to peer review, with the expectation of major revision or rejection as submitted.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes ViSR-KGC, a multimodal knowledge graph completion (MMKGC) method that first learns multimodal entity representations with an IMF-style encoder, extracts a query-aware subgraph via embedding-based relevance scoring, renders the subgraph as an image using a Graphviz layout, and then prompts the Qwen3VL vision-language model with the subgraph image, entity images, textual serialization, and a candidate list to predict the missing entity. Experiments on FB15K-237 and DB15K report Hit@1/Hit@3 improvements over structure-only, text-enhanced, multimodal embedding, LLM-based, and VLM-based baselines, together with ablations and hyperparameter analyses. The paper claims that visual subgraph topology and the VLM's internal knowledge jointly contribute, with roughly 10% of correct answers falling outside the embedding candidate list.

Significance. If the empirical claims were validated under a matched evaluation protocol, the paper would make a useful contribution: it is among the first to convert KG subgraphs into rendered images for VLM-based link prediction, and the out-of-candidate-set examples provide a concrete, falsifiable demonstration of VLM commonsense reasoning beyond embedding scoring. The pipeline is clearly described, and the ablation study covers several components. However, the current evidence does not support the headline SOTA claim because the test set is nonstandard, baseline numbers may come from different protocols, and no code or data are released; in addition, the visual modality's marginal contribution in the ablation is small and not shown to be statistically significant. The significance is therefore conditional on a major reworking of the experimental validation.

major comments (4)

[Section 4.1, Table 2] The central SOTA claim is not supported by a matched comparison. The test sets are filtered to triples with a unique correct answer (FB15K-237: 6,102 test triples instead of the standard 20,466), while Table 2 states that some baseline results are directly copied from the original papers 'under the same settings.' No code, data, or supplementary material is provided to verify that the original baseline evaluations used this filtered subset, and the supplementary link in the abstract is a literal placeholder. If the copied baseline numbers come from unfiltered test sets, the ViSR-KGC advantages in Table 2 could be entirely due to the easier evaluation subset. The authors need to rerun every baseline on exactly the same filtered split, report full unfiltered results as well, and release code and data to demonstrate protocol alignment.
[Section 4.6, Table 3] The core claim that visual subgraph structure provides critical reasoning signals is weakly supported by the ablation. Removing the subgraph image lowers DB15K tail Hit@1 from 0.6715 to 0.6667 and head Hit@1 from 0.3589 to 0.3542, whereas removing the subgraph text lowers tail Hit@1 to 0.5075. The paper reports no repeated runs, seeds, or error bars for stochastic VLM decoding, so a drop of about 0.005 is within plausible run-to-run variation. This is not enough to establish that the VLM uses the rendered image rather than the textual serialization. The authors should report variance over multiple decoding runs and provide case-level evidence where the image changes the prediction, or temper the visual-reasoning claims.
[Section 4.7, Figures 4 and 5] The hyperparameter analysis does not state which split is used. If the curves for kmax and lambda are computed on the test set, then the default values are selected on test data and the reported numbers are optimistically biased. The paper should explicitly state that validation was used for tuning or should re-run the analysis on the validation split; the same clarification is needed for the layout comparison said to be in the Supplementary Materials, which is not currently available.
[Section 3.6, Section 4.5] The candidate-list and generation protocol is underspecified. Equation (8) defines prediction as argmax over the entity set, but the implementation uses prompt-based generation; the paper does not report the candidate list size K, the decoding parameters (temperature, top-p, number of samples), or how generated text is mapped to entities. The claims that 9.96% and 11.24% of correct predictions fall outside the candidate set depend on these details and cannot be reproduced. Please specify the full protocol and, ideally, release the prompts and model outputs.

minor comments (6)

[Abstract, Section 1] The supplementary link in the abstract and introduction is the literal placeholder 'http://arxiv.org/abs/xxxx.xxxxx' and must be replaced with a working URL.
[Section 4.2] There is a typo in the first paragraph: 'structral' should be 'structural'.
[Table 2] The NativE rows report missing Hit@3 values marked with '-' on both datasets; the paper should explain whether these results are unavailable or were not reported in the original paper.
[Section 4.5] In the paragraph beginning 'However,Visualizing graph structure', there is a missing space and an incorrect capitalization; it should read 'However, visualizing graph structure...'.
[Section 4.6] The ablation study is reported only on DB15K; adding FB15K-237 ablation results would strengthen the claim that the component contributions generalize across datasets.
[References] Reference [25] is listed with '[n. d.]' even though the text supplies a 2018 publication year; the bibliographic metadata should be completed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: held-out empirical evaluation, with explicit evidence of predictions outside the encoder's candidate list.

full rationale

The central claim is an empirical comparison on held-out test triples, not a derivation from the model's own outputs. The pipeline is: train IMF on training triples; extract a query subgraph from the training graph using IMF embeddings; render it; prompt Qwen3-VL with subgraph image, entity images, text serialization, and a candidate list; measure Hit@1/Hit@3. No equation defines the prediction as a function of the training loss or of the candidate list; Equation (8) conditions on the prompt and allows any entity in E. The paper provides direct evidence against the main circularity risk (the VLM merely re-emitting the encoder's candidate): Section 4.5 states that 'Around 9.96% and 11.24% of correct predicted entities on FB15K-237 and DB15K respectively, fall outside the candidate set.' Ablations (Table 3) show that removing the subgraph text, subgraph image, candidate entities, or entity images each changes performance, and replacing the IMF backbone with ConvE/ConvKB/HGNN-IMA changes results, so the components are not tautologically equivalent. The cited self-works (HGNN-IMA [16], SLiNT [42], and the VLM graph benchmark [1]) are baselines or related work, not load-bearing premises; the uniqueness or correctness of the method does not rest on them. The only notable deficiencies are non-circular: the supplementary link is the literal placeholder 'http://arxiv.org/abs/xxxx.xxxxx', and Table 2 copies some baseline numbers from original papers while the test set is filtered to unique-answer triples (6,102 FB15K-237 test triples vs. the standard 20,466), so protocol alignment is unverified. These are evaluation-transparency and reproducibility concerns, not examples of the prediction reducing to its inputs.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new theoretical entities. Its central claim depends on several externally fitted components and on domain assumptions about VLM perception and evaluation comparability.

free parameters (7)

lambda = 0.5
Balances relation similarity and entity similarity in subgraph edge scoring (Eq. 4); tuned via hyperparameter analysis on DB15K validation (Figure 5).
gamma = 0.1
Bonus weight for edges sharing an element with the query (Eq. 6); chosen by hand in Section 4.4.
k_rel = 10
Number of top relation-consistent edges selected for the subgraph; set in Section 4.4.
k_nei = 10
Number of top neighbor edges selected for the subgraph; set in Section 4.4.
k_max = 15
Maximum number of edges in the extracted subgraph; tuned via hyperparameter analysis (Figure 4).
layout strategy = dot
Graphviz layout chosen after an empirical comparison that is only described in the missing supplementary materials.
IMF encoder and Qwen3VL backbone
Pretrained multimodal encoder and VLM are external components chosen to maximize validation performance; the ablation shows the pipeline is sensitive to this choice.

assumptions (4)

domain assumption A VLM can extract meaningful relational information from a rendered graph image.
The whole method relies on the VLM interpreting the Graphviz-rendered subgraph image; the ablation (w/o subgraph image) gives only weak support for this.
domain assumption Unique-answer filtering produces a valid and comparable evaluation benchmark for KGC.
Section 4.1 filters test triples to those with a unique correct answer, removing ambiguous cases; this changes the task difficulty and comparability.
domain assumption Baseline numbers copied from other papers were produced under the same filtered evaluation protocol.
Section 4.2 says baseline results are copied 'under the same settings', but no evidence is given that the original papers used the same filter.
domain assumption The multimodal encoder IMF provides reliable entity embeddings for subgraph extraction and candidate ranking.
IMF is an external method adopted in Section 3.3; the ablation with weaker encoders shows the pipeline is sensitive to this choice.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ViSR-KGC: Visual Subgraph Reasoning with Vision-Language Models for Multimodal Knowledge Graph Completion." pith.science (2026). https://pith.science/paper/XVXHZEE6

@misc{pith2026260805833,
  author       = {Pith},
  title        = {Pith review of: ViSR-KGC: Visual Subgraph Reasoning with Vision-Language Models for Multimodal Knowledge Graph Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XVXHZEE6}},
  note         = {Machine review of arXiv:2608.05833}
}

read the original abstract

Knowledge graph completion (KGC) aims to infer missing entities or relations from incomplete graph structures, and has evolved into multimodal knowledge graph completion (MMKGC), where entities are associated with multiple modalities such as text and images. Traditional representation learning approaches follow the embedding-based paradigm and may struggle when relation-specific evidence is limited. Meanwhile, LLM-based reasoning methods typically linearize graph structures into textual prompts, which obscures structural topology and neglects vital visual information. While vision-language models (VLMs) excel at multimodal reasoning, they cannot natively interpret structured graph topology, particularly when it comes to knowledge graphs where nodes and edges carry complex semantics. To bridge this gap, we propose ViSR-KGC, a visual subgraph reasoning approach for KGC. It integrates three complementary capabilities to capture semantic correlations: identifying global topology dependencies via representation learning, analyzing local multimodal evidence using VLMs, and providing necessary commonsense knowledge inherent in pre-trained models. Based on learned multimodal embeddings, our framework first extracts a compact and query-aware subgraph from the MMKG. Then, this subgraph is transformed into a visually interpretable image using a layout strategy selected through empirical comparison.Finally, the visualized subgraph, entity images, textual descriptions, and candidate answers are combined into a unified prompt, enabling the VLM to infer the missing entity.

Figures

Figures reproduced from arXiv: 2608.05833 by the authors.

**Figure 2.** Overview of the proposed ViSR-KGC framework. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Example of link prediction by ViSR-KGC compared to LLM-based methods. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Effect of 𝑘max on Hit@1 and Hit@3 on DB15K [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

**Figure 5.** Figure 5: Effect of 𝜆 on Hit@1 and Hit@3 on DB15K. measure would weaken the model’s ability to distinguish entities. Therefore, 𝜆 = 0.5 is a tradeoff value and used as default. 5 Conclusion This paper delves into an underexplored area, multimodal knowledge graph completion (MMK…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 41 canonical work pages

[1]

Qihang Ai, Jiafan Li, Jincheng Dai, Jianwu Zhou, Lemao Liu, Haiyun Jiang, and Shuming Shi. 2024. Advancement in Graph Understanding: A Multimodal Benchmark and Fine-Tuning of Vision-Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (ACL). 7485–7501

work page 2024
[2]

Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

arXiv 2025
[3]

Ivana Balazevic, Carl Allen, and Timothy Hospedales. 2019. TuckER: Tensor Fac- torization for Knowledge Graph Completion. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 5185–5194

work page 2019
[4]

Antoine Bordes, Nicolas Usunier, Alberto Garcia-Durán, Jason Weston, and Ok- sana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. InProceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2 (NeurIPS). 2787–2795

work page 2013
[5]

Hammerla

Dan Busbridge, Dane Sherburn, Pietro Cavallo, and Nils Y. Hammerla. 2019. Relational Graph Attention Networks. arXiv:1904.05811 [cs.LG]

arXiv 2019
[6]

Pan, Hua- jun Chen, and Wen Zhang

Zhuo Chen, Yin Fang, Yichi Zhang, Lingbing Guo, Jiaoyan Chen, Jeff Z. Pan, Hua- jun Chen, and Wen Zhang. 2025. Noise-powered Multi-modal Knowledge Graph Representation Framework. InProceedings of the 31st International Conference on Computational Linguistics (COLING). 141–155

work page 2025
[7]

Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2D knowledge graph embeddings. InProceedings of the Thirty- Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Ap- plications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligenc...

work page 2018
[8]

Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn, Ni Lao, Kevin Mur- phy, Thomas Strohmann, Shaohua Sun, and Wei Zhang. 2014. Knowledge Vault: A Web-Scale Approach to Probabilistic Knowledge Fusion. InProceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD). 601–610

work page 2014

Show all 47 references

[9]

Dongzhe Fan, Yi Fang, Jiajin Liu, Djellel Difallah, and Qiaoyu Tan. 2025. MLaGA: Multimodal Large Language and Graph Assistant.CoRRabs/2506.02568 (2025). arXiv:2506.02568

2025 arXiv
[10]

Gansner and Stephen C

Emden R. Gansner and Stephen C. North. 2000. An open graph visualization system and its applications to software engineering.Softw. Pract. Exper.30, 11 (Sept. 2000), 1203–1233

2000
[11]

Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Xu Tang, Yao Hu, and Shaohui Lin. 2026. Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models. In14th International Conference on Learning Representations (ICLR)

2026
[12]

Wei Huang, Peining Li, Meiyu Liang, Xu Hou, Junping Du, Yingxia Shao, Guanhua Ye, Wu Liu, Kangkang Lu, and Yang Yu. 2026. ELMM: Efficient Lightweight Mul- timodal Large Language Models for Multimodal Knowledge Graph Completion. arXiv:2510.16753 [cs.AI]

2026
[13]

Thanh Le, Nam Le, and Bac Le. 2023. Knowledge graph embedding by relational rotation and complex convolution for link prediction.Expert Syst. Appl.214, C (March 2023), 23 pages

2023
[14]

Junlin Lee, Yequan Wang, Jing Li, and Min Zhang. 2024. Multimodal Reasoning with Multimodal Knowledge Graph. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (ACL). 10767–10782

2024
[15]

Jens Lehmann, Robert Isele, Max Jakob, Anja Jentzsch, Dimitris Kontokostas, Pablo N Mendes, Sebastian Hellmann, Mohamed Morsey, Patrick Van Kleef, Sören Auer, et al. 2015. Dbpedia–a large-scale, multilingual knowledge base extracted from wikipedia.Semantic web6, 2 (2015), 167–195

2015
[16]

Jiafan Li, Jiaqi Zhu, Liang Chang, Yilin Li, Miaomiao Li, Yang Wang, Yi Yang, and Hongan Wang. 2025. Representation Learning with Mutual Influence of Modalities for Node Classification in Multi-Modal Heterogeneous Networks. InProceedings of the Thirty-Fourth International Join...

2025
[17]

Xinhang Li, Xiangyu Zhao, Jiaxing Xu, Yong Zhang, and Chunxiao Xing. 2023. IMF: Interactive Multimodal Fusion Model for Link Prediction. InProceedings of the ACM Web Conference 2023 (WWW). 2572–2580

2023
[18]

Yunxin Li, Baotian Hu, Haoyuan Shi, Wei Wang, Longyue Wang, and Min Zhang
[19]

Zhen Li, Jibin Wang, Zhuo Chen, Kun Wu, Meng Ai, Leike An, Liqiang Wang, and Haoxuan Li. 2025. Unifying Within and Across: Intra-Modality Multi-View Fusion and Inter-Modality Alignment for Knowledge Graph Completion. InProceedings of the IEEE International Conference on Acoust...

2025
[20]

Yankai Lin, Zhiyuan Liu, Maosong Sun, Yang Liu, and Xuan Zhu. 2015. Learning entity and relation embeddings for knowledge graph completion. InProceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence (AAAI). 2181–2187

2015
[21]

Yang Liu, Zequn Sun, Guangyao Li, and Wei Hu. 2022. I Know What You Do Not Know: Knowledge Graph Embedding via Co-distillation Learning. InProceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM), Mohammad Al Hasan and Li Xiong (Eds.). ...

2022
[22]

Yang Liu, Xiaobin Tian, Zequn Sun, and Wei Hu. 2024. Finetuning Generative Large Language Models with Discrimination Instructions for Knowledge Graph Completion. InThe Semantic Web - ISWC 2024 - 23rd International Semantic Web Conference, Baltimore (ISWC) (Lecture Notes in Com...

2024
[23]

Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. 2025. Visual-RFT: Visual Reinforcement Fine-Tuning. InIEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2034– 2044

2025
[24]

Ying Ma, Owen Burns, Mingqiu Wang, Gang Li, Nan Du, Laurent El Shafey, Liqiang Wang, Izhak Shafran, and Hagen Soltau. 2024. Knowledge Graph Reason- ing with Self-supervised Reinforcement Learning.CoRRabs/2405.13640 (2024). arXiv:2405.13640

2024 arXiv
[25]

Dai Quoc Nguyen, Tu Dinh Nguyen, Dat Quoc Nguyen, and Dinh Phung. [n. d.]. A Novel Embedding Model for Knowledge Base Completion Based on Convolu- tional Neural Network. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Lin...

2018
[26]

Maximilian Nickel, Volker Tresp, and Hans-Peter Kriegel. 2011. A three-way model for collective learning on multi-relational data. InProceedings of the 28th International Conference on International Conference on Machine Learning (ICML). 809–816

2011
[27]

Talukdar

Apoorv Saxena, Aditay Tripathi, and Partha P. Talukdar. 2020. Improving Multi- hop Question Answering over Knowledge Graphs using Knowledge Base Em- beddings. InProceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics (ACL). 4498–4507

2020
[28]

Yuanfu Sun, Kang Li, Pengkang Guo, Jiajin Liu, and Qiaoyu Tan. 2026. Mario: Mul- timodal Graph Reasoning with Large Language Models. arXiv:2603.05181 [cs.CV]

2026
[29]

Kristina Toutanova, Danqi Chen, Patrick Pantel, Hoifung Poon, Pallavi Choud- hury, and Michael Gamon. 2015. Representing Text for Joint Embedding of Text and Knowledge Bases. InProceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP). 1499–1509

2015
[30]

Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. InProceedings of the 33rd International Conference on International Conference on Machine Learning - Volume 48(New York, NY, USA)(ICML)...

2016
[31]

Bo Wang, Tao Shen, Guodong Long, Tianyi Zhou, Ying Wang, and Yi Chang. [n. d.]. Structure-Augmented Text Representation Learning for Efficient Knowl- edge Graph Completion. InWWW ’21: The Web Conference 2021 (WWW). 1737– 1748

2021
[32]

Yang, Yuan Gao, Jihong Ding, Xiaokang Zhou, and Huazhong Liu

Hao Wang, Jing Yang, Laurence T. Yang, Yuan Gao, Jihong Ding, Xiaokang Zhou, and Huazhong Liu. 2024. MvTuckER: Multi-view knowledge graphs represen- tation learning based on tensor Tucker model.Information Fusion106 (2024), 102249

2024
[33]

Liang Wang, Wei Zhao, Zhuoyu Wei, and Jingming Liu. 2022. SimKGC: Simple Contrastive Knowledge Graph Completion with Pre-trained Language Models. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (ACL). 4281–4294

2022
[34]

Meng Wang, Sen Wang, Han Yang, Zheng Zhang, Xi Chen, and Guilin Qi. 2021. Is Visual Context Really Helpful for Knowledge Graph? A Representation Learning Perspective. InProceedings of the 29th ACM International Conference on Multime- dia (MM). 2735–2743. MM ’26, November 10–14...

2021
[35]

Minzheng Wang, Nan Xu, Jiahao Zhao, Yin Luo, and Wenji Mao. 2024. PromISe: Releasing the Capabilities of LLMs with Prompt Introspective Search. InProceed- ings of the 2024 Joint International Conference on Computational Linguistics, Lan- guage Resources and Evaluation (LREC-CO...

2024
[36]

Xiang Wang, Xiangnan He, Yixin Cao, Meng Liu, and Tat-Seng Chua. 2019. KGAT: Knowledge Graph Attention Network for Recommendation. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD). ACM, 950–958

2019
[37]

Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng Chen. 2014. Knowledge graph embedding by translating on hyperplanes. InProceedings of the Twenty- Eighth AAAI Conference on Artificial Intelligence (AAAI). 1112–1119

2014
[38]

Kwok, and Yu Zhang

Yanbin Wei, Shuai Fu, Weisen Jiang, Zejian Zhang, Zhixiong Zeng, Qi Wu, James T. Kwok, and Yu Zhang. 2024. GITA: Graph to Visual and Textual Integration for Vision-Language Graph Reasoning. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Inf...

2024
[39]

Kwok, and Yu Zhang

Yanbin Wei, Qiushi Huang, James T. Kwok, and Yu Zhang. 2024. KICGPT: Large Language Model with Knowledge in Context for Knowledge Graph Completion. CoRRabs/2402.02389 (2024). arXiv:2402.02389

2024 arXiv
[40]

Ruobing Xie, Zhiyuan Liu, Huanbo Luan, and Maosong Sun. 2017. Image- embodied knowledge representation learning. InProceedings of the 26th Interna- tional Joint Conference on Artificial Intelligence (IJCAI). 3140–3146

2017
[41]

Bishan Yang, Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. Em- bedding Entities and Relations for Learning and Inference in Knowledge Bases. In3rd International Conference on Learning Representations (ICLR)

2015
[42]

Mengxue Yang, Chun Yang, Jiaqi Zhu, Jiafan Li, Jingqi Zhang, Yuyang Li, and Ying Li. 2025. SLiNT: Structure-aware Language Model with Injection and Contrastive Training for Knowledge Graph Completion. InFindings of the Association for Computational Linguistics: EMNLP 2025. 13658–13671

2025
[43]

Weihang Zhang and Ovidiu Serban. [n. d.]. LLM-based Reranking and Validation of Knowledge Graph Completion. InProceedings of the 6th International Workshop on Knowledge Graph Construction co-located with 22nd Extended Semantic Web Conference (ESWC) (CEUR Workshop Proceedings)....
[44]

Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Binbin Hu, Ziqi Liu, Wen Zhang, and Huajun Chen. 2024. NativE: Multi-modal Knowledge Graph Comple- tion in the Wild. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retriev...

2024
[45]

Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Binbin Hu, Ziqi Liu, Wen Zhang, and Huajun Chen. 2025. Tokenization, Fusion, and Augmentation: To- wards Fine-grained Multi-modal Entity Representation. InProceedings of the AAAI Conference on Artificial Intelligence (AAAI), Vol...

2025
[46]

Yongchao Zhou, Andrei Ioan Muresanu, Ziwen Han, Keiran Paster, Silviu Pitis, Harris Chan, and Jimmy Ba. 2023. Large Language Models are Human-Level Prompt Engineers. InThe Eleventh International Conference on Learning Repre- sentations. https://openreview.net/forum?id=92gvk82DE-

2023
[2024]

InProceedings of the 41st International Conference on Machine Learning (ICML)

VisionGraph: leveraging large multimodal models for graph theory prob- lems in visual context. InProceedings of the 41st International Conference on Machine Learning (ICML). Article 1118, 17 pages

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Qihang Ai, Jiafan Li, Jincheng Dai, Jianwu Zhou, Lemao Liu, Haiyun Jiang, and Shuming Shi. 2024. Advancement in Graph Understanding: A Multimodal Benchmark and Fine-Tuning of Vision-Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (ACL). 7485–7501

work page 2024

[2] [2]

Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

arXiv 2025

[3] [3]

Ivana Balazevic, Carl Allen, and Timothy Hospedales. 2019. TuckER: Tensor Fac- torization for Knowledge Graph Completion. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 5185–5194

work page 2019

[4] [4]

Antoine Bordes, Nicolas Usunier, Alberto Garcia-Durán, Jason Weston, and Ok- sana Yakhnenko. 2013. Translating embeddings for modeling multi-relational data. InProceedings of the 27th International Conference on Neural Information Processing Systems - Volume 2 (NeurIPS). 2787–2795

work page 2013

[5] [5]

Hammerla

Dan Busbridge, Dane Sherburn, Pietro Cavallo, and Nils Y. Hammerla. 2019. Relational Graph Attention Networks. arXiv:1904.05811 [cs.LG]

arXiv 2019

[6] [6]

Pan, Hua- jun Chen, and Wen Zhang

Zhuo Chen, Yin Fang, Yichi Zhang, Lingbing Guo, Jiaoyan Chen, Jeff Z. Pan, Hua- jun Chen, and Wen Zhang. 2025. Noise-powered Multi-modal Knowledge Graph Representation Framework. InProceedings of the 31st International Conference on Computational Linguistics (COLING). 141–155

work page 2025

[7] [7]

Tim Dettmers, Pasquale Minervini, Pontus Stenetorp, and Sebastian Riedel. 2018. Convolutional 2D knowledge graph embeddings. InProceedings of the Thirty- Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Ap- plications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligenc...

work page 2018

[8] [8]

Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn, Ni Lao, Kevin Mur- phy, Thomas Strohmann, Shaohua Sun, and Wei Zhang. 2014. Knowledge Vault: A Web-Scale Approach to Probabilistic Knowledge Fusion. InProceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD). 601–610

work page 2014