Dual-Path LLM Reasoning for Multimodal Few-Shot Knowledge Graph Completion

# Dual-Path LLM Reasoning for Multimodal Few-Shot Knowledge Graph Completion
Source: [https://arxiv.org/html/2607.26909](https://arxiv.org/html/2607.26909)
Jinlan Liu1, Zhiying Tu1,2,3\\corresponding, Yongchao Xing1, Yicheng Liu1, Bolin Zhang1,2,3, Dianbo Sui1,2, Dianhui Chu1,2, Hongliang Sun1,2,3\\corresponding

###### Abstract

Knowledge graph completion \(KGC\) aims to infer missing facts in knowledge graphs \(KGs\), thereby improving their completeness and supporting downstream intelligent applications\. However, emerging entities and relations in real\-world deployments make inductive KGC difficult, especially under few\-shot and zero\-shot settings\. Multimodal information and Large Language Model \(LLM\)\-derived priors can enrich sparse relational contexts, but they may also introduce noisy or hallucinated evidence\. To address these issues, we propose DuPLeR, aDual\-PathLLMReasoning framework for multimodal few\-shot KGC\. DuPLeR builds a calibrated relation graph by combining multimodal LLM\-derived type priors with factual support structures, and performs dual\-level structural reasoning over the refined relation topology\. Moreover, a dual\-pathway multimodal enhancement module regulates message passing with query\-relevant multimodal signals and supplements entity representations after graph propagation\. Experiments on eight inductive variants of two multimodal KG \(MMKG\) benchmarks show that DuPLeR achieves robust performance in data\-scarce KGC scenarios\.

## 1Introduction

KGC, also known as link prediction, aims to infer missing facts in a KG, playing a critical role in maintaining the coverage and reliability of KGs\(Yuanet al\.[2025](https://arxiv.org/html/2607.26909#bib.bib46)\)\. Fact triples extracted automatically from massive logs, customer interactions, and so on are often incomplete, leaving potential relationships between entities unlinked\(Huet al\.[2024](https://arxiv.org/html/2607.26909#bib.bib2)\)\. This limits the effectiveness of KGs in enabling downstream tasks\. By performing KGC, missing links can be inferred and the KGs can be enriched, thereby enhancing their completeness and utility\.

Most existing KGC methods operate in a transductive setting, where all entities and relations are observed during training and the goal is to predict missing links among them\. However, the real world continuously witnesses the emergence of new entities that are incrementally added to the existing KG as new knowledge\(Zhanget al\.[2023](https://arxiv.org/html/2607.26909#bib.bib45)\)\. Traditional transductive KGC methods fail to update the knowledge representations of these newly added entities in a timely manner, requiring link prediction on unseen entities—known as inductive KGC\(Zhanget al\.[2024c](https://arxiv.org/html/2607.26909#bib.bib37)\)\. This capability is crucial for scalable intelligence and reducing annotation costs\. Nevertheless, inductive KGC is fundamentally constrained by the scarcity of reliable contextual evidence for emerging entities, as emerging entities often provide only limited local neighborhoods and, in some cases, only a handful of support triples\.

Multimodal information, such as textual descriptions and associated images of real\-world items, can serve as auxiliary evidence to supplement the missing contextual information in sparse triples\. By providing richer semantic signals beyond structural triples alone, multimodal cues enable models to better capture latent relationships between entities, thereby enhancing KGC performance\. Existing multimodal KGC methods have explored adaptive fusion mechanisms to selectively integrate heterogeneous modality\-specific evidence\(Chenet al\.[2022](https://arxiv.org/html/2607.26909#bib.bib25); Shanget al\.[2024](https://arxiv.org/html/2607.26909#bib.bib4); Weiet al\.[2024](https://arxiv.org/html/2607.26909#bib.bib1)\), with some further incorporating multimodal neighborhood information to enrich entity or relation representations\(Shanget al\.[2024](https://arxiv.org/html/2607.26909#bib.bib4); Weiet al\.[2024](https://arxiv.org/html/2607.26909#bib.bib1)\)\. However, multimodal integration inevitably introduces cross\-modal heterogeneity and noise, which can compromise the reliability of relational reasoning\(Yanget al\.[2025](https://arxiv.org/html/2607.26909#bib.bib40); Kimet al\.[2025](https://arxiv.org/html/2607.26909#bib.bib38); Zhaoet al\.[2024](https://arxiv.org/html/2607.26909#bib.bib34)\)\. This challenge is further amplified in inductive KGC settings\. In scenarios involving emerging entities, the few\-shot and data\-scarce nature of the task means that noisy or misaligned modalities, especially visual or operational signals, may have a disproportionate impact on entity representations, distorting the representation space and ultimately degrading link prediction accuracy\.

Given the scarcity of evidence that constrains inductive KGC, LLMs provide a promising way to enrich sparse relational contexts in low\-resource settings, owing to their encoded world knowledge and strong generalization ability\. Existing methods have explored the use of LLMs to extract relational context for KGC reasoning\(Wanget al\.[2024](https://arxiv.org/html/2607.26909#bib.bib19)\), but they do not explicitly consider modality\-specific reliability or granularity mismatch in MMKGs\. In MMKGs, the semantic priors derived from LLMs may also provide complementary high\-level cues for bridging heterogeneous modalities and reducing noisy signals\. However, directly combining LLM priors with multimodal information is still challenging\. Different modalities describe relational contexts at different levels of granularity, and their reliability may vary significantly under limited supervision\(Chenet al\.[2024](https://arxiv.org/html/2607.26909#bib.bib27)\)\. Therefore, the model should be able to selectively assess which signals are more trustworthy for a given prediction\. This motivates a principled adaptive integration mechanism that can align LLM semantic priors with multimodal cues and improve inductive link prediction for emerging entities under few\-shot and zero\-shot conditions\.

To address these challenges, we propose DuPLeR, a novel framework for inductive multimodal KGC in few\-shot and zero\-shot settings\. First, we leverage LLMs for relation semantic enrichment to alleviate resource scarcity in few\-shot scenarios\. To mitigate model hallucination, we design a topology refinement module that utilizes factual knowledge to denoise and enrich the LLM\-expanded information, thereby generating a calibrated relation graph for downstream reasoning\. Subsequently, DuPLeR introduces a dual\-pathway multimodal enhancement module, which leverages query\-relevant and query\-irrelevant multimodal information to inject multimodal cues during and after the message passing process, respectively, thereby selectively utilizing multimodal information to enhance entity representations\. Finally, DuPLeR performs dual\-level structural reasoning to simultaneously capture inter\-relation patterns and entity\-level structural evidence\. We conduct extensive experiments on two public MMKG benchmarks, and the results demonstrate that DuPLeR effectively combines multimodal LLM\-derived relational priors with multimodal signals to achieve more precise link prediction\. In summary, our paper makes the following contributions:

- •We propose DuPLeR, which integrates multimodal LLM\-derived relational priors with structural reasoning to mitigate representation distortion caused by sparse and noisy supervision, thereby improving inductive KGC\.
- •We design a dual\-pathway multimodal enhancement mechanism that injects query\-relevant multimodal evidence during message passing and incorporates query\-independent multimodal semantics after propagation, enabling more robust multimodal fusion\.
- •Extensive experiments on eight variants of two MMKG benchmarks show that DuPLeR consistently improves inductive generalization and robustness\.

## 2Related Work

### 2\.1Inductive KGC

Inductive KGC aims to generalize link prediction to unseen entities and relations by learning transferable structural patterns\. NBFNet captures relation\-aware paths through graph message passing\(Zhuet al\.[2021](https://arxiv.org/html/2607.26909#bib.bib22)\), while InGram and ULTRA construct relation graphs to model transferable inter\-relation dependencies\(Leeet al\.[2023](https://arxiv.org/html/2607.26909#bib.bib20); Galkinet al\.[2023](https://arxiv.org/html/2607.26909#bib.bib18)\)\. Recent methods further introduce semantic knowledge to alleviate sparse structural evidence\. ProLINK uses LLM\-derived type priors to augment relation graphs for low\-resource reasoning\(Wanget al\.[2024](https://arxiv.org/html/2607.26909#bib.bib19)\), whereas SEMMA combines LLM\-enriched relation semantics with structural relation graphs in fully inductive settings\(Arunet al\.[2025](https://arxiv.org/html/2607.26909#bib.bib6)\)\. However, these methods mainly focus on relation\-level textual semantics\. DuPLeR instead constructs a factually calibrated relation topology from multimodal type priors and further incorporates entity\-level multimodal evidence\.

### 2\.2Multimodal KGC

Multimodal KGC improves entity representations by combining structural, textual, and visual information\. LAFA jointly models link\-aware modality fusion and neighborhood aggregation\(Shanget al\.[2024](https://arxiv.org/html/2607.26909#bib.bib4)\), while MCKGC captures heterogeneous multimodal patterns in mixed\-curvature spaces\(Gaoet al\.[2025](https://arxiv.org/html/2607.26909#bib.bib8)\)\. Recent approaches increasingly consider modality\-specific reliability: NativE adopts relation\-guided adaptive fusion to address modality imbalance\(Zhanget al\.[2024b](https://arxiv.org/html/2607.26909#bib.bib7)\), and MMSN combines multimodal neighborhood aggregation with meta\-learning for few\-shot KGC\(Weiet al\.[2024](https://arxiv.org/html/2607.26909#bib.bib1)\)\. In particular, multimodal LLMs are prone to being overly influenced by a single modality during content generation\(Zhuet al\.[2026b](https://arxiv.org/html/2607.26909#bib.bib9)\)\. Different from representation\-level fusion, DuPLeR instead separates multimodal integration into query\-conditioned regulation during structural message passing and holistic supplementation after graph propagation\.

## 3Problem Definition

We study few\-shot multimodal KGC under inductive settings\. Let𝒢=\(ℰ,ℛ,𝒮\)\\mathcal\{G\}=\(\\mathcal\{E\},\\mathcal\{R\},\\mathcal\{S\}\)denote an MMKG, where𝒮⊆ℰ×ℛ×ℰ\\mathcal\{S\}\\subseteq\\mathcal\{E\}\\times\\mathcal\{R\}\\times\\mathcal\{E\}is the set of observed triples\. Each entitye∈ℰe\\in\\mathcal\{E\}is associated with multimodal side information, including a textual descriptionttand a set of images\{vk\}k=1K\\\{v\_\{k\}\\\}\_\{k=1\}^\{K\}\. It is worth noting that under the settings of this paper, entities and relations appearing at test time are disjoint from those seen during training\. For a target relationr∈ℛr\\in\\mathcal\{R\}, only a small support set𝒮r=\{\(hi,r,ti\)\}i=1N\\mathcal\{S\}\_\{r\}=\\\{\(h\_\{i\},r,t\_\{i\}\)\\\}\_\{i=1\}^\{N\}withN≪\|𝒮\|N\\ll\|\\mathcal\{S\}\|\(orN=0N=0\) is available\. Given a query request\(q,r,?\)\(q,r,?\), the goal is to retrieve the appropriate tail entity by leveraging limited support triples\.

![Refer to caption](https://arxiv.org/html/2607.26909v1/x1.png)Figure 1:The overall framework of DuPLeR\. It leverages multimodal LLM\-derived relational priors for structural reasoning and employs a dual\-path mechanism to inject multimodal signals, ultimately refining entity representations for link prediction\.
## 4Methodology

In this section, we propose DuPLeR \(Figure[1](https://arxiv.org/html/2607.26909#S3.F1)\), a framework that leverages multimodal LLM\-derived priors and entity\-level multimodal features to improve KGC under scarce triplet resources\. Given a query relation, DuPLeR constructs a relation\-aware context from sparse KGs and captures transferable relation–entity patterns through dual\-level structural reasoning \(Section[4\.1](https://arxiv.org/html/2607.26909#S4.SS1)\)\. This process is enhanced by multimodal LLM\-derived relational priors, which provide semantic cues for relation topology construction and calibration \(Section[4\.2](https://arxiv.org/html/2607.26909#S4.SS2)\)\. A dual\-pathway multimodal module then injects entity\-level multimodal information to refine representations and support robust entity matching \(Section[4\.3](https://arxiv.org/html/2607.26909#S4.SS3)\)\. Finally, DuPLeR is pretrained on multiple MMKGs and generalizes to unseen graphs without retraining \(Section[4\.4](https://arxiv.org/html/2607.26909#S4.SS4)\)\.

### 4\.1Dual\-level Structural Reasoning

In few\-shot scenarios, each observed triple provides critical structural evidence\. To effectively exploit such limited supervision, we perform dual\-level structural reasoning over relations and entities\. At the relation level, context\-enhanced co\-occurrence relation graph encoder \(CCoRGE\) constructs a high\-order relation graph and performs graph propagation to capture transferable interaction patterns among relations\. At the entity level, multimodal\-aware prediction network \(MAPNet\) uses the resulting relation embeddings to guide message passing over the KG, enabling the model to identify relation\-aware structural paths and score candidate entities for link prediction\. The two reasoning levels are complementary: the relation graph captures transferable interaction patterns that are less dependent on specific entity identities, whereas the entity graph grounds these patterns in query\-specific factual paths\.

#### CCoRGE

The high\-order relation graph is denoted as𝒢R=\(𝒱R,ℰR\)\\mathcal\{G\}\_\{R\}=\(\\mathcal\{V\}\_\{R\},\\mathcal\{E\}\_\{R\}\), where the node set𝒱R\\mathcal\{V\}\_\{R\}corresponds to the set of relationsℛ\\mathcal\{R\}in𝒢\\mathcal\{G\}\. The edges in𝒢R\\mathcal\{G\}\_\{R\}capture the topological co\-occurrence regularities between relations\. We define four distinct edge types,𝒯=\{h​\-​h,h​\-​t,t​\-​h,t​\-​t\}\\mathcal\{T\}=\\\{h\\text\{\-\}h,h\\text\{\-\}t,t\\text\{\-\}h,t\\text\{\-\}t\\\}, representing how two relations interact via shared entities\. Formally, fora,b∈\{h,t\}a,b\\in\\\{h,t\\\}, a directed edge\(ri,rj,a​\-​b\)∈ℰR\(r\_\{i\},r\_\{j\},a\\text\{\-\}b\)\\in\\mathcal\{E\}\_\{R\}exists if the same entity appears at positionaain a triple involvingrir\_\{i\}and at positionbbin a triple involvingrjr\_\{j\}\. For example, an edge\(ri,rj\)\(r\_\{i\},r\_\{j\}\)of typeh​\-​t\\ h\\text\{\-\}texists if there exists an entitye∈ℰe\\in\\mathcal\{E\}such that:

\(e,ri,xi\)∈𝒮∧\(xj,rj,e\)∈𝒮\.\(e,r\_\{i\},x\_\{i\}\)\\in\\mathcal\{S\}\\ \\land\\ \(x\_\{j\},r\_\{j\},e\)\\in\\mathcal\{S\}\.\(1\)During pre\-training,𝒢R\\mathcal\{G\}\_\{R\}is constructed from𝒢\\mathcal\{G\}\. At reasoning time, it is augmented with semantic connections derived from the LLM\-predicted head\- and tail\-entity types, as detailed in Section[4\.2](https://arxiv.org/html/2607.26909#S4.SS2)\. Finally, multi\-layer GNN is employed to learn relation representations over𝒢R\\mathcal\{G\}\_\{R\}\.

#### MAPNet

To ground transferable relation patterns in query\-specific evidence while avoiding indiscriminate multimodal fusion, MAPNet performs relation\-aware propagation over the entity graph and incorporates the two multimodal representations at complementary stages\. During entity encoding, the initial representation of the source entity is obtained by fusing the CCoRGE representation of the query relation with the relation\-aware multimodal representation of the source entity\. All other entity states are initialized to zero\. The CCoRGE representations of all relations are further projected into layer\-specific edge features for entity propagation\. Formally, at thell\-th layer, MAPNet constructs a relation\-aware message for each neighborv∈𝒩​\(e\)v\\in\\mathcal\{N\}\(e\)by concatenating its previous\-layer representation𝐡v\(l−1\)\\mathbf\{h\}\_\{v\}^\{\(l\-1\)\}with the corresponding relation embedding𝐫v,e\(l\)\\mathbf\{r\}\_\{v,e\}^\{\(l\)\}, followed by a linear transformation parameterized by𝐖M\(l\)\\mathbf\{W\}\_\{M\}^\{\(l\)\}\. These messages are summed over all neighbors to obtain the neighborhood representation𝐚e\(l\)\\mathbf\{a\}\_\{e\}^\{\(l\)\}\. The previous\-layer state𝐡e\(l−1\)\\mathbf\{h\}\_\{e\}^\{\(l\-1\)\}and the aggregated neighborhood representation𝐚e\(l\)\\mathbf\{a\}\_\{e\}^\{\(l\)\}are first structurally fused, after which the resulting representation is modulated by the multimodal context𝐡mult\(e\)\\mathbf\{h\}\_\{\\mathrm\{mult\}\}^\{\(e\)\}:

𝐡e\(l\)=ϕmult\(l\)​\(ϕstr\(l\)​\(𝐡e\(l−1\),𝐚e\(l\)\),𝐡mult\(e\)\),\\mathbf\{h\}\_\{e\}^\{\(l\)\}=\\phi\_\{\\mathrm\{mult\}\}^\{\(l\)\}\\left\(\\phi\_\{\\mathrm\{str\}\}^\{\(l\)\}\\left\(\\mathbf\{h\}\_\{e\}^\{\(l\-1\)\},\\mathbf\{a\}\_\{e\}^\{\(l\)\}\\right\),\\mathbf\{h\}\_\{\\mathrm\{mult\}\}^\{\(e\)\}\\right\),\(2\)whereϕstr\(l\)​\(⋅\)\\phi\_\{\\mathrm\{str\}\}^\{\(l\)\}\(\\cdot\)andϕmult\(l\)​\(⋅\)\\phi\_\{\\mathrm\{mult\}\}^\{\(l\)\}\(\\cdot\)denote the structural fusion and multimodal residual modulation at thell\-th layer, respectively\. AfterLLlayers of structural propagation, the final scores​\(e\)s\(e\)of entityeeis computed by concatenating the hidden states from all GNN layers with the query\-agnostic multimodal embedding𝐳out\(e\)\\mathbf\{z\}\_\{\\mathrm\{out\}\}^\{\(e\)\}:

s​\(e\)=MLP⁡\(\(⨁l=0L𝐡e\(l\)\)∥𝐳out\(e\)\),s\(e\)=\\operatorname\{MLP\}\\left\(\\left\(\\bigoplus\\nolimits\_\{l=0\}^\{L\}\\mathbf\{h\}\_\{e\}^\{\(l\)\}\\right\)\\parallel\\mathbf\{z\}\_\{\\mathrm\{out\}\}^\{\(e\)\}\\right\),\(3\)where⨁\\bigoplusdenotes feature\-wise concatenation across layers\. The detailed constructions of𝐡mult\(e\)\\mathbf\{h\}\_\{\\mathrm\{mult\}\}^\{\(e\)\}and𝐳out\(e\)\\mathbf\{z\}\_\{\\mathrm\{out\}\}^\{\(e\)\}are provided in Section[4\.3](https://arxiv.org/html/2607.26909#S4.SS3)\. Finally, we optimize the model using the standard binary cross\-entropy \(BCE\) loss; its detailed formulation is provided in Appendix[A](https://arxiv.org/html/2607.26909#A1)\.

### 4\.2Prior Knowledge Injection

#### Multimodal LLM\-based Type Predictor

Relation modeling is pivotal to KGC\. Under few\-shot settings, the scarcity of relation\-associated triples makes it difficult to learn reliable relation representations\(Zhanget al\.[2019](https://arxiv.org/html/2607.26909#bib.bib28)\)\. This challenge becomes more pronounced in zero\-shot scenarios, where conventional methods are often unable to infer meaningful representations for unseen relations without support triples\(Chenet al\.[2023](https://arxiv.org/html/2607.26909#bib.bib29)\)\. To address this issue, we leverage carefully constructed prompts—augmented by sparse relational information—to infer the plausible types of head and tail entities, thereby enriching the relational graph and obtaining reliable relational representations\. The detailed prompts and output schema used for multimodal LLMs are illustrated in Figure[2](https://arxiv.org/html/2607.26909#S4.F2)\. Specifically, entity images are treated as optional evidence rather than uniformly reliable cues\. Since images may contain abundant details that are irrelevant to the target relation, the prompt instructs the multimodal LLM to use visual information only when it is relevant to the relation and consistent with the accompanying textual context, and to otherwise ignore it, thereby reducing visual noise\. For pure\-text LLMs, the visual inputs are omitted while the remaining prompt structure is retained\. To regularize type inference, we constrain the outputs using a generic entity type list under three progressively relaxed settings: Mandatory, Guided, and Unrestricted\. For each relation, the model returns a list of plausible types for the head entities and another list for the tail entities\. We further provide relational context in three formats—Relation\-only, Triple\-only, and Fusion—to examine the trade\-off between type inference quality and prompting overhead\. The predicted head\- and tail\-type sets of each relation are used to construct𝒢sem\\mathcal\{G\}\_\{\\text\{sem\}\}based on type\-set overlap, following the positional connectivity scheme of CCoRGE \(Section[4\.1](https://arxiv.org/html/2607.26909#S4.SS1.SSSx1)\)\. Overall, this prompting strategy aims to enrich sparse relational contexts and provide reliable semantic priors for downstream inductive KGC\.

![Refer to caption](https://arxiv.org/html/2607.26909v1/x2.png)Figure 2:Multimodal LLM Prompts
#### Relational Topology Refinement

To mitigate hallucinations in multimodal LLMs, we introduce a factual\-triple\-based refinement step that filters potentially spurious outputs and enriches them with reliable information\. We begin by constructing a structural relation graph𝒢ins\\mathcal\{G\}\_\{\\text\{ins\}\}for the target relationrqr\_\{q\}based on the support triples and the background KG, following Section[4\.1](https://arxiv.org/html/2607.26909#S4.SS1.SSSx1)\. We then merge𝒢sem\\mathcal\{G\}\_\{\\text\{sem\}\}and𝒢ins\\mathcal\{G\}\_\{\\text\{ins\}\}to obtain the preliminary relation graph𝒢ori\\mathcal\{G\}\_\{\\text\{ori\}\}\. To reduce noisy relation neighbors introduced during graph augmentation, we apply an indirect noise filtering strategy to the first\-order neighbors ofrqr\_\{q\}in𝒢ori\\mathcal\{G\}\_\{\\text\{ori\}\}\. For each neighborv∈𝒩ori​\(rq\)v\\in\\mathcal\{N\}\_\{\\text\{ori\}\}\(r\_\{q\}\), we measure its noise degree by comparing its non\-target connections in the merged graph and the structural graph:

n​\(v\)=\|E¯ori​\(v\)\|−\|E¯ins​\(v\)\|,n\(v\)=\\left\|\\bar\{E\}\_\{\\text\{ori\}\}\(v\)\\right\|\-\\left\|\\bar\{E\}\_\{\\text\{ins\}\}\(v\)\\right\|,\(4\)whereE¯ori​\(v\)\\bar\{E\}\_\{\\text\{ori\}\}\(v\)andE¯ins​\(v\)\\bar\{E\}\_\{\\text\{ins\}\}\(v\)denote the sets of edges incident tovvin𝒢ori\\mathcal\{G\}\_\{\\text\{ori\}\}and𝒢ins\\mathcal\{G\}\_\{\\text\{ins\}\}, respectively, excluding the edge connectingvvto the target relationrqr\_\{q\}\. Given a predefined noise thresholdτ\\tau, the edges connectingrqr\_\{q\}to neighbors withn​\(v\)≥τn\(v\)\\geq\\tauare collected into the noisy\-edge set:

ℰnoise=\{\(rq,v\)∈ℰori∣v∈𝒩ori​\(rq\),n​\(v\)≥τ\}\.\\mathcal\{E\}\_\{\\mathrm\{noise\}\}=\\left\\\{\(r\_\{q\},v\)\\in\\mathcal\{E\}\_\{\\mathrm\{ori\}\}\\mid v\\in\\mathcal\{N\}\_\{\\mathrm\{ori\}\}\(r\_\{q\}\),\\;n\(v\)\\geq\\tau\\right\\\}\.\(5\)The final denoised relation graph𝒢R\\mathcal\{G\}\_\{R\}is obtained by removing these edges from the merged relation graph:

𝒢R=\(𝒱ori,ℰori∖ℰnoise\)\.\\mathcal\{G\}\_\{R\}=\\left\(\\mathcal\{V\}\_\{\\mathrm\{ori\}\},\\mathcal\{E\}\_\{\\mathrm\{ori\}\}\\setminus\\mathcal\{E\}\_\{\\mathrm\{noise\}\}\\right\)\.\(6\)A largerτ\\tauretains more auxiliary semantic connections, whereas a smallerτ\\tauresults in more aggressive noise suppression\.

### 4\.3Dual\-pathway Multimodal Enhancement

While relation graphs inferred by multimodal LLMs offer rich relational context, they remain unreliable as standalone evidence and must be anchored by structural message passing\(Zhuet al\.[2026a](https://arxiv.org/html/2607.26909#bib.bib10)\)\. To complement this LLM\-derived relational context with grounded entity\-level semantics, multimodal information is essential; however, indiscriminate fusion may introduce substantial query\-irrelevant noise and distort structural message passing\(Zhanget al\.[2024a](https://arxiv.org/html/2607.26909#bib.bib44); Chenet al\.[2024](https://arxiv.org/html/2607.26909#bib.bib27)\)\. Injecting query\-specific multimodal information during message passing can suppress modality cues irrelevant to the current relation, but such relation\-conditioned filtering may discard global semantics useful for entity identification\. Conversely, query\-agnostic fusion preserves comprehensive entity\-level semantics, yet directly introducing these unfiltered features into message passing may interfere with relation\-specific structural paths\. Neither pathway alone can simultaneously maintain relation\-aware structural reasoning and preserve complete entity\-level semantics\. Therefore, DuPLeR employs two complementary pathways: Query\-Specific Regulation provides controlled multimodal guidance during structural propagation, while Query\-Agnostic LMF\-Max compensates for the semantic loss induced by relation\-conditioned filtering by restoring comprehensive entity\-level semantics after propagation\.

#### Query\-Specific Regulation

To instantiate the relation\-aware regulation, we dynamically fuse text and visual features conditioned on the query relation\. After projecting normalized features into a shared space, we obtain a unified visual representation𝐡vis\\mathbf\{h\}\_\{\\text\{vis\}\}via relation\-guided low\-rank attention\. Thekk\-th image attention weight is:

αk=softmaxk⁡\(\(𝐖q​𝐪\)⊤​\(𝐖v​𝐟vk\)\),\\alpha\_\{k\}=\\operatorname\{softmax\}\_\{k\}\\left\(\(\\mathbf\{W\}\_\{q\}\\mathbf\{q\}\)^\{\\top\}\(\\mathbf\{W\}\_\{v\}\\mathbf\{f\}\_\{v\_\{k\}\}\)\\right\),\(7\)where𝐖q\\mathbf\{W\}\_\{q\}and𝐖v\\mathbf\{W\}\_\{v\}are low\-rank projection matrices\. To precisely control modality contributions, we deploy modality\-specific low\-rank gates,𝐠t\\mathbf\{g\}\_\{t\}and𝐠v\\mathbf\{g\}\_\{v\}, which are parameterized by the query relation:

𝐡fused=𝐡t⊙𝐠t\+𝐡vis⊙𝐠v\+𝐡t\.\\mathbf\{h\}\_\{\\text\{fused\}\}=\\mathbf\{h\}\_\{t\}\\odot\\mathbf\{g\}\_\{t\}\+\\mathbf\{h\}\_\{\\text\{vis\}\}\\odot\\mathbf\{g\}\_\{v\}\+\\mathbf\{h\}\_\{t\}\.\(8\)Unlike standard gating mechanisms, our relation\-aware gating coupled with a residual text connection is specifically designed to alleviate modality imbalance\. We then apply layer normalization and dropout to𝐡fused\\mathbf\{h\}\_\{\\text\{fused\}\}to obtain the relation\-guided multimodal embedding𝐡mult\\mathbf\{h\}\_\{\\text\{mult\}\}\. Rather than directly appending multimodal features, which may dilute graph\-topological information, we use𝐡mult\\mathbf\{h\}\_\{\\text\{mult\}\}as a soft regulator to modulate the structurally updated entity representation at each layer, as formulated in Equation \([2](https://arxiv.org/html/2607.26909#S4.E2)\)\. This customized modulation ensures that structural reasoning remains dominant while being smoothly guided by query\-relevant multimodal cues\. A formal proof is provided in the supplementary material\.

#### Query\-Agnostic LMF\-Max

To compensate for the aforementioned information bottleneck, our post\-fusion pathway captures salient visual evidence while excluding invalid image inputs\. Specifically, invalid image positions indicated by the validity mask are filled with a large negative value before dimension\-wise max pooling, while entities with no valid images are assigned a zero visual representation, yielding𝐟~v\\tilde\{\\mathbf\{f\}\}\_\{v\}\. We then project the textual feature and𝐟~v\\tilde\{\\mathbf\{f\}\}\_\{v\}into a low\-rank space, obtaining𝐓,𝐕∈ℝR×d\\mathbf\{T\},\\mathbf\{V\}\\in\\mathbb\{R\}^\{R\\times d\}\. Unlike vanilla LMF, which treats all low\-rank components equally during aggregation, LMF\-Max regards each rank as a distinct semantic subspace and assigns it a learnable weight𝐰∈ℝR\\mathbf\{w\}\\in\\mathbb\{R\}^\{R\}:

𝐳=∑r=1Rwr​\(𝐓​\[r\]⊙𝐕​\[r\]\)\+𝐛\.\\mathbf\{z\}=\\sum\\nolimits\_\{r=1\}^\{R\}w\_\{r\}\\left\(\\mathbf\{T\}\[r\]\\odot\\mathbf\{V\}\[r\]\\right\)\+\\mathbf\{b\}\.\(9\)Masked max pooling preserves the most salient valid visual signals, while rank\-wise weighting selectively emphasizes informative cross\-modal interactions\. This pathway remains query\-agnostic, providing complementary entity\-level semantics that are incorporated into the final prediction as formulated in Equation \([3](https://arxiv.org/html/2607.26909#S4.E3)\)\.

### 4\.4Pre\-training and Inference

During pre\-training, DuPLeR is jointly optimized across multiple multimodal knowledge graphs, with all trainable components fully shared across graphs\. In addition, to balance the varying difficulty levels across graphs, DuPLeR adopts a loss\-guided dynamic dataset weighting strategy, as detailed in Appendix[B](https://arxiv.org/html/2607.26909#A2)\.

During inference, DuPLeR derives representations of unseen relations based on the outputs of multimodal LLMs\. These relation representations are subsequently injected into the knowledge graph to obtain representations of unseen entities\. Consequently, DuPLeR requires no retraining on the target graph, thereby enabling more precise and context\-aware link prediction under data\-scarce conditions\.

## 5Experiment

Table 1:Few\-shot Hits@10 results \(%\) on FB15K\-IMG\-R and OpenBG\-IMG\-R, where 1s and 3s denote the 1\-shot and 3\-shot settings\. Bold and underline indicate the best overall and best baseline results, respectively\.### 5\.1Experiment Settings

#### Datasets

To evaluate DuPLeR, we use five MMKG datasets: FB15K\-IMG\-R, DB15K\-EMB\-R, and YAGO15K\-EMB\-R derived from MMKB\(García\-Durán and Niepert[2018](https://arxiv.org/html/2607.26909#bib.bib11); Oñoro\-Rubioet al\.[2019](https://arxiv.org/html/2607.26909#bib.bib12)\); OpenBG\-IMG\-R from the OpenBG benchmarks\(Denget al\.[2023](https://arxiv.org/html/2607.26909#bib.bib14)\); and CoDEx\-IMG\-R, which we construct from the publicly available CoDEx dataset\(Safavi and Koutra[2020](https://arxiv.org/html/2607.26909#bib.bib13)\)\. Here, the suffix “EMB” denotes datasets in which the original images are represented by pre\-extracted visual embeddings, whereas “IMG” denotes datasets that provide the raw image files\.

DB15K\-EMB\-R, YAGO15K\-EMB\-R, and CoDEx\-IMG\-R are used for pre\-training with standard transductive splits\. FB15K\-IMG\-R and OpenBG\-IMG\-R are used for few\-shot link prediction\. Following InGram\(Leeet al\.[2023](https://arxiv.org/html/2607.26909#bib.bib20)\), we create four variants for each dataset according to the proportions of triples involving new relations in the inference graph, with ratios of 0\.25, 0\.50, 0\.75, and 1\.00\. For each variant, three random seeds are used to generate 0\-shot, 1\-shot, and 3\-shot settings, resulting in 72 few\-shot datasets for evaluation\.

#### Baselines

To evaluate DuPLeR, we compare it with three groups of baselines: \(1\)Traditional KG Embedding methods: DistMult\(Yanget al\.[2014](https://arxiv.org/html/2607.26909#bib.bib16)\), which uses diagonal relation matrices for bilinear scoring, and RotatE\(Sunet al\.[2019](https://arxiv.org/html/2607.26909#bib.bib17)\), which represents relations as rotations in complex space\. \(2\)Unimodal models: InGram\(Leeet al\.[2023](https://arxiv.org/html/2607.26909#bib.bib20)\), which learns relation and entity representations by aggregating neighborhood information over relation\-weighted KG structures, and ULTRA\(Galkinet al\.[2023](https://arxiv.org/html/2607.26909#bib.bib18)\), which builds relation graphs from triples to capture inter\-relation dependencies, and ProLINK\(Wanget al\.[2024](https://arxiv.org/html/2607.26909#bib.bib19)\), which extends ULTRA with LLM\-derived priors for few\-shot reasoning\. \(3\)Multimodal models: MMSN\(Weiet al\.[2024](https://arxiv.org/html/2607.26909#bib.bib1)\), which combines multimodal neighbor aggregation with meta\-learning\-based cross\-modal interaction modeling\.

#### Experimental Settings

In the comparative experiments, we followed the original settings of the baseline models and adapted them for the few\-shot dataset setup\. DuPLeR is implemented with a three\-layer DistMult\-based GNN for relation encoding and a six\-layer RotatE\-based GNN for entity encoding\. The low\-rank fusion rank, bottleneck dimension, and embedding dimension are set to 2, 64, and 64, respectively\. The model is trained with AdamW using an initial learning rate of 0\.003 and a cosine annealing scheduler\. For evaluation, we use three LLMs, Llama2\-7B\(Touvron and others[2023](https://arxiv.org/html/2607.26909#bib.bib26)\), Qwen3\-8B\(Team[2025](https://arxiv.org/html/2607.26909#bib.bib23)\), and DeepSeek\-V3\.2\(DeepSeek\-AI[2025](https://arxiv.org/html/2607.26909#bib.bib24)\), along with the VLM Qwen3\-VL\-8B\-Instruct\(Team[2025](https://arxiv.org/html/2607.26909#bib.bib23)\), abbreviated as Qwen3\-VL\-8B in subsequent tables\. We use Hits@10 as the evaluation metric, which measures the proportion of test queries whose ground\-truth tail entity appears among the top 10 predicted entities\.

### 5\.2Main Results

We conducted experiments under the aforementioned settings, and the results are presented in Table[1](https://arxiv.org/html/2607.26909#S5.T1)\. It should be noted that, for traditional KG Embedding models, we employed the OpenKE toolkit\(Hanet al\.[2018](https://arxiv.org/html/2607.26909#bib.bib21)\)\. All results reported in the table are averaged over three runs with different random seeds on the few\-shot datasets\.

The following conclusions can be drawn from the table: \(1\) Superiority of LLM Integration: ProLINK and DuPLeR significantly outperform traditional baselines across most settings\. For instance, on FB15K\-IMG\-R v1, DuPLeR achieves 71\.77% compared to ULTRA’s 27%, validating that LLM\-injected prior knowledge effectively enriches sparse context for KGC\. \(2\) Effectiveness of Dual\-Path Enhancement: DuPLeR generally surpasses ProLINK, demonstrating that the dual\-path multimodal enhancement module successfully leverages entity\-level multimodal information to complement the relational context provided by LLMs\. \(3\) Model robustness and backbone diversity: Peak performance varies across different LLM backbones \(e\.g\., DeepSeek\-V3\.2 leads in only specific cases\)\. This diversity reflects differences in pre\-training corpora and highlights DuPLeR’s robustness; it maintains high Hits@10 performance via multimodal cues even when LLM\-derived relational context is insufficient, thereby validating the effectiveness of the dual\-path multimodal enhancement module\.

### 5\.3Zero\-Shot Performance

The motivation behind incorporating multimodal LLMs is to alleviate information scarcity and provide high\-quality relational representations—an effect that becomes particularly pronounced in a zero\-shot setting\. To demonstrate the generalizability of these representations across broad world knowledge, we evaluated the zero\-shot performance on FB15K\-IMG\-R\. The results comparing several baseline models alongside DuPLeR are presented in Table[2](https://arxiv.org/html/2607.26909#S5.T2)\. It is worth noting that, under this specific setting, the prompts provided to all models consisted solely of textual information containing the relation descriptions\.

Table 2:0\-shot evaluation results on FB15K\-IMG\-R \(%\)\. Bold numbers indicate the best overall performance; underlined numbers indicate the best among baseline models\.The zero\-shot experimental results corroborate the conclusions of the main experiments—namely, that incorporating LLM prior knowledge effectively mitigates the problem of information sparsity\. Notably, DuPLeR shows a clear improvement over ProLINK, which also utilizes LLMs, with the enhancement being particularly pronounced for the relatively weaker Llama2\-7B model\. Furthermore, in the v1 version, the performance achieved using the compact Qwen3\-8B model surpassed that of the full\-scale DeepSeek\-V3\.2 model\. This suggests that DuPLeR’s dual\-path multimodal information injection module effectively complements the relational context enriched by LLMs, thereby further mitigating the issue of information scarcity within a zero\-shot setting and enabling more precise predictions\.

### 5\.4Ablation Study

Table 3:Ablation evaluation results on the OpenBG\-IMG\-R V1 dataset \(%\) with Qwen3\-VL\-8B\.Constrained by page limits, we conduct our ablation studies exclusively on OpenBG\-IMG\-R V1 using Hits@10 under 0\-shot, 1\-shot, and 3\-shot settings\. Three DuPLeR variants are evaluated: w/o multimodal LLM, which removes multimodal LLM prompting during evaluation; w/o Message Passing, which disables multimodal process injection during pre\-training; and w/o LMF\-Max, which disables multimodal post\-injection during pre\-training\. These variants examine the contribution of large\-model relational knowledge and multimodal injection strategies\. As shown in Table[3](https://arxiv.org/html/2607.26909#S5.T3), all modules contribute to DuPLeR\. Removing the multimodal LLM prior causes a∼28\.2%\\sim 28\.2\\%relative performance drop in zero\-shot setting, highlighting its vital role in enriching relational context without reference examples\. This drop diminishes in few\-shot scenarios, where explicit empirical demonstrations partially replace the LLM’s semantic prior, leading to a diminishing marginal effect\. Notably, this degradation is less severe than ablating the Dual\-pathway Multimodal Enhancement module\. This suggests robust multimodal entity representations can partially offset limited relational context, underscoring the dual\-path injection’s importance in capturing multimodal signals\.

### 5\.5Hyperparameter Analysis

To systematically evaluate the impact of hyperparameter variations on DuPLeR, we analyze four architecture\-related hyperparameters—bottleneck dimension, low\-rank decomposition rank, CCoRGE depth, and MAPNet depth—as well as the training hyperparameter learning rate\. In our analysis, we select the FB15K\-IMG\-R V1 and OpenBG\-IMG\-R V1 datasets and report performance under the 1\-shot setting using the Hits@10 metric, as 1\-shot is more sensitive compared to 0\-shot and 3\-shot\. The results are shown in Figure[3](https://arxiv.org/html/2607.26909#S5.F3)\.

![Refer to caption](https://arxiv.org/html/2607.26909v1/x3.png)Figure 3:Hyperparameter sensitivity analysis\. We analyze 1\-shot Hits@10 performance on FB15K\-IMG\-R V1 and OpenBG\-IMG\-R V1 while varying one hyperparameter at a time \(low\-rank rank, bottleneck dimension, CCoRGE depth, MAPNet depth, learning rate\)\.As shown in the figure, adjusting the learning rate has minimal impact on FB15K but substantially improves OpenBG performance, indicating that a larger learning rate can enhance generalization\. Increasing the bottleneck dimension also improves performance, while too small a dimension \(e\.g\., 16\) excessively compresses multimodal semantics, weakening relation\-aware gating and structural reasoning\. The rank of low\-rank interaction affects few\-shot generalization: lower rank constrains cross\-modal interactions and introduces stronger structural bias, whereas excessively large rank may cause overfitting\. CCoRGE depth is critical for modeling relational context; shallow networks fail to capture sufficient context, while overly deep networks may overfit on simpler datasets\. MAPNet depth has a smaller but consistently positive effect, expanding the structural receptive field and enabling richer path reasoning for more precise link predictions\.

## 6Conclusion

This paper proposes DuPLeR, a novel framework for few\-shot MMKG completion\. DuPLeR combines multimodal LLM\-derived relational priors, topology\-refined structural reasoning, and dual\-path multimodal enhancement to improve link prediction under data\-scarce settings\. Extensive experiments demonstrate that LLM\-derived priors provide useful relational context, while relation\-aware process injection and LMF\-Max post fusion further enhance entity representations\. Future work will explore higher\-quality MMKG construction and more domain\-adaptive multimodal LLM selection and prompting strategies\.

## References

- SEMMA: a semantic aware knowledge graph foundation model\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 31813–31836\.Cited by:[§2\.1](https://arxiv.org/html/2607.26909#S2.SS1.p1.1)\.
- M\. Chen, W\. Zhang, Y\. Geng, Z\. Xu, J\. Z\. Pan, and H\. Chen \(2023\)Generalizing to unseen elements: a survey on knowledge extrapolation for knowledge graphs\.InProceedings of the Thirty\-Second International Joint Conference on Artificial Intelligence, IJCAI\-23,E\. Elkind \(Ed\.\),pp\. 6574–6582\.Note:Survey TrackExternal Links:[Document](https://dx.doi.org/10.24963/ijcai.2023/737)Cited by:[§4\.2](https://arxiv.org/html/2607.26909#S4.SS2.SSSx1.p1.1)\.
- X\. Chen, N\. Zhang, L\. Li, S\. Deng, C\. Tan, C\. Xu, F\. Huang, L\. Si, and H\. Chen \(2022\)Hybrid transformer with multi\-level fusion for multimodal knowledge graph completion\.InSIGIR ’22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid, Spain, July 11 \- 15, 2022,pp\. 904–915\.External Links:[Document](https://dx.doi.org/10.1145/3477495.3531992)Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p3.1)\.
- Z\. Chen, Y\. Zhang, Y\. Fang, Y\. Geng, L\. Guo, X\. Chen, Q\. Li, W\. Zhang, J\. Chen, Y\. Zhu, J\. Li, X\. Liu, J\. Z\. Pan, N\. Zhang, and H\. Chen \(2024\)Knowledge graphs meet multi\-modal learning: A comprehensive survey\.CoRRabs/2402\.05391\.Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p4.1),[§4\.3](https://arxiv.org/html/2607.26909#S4.SS3.p1.1)\.
- DeepSeek\-AI \(2025\)DeepSeek\-v3\.2: pushing the frontier of open large language models\.Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx3.p1.1)\.
- S\. Deng, C\. Wang, Z\. Li, N\. Zhang, Z\. Dai, H\. Chen, F\. Xiong, M\. Yan, Q\. Chen, M\. Chen, J\. Chen, J\. Z\. Pan, B\. Hooi, and H\. Chen \(2023\)Construction and applications of billion\-scale pre\-trained multimodal business knowledge graph\.InICDE,pp\.\.External Links:[Link](https://doi.org/10.48550/arXiv.2209.15214)Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx1.p1.1)\.
- M\. Galkin, X\. Yuan, H\. Mostafa, J\. Tang, and Z\. Zhu \(2023\)Towards foundation models for knowledge graph reasoning\.InInternational Conference on Learning Representations,Cited by:[§2\.1](https://arxiv.org/html/2607.26909#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx2.p1.1)\.
- Y\. Gao, F\. Zhang, Z\. Zhang, X\. Min, and F\. Zhuang \(2025\)Mixed\-curvature multi\-modal knowledge graph completion\.InProceedings of the Thirty\-Ninth AAAI Conference on Artificial Intelligence and Thirty\-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence,AAAI’25/IAAI’25/EAAI’25\.External Links:ISBN 978\-1\-57735\-897\-8,[Document](https://dx.doi.org/10.1609/aaai.v39i11.33273)Cited by:[§2\.2](https://arxiv.org/html/2607.26909#S2.SS2.p1.1)\.
- A\. García\-Durán and M\. Niepert \(2018\)KBlrn: end\-to\-end learning of knowledge base representations with latent, relational, and numerical features\.InProceedings of the Thirty\-Fourth Conference on Uncertainty in Artificial Intelligence, UAI,pp\. 372–381\.Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx1.p1.1)\.
- X\. Han, S\. Cao, L\. Xin, Y\. Lin, Z\. Liu, M\. Sun, and J\. Li \(2018\)OpenKE: an open toolkit for knowledge embedding\.InProceedings of EMNLP,Cited by:[§5\.2](https://arxiv.org/html/2607.26909#S5.SS2.p1.1)\.
- J\. Hu, H\. Yang, F\. Teng, S\. Du, and T\. Li \(2024\)A knowledge graph completion model based on triple level interaction and contrastive learning\.Pattern Recognition156,pp\. 110783\.External Links:ISSN 0031\-3203,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.patcog.2024.110783)Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p1.1)\.
- J\. Kim, G\. Lee, T\. Kim, and K\. Shin \(2025\)KGMEL: knowledge graph\-enhanced multimodal entity linking\.InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval,pp\. 3015–3019\.Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p3.1)\.
- J\. Lee, C\. Chung, and J\. J\. Whang \(2023\)InGram: inductive knowledge graph embedding via relation graphs\.InProceedings of the 40th International Conference on Machine Learning,pp\. 18796–18809\.Cited by:[§2\.1](https://arxiv.org/html/2607.26909#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx1.p2.1),[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx2.p1.1)\.
- D\. Oñoro\-Rubio, M\. Niepert, A\. García\-Durán, R\. Gonzalez\-Sanchez, and R\. J\. López\-Sastre \(2019\)Answering visual\-relational queries in web\-extracted knowledge graphs\.InAKBC,Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx1.p1.1)\.
- T\. Safavi and D\. Koutra \(2020\)CoDEx: a Comprehensive Knowledge Graph Completion Benchmark\.InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\),misc,pp\. 8328–8350\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.669)Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx1.p1.1)\.
- B\. Shang, Y\. Zhao, J\. Liu, and D\. Wang \(2024\)LAFA: multimodal knowledge graph completion with link aware fusion and aggregation\.38,pp\. 8957–8965\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/28744),[Document](https://dx.doi.org/10.1609/aaai.v38i8.28744)Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p3.1),[§2\.2](https://arxiv.org/html/2607.26909#S2.SS2.p1.1)\.
- Z\. Sun, Z\. Deng, J\. Nie, and J\. Tang \(2019\)RotatE: knowledge graph embedding by relational rotation in complex space\.InInternational Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=HkgEQnRqYQ)Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx2.p1.1)\.
- Q\. Team \(2025\)Qwen3 technical report\.External Links:2505\.09388,[Link](https://arxiv.org/abs/2505.09388)Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx3.p1.1)\.
- H\. Touvronet al\.\(2023\)Llama 2: open foundation and fine\-tuned chat models\.ArXivabs/2307\.09288\.External Links:[Link](https://api.semanticscholar.org/CorpusID:259950998)Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx3.p1.1)\.
- K\. Wang, Y\. Xu, Z\. Wu, and S\. Luo \(2024\)LLM as prompter: low\-resource inductive reasoning on arbitrary knowledge graphs\.InFindings of the Association for Computational Linguistics ACL 2024,Bangkok, Thailand and virtual meeting,pp\. 3742–3759\.External Links:[Link](https://aclanthology.org/2024.findings-acl.224)Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p4.1),[§2\.1](https://arxiv.org/html/2607.26909#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx2.p1.1)\.
- Y\. Wei, W\. Chen, X\. Zhang, P\. Zhao, J\. Qu, and L\. Zhao \(2024\)Multi\-modal siamese network for few\-shot knowledge graph completion\.In2024 IEEE 40th International Conference on Data Engineering \(ICDE\),pp\. 719–732\.Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p3.1),[§2\.2](https://arxiv.org/html/2607.26909#S2.SS2.p1.1),[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx2.p1.1)\.
- B\. Yang, W\. Yih, X\. He, J\. Gao, and L\. Deng \(2014\)Embedding entities and relations for learning and inference in knowledge bases\.InInternational Conference on Learning Representations,External Links:[Link](https://api.semanticscholar.org/CorpusID:2768038)Cited by:[§5\.1](https://arxiv.org/html/2607.26909#S5.SS1.SSSx2.p1.1)\.
- S\. Yang, J\. Yang, X\. Jiang, Y\. Gao, L\. T\. Yang, R\. Luo, and J\. Yang \(2025\)Towards multimodal inductive learning: adaptively embedding mmkg via prototypes\.InProceedings of the ACM on Web Conference 2025,pp\. 109–118\.Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p3.1)\.
- D\. Yuan, S\. Zhou, X\. Chen, D\. Wang, K\. Liang, X\. Liu, and J\. Huang \(2025\)Knowledge graph completion with relation\-aware anchor enhancement\.InProceedings of the Thirty\-Ninth AAAI Conference on Artificial Intelligence and Thirty\-Seventh Conference on Innovative Applications of Artificial Intelligence and Fifteenth Symposium on Educational Advances in Artificial Intelligence,AAAI’25/IAAI’25/EAAI’25\.External Links:ISBN 978\-1\-57735\-897\-8,[Link](https://doi.org/10.1609/aaai.v39i14.33672),[Document](https://dx.doi.org/10.1609/aaai.v39i14.33672)Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p1.1)\.
- C\. Zhang, H\. Yao, C\. Huang, M\. Jiang, Z\. Li, and N\. V\. Chawla \(2019\)Few\-shot knowledge graph completion\.External Links:1911\.11298,[Link](https://arxiv.org/abs/1911.11298)Cited by:[§4\.2](https://arxiv.org/html/2607.26909#S4.SS2.SSSx1.p1.1)\.
- Q\. Zhang, Y\. Wei, Z\. Han, H\. Fu, X\. Peng, C\. Deng, Q\. Hu, C\. Xu, J\. Wen, D\. Hu, and C\. Zhang \(2024a\)Multimodal fusion on low\-quality data: a comprehensive survey\.External Links:2404\.18947,[Link](https://arxiv.org/abs/2404.18947)Cited by:[§4\.3](https://arxiv.org/html/2607.26909#S4.SS3.p1.1)\.
- W\. Zhang, Z\. Yao, M\. Chen, Z\. Huang, and H\. Chen \(2023\)NeuralKG\-ind: a python library for inductive knowledge graph representation learning\.External Links:2304\.14678,[Link](https://arxiv.org/abs/2304.14678)Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p2.1)\.
- Y\. Zhang, Z\. Chen, L\. Guo, Y\. Xu, B\. Hu, Z\. Liu, W\. Zhang, and H\. Chen \(2024b\)Native: multi\-modal knowledge graph completion in the wild\.InProceedings of the 47th international ACM SIGIR conference on research and development in information retrieval,pp\. 91–101\.Cited by:[§2\.2](https://arxiv.org/html/2607.26909#S2.SS2.p1.1)\.
- Y\. Zhang, W\. Chen, X\. Chen, Q\. Ma, and L\. Zhao \(2024c\)Inductive link prediction for sequential\-emerging knowledge graph\.In2024 IEEE 40th International Conference on Data Engineering \(ICDE\),pp\. 5753–5766\.Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p2.1)\.
- Y\. Zhao, Y\. Zhang, B\. Zhou, X\. Qian, K\. Song, and X\. Cai \(2024\)Contrast then memorize: semantic neighbor retrieval\-enhanced inductive multimodal knowledge graph completion\.InProceedings of the 47th international ACM SIGIR conference on research and development in information retrieval,pp\. 102–111\.Cited by:[§1](https://arxiv.org/html/2607.26909#S1.p3.1)\.
- X\. Zhu, C\. Yang, Q\. Wang, Z\. Guo, Y\. Wang, Z\. Liu, C\. Wang, and C\. Shi \(2026a\)Graph4LLM: a systematic survey of graph\-enhanced large language models\.TechRxiv2026\(0213\),pp\.\.External Links:[Document](https://dx.doi.org/10.36227/techrxiv.177100285.57294132/v1)Cited by:[§4\.3](https://arxiv.org/html/2607.26909#S4.SS3.p1.1)\.
- Z\. Zhu, Z\. Zhang, L\. Xhonneux, and J\. Tang \(2021\)Neural bellman\-ford networks: a general graph neural network framework for link prediction\.Advances in Neural Information Processing Systems34\.Cited by:[§2\.1](https://arxiv.org/html/2607.26909#S2.SS1.p1.1)\.
- Z\. Zhu, J\. Liang, S\. Jiang, J\. Fu, M\. Liu, G\. Sun, S\. Ng, and B\. Qin \(2026b\)Analyzing reasoning consistency in large multimodal models under cross\-modal conflicts\.External Links:2601\.04073,[Link](https://arxiv.org/abs/2601.04073)Cited by:[§2\.2](https://arxiv.org/html/2607.26909#S2.SS2.p1.1)\.

## Appendix ATraining Objective

The model is trained using the binary cross\-entropy loss:

ℒ=−1\|𝒮t\|∑\(q,r,e\)∈𝒮t\[\\displaystyle\\mathcal\{L\}=\-\\frac\{1\}\{\|\\mathcal\{S\}\_\{t\}\|\}\\sum\_\{\(q,r,e\)\\in\\mathcal\{S\}\_\{t\}\}\\Big\[ye​log⁡\(σ​\(s​\(e\)\)\)\\displaystyle y\_\{e\}\\log\(\\sigma\(s\(e\)\)\)\(10\)\+\(1−ye\)log\(1−σ\(s\(e\)\)\)\],\\displaystyle\+\(1\-y\_\{e\}\)\\log\(1\-\\sigma\(s\(e\)\)\)\\Big\],where𝒮t\\mathcal\{S\}\_\{t\}denotes the training set,ye∈\{0,1\}y\_\{e\}\\in\\\{0,1\\\}indicates whether entityeeis the ground\-truth tail entity,s​\(e\)s\(e\)is its prediction score, andσ​\(⋅\)\\sigma\(\\cdot\)denotes the sigmoid function\.

## Appendix BLoss\-Guided Dynamic Graph Sampling

DuPLeR is jointly pre\-trained on multiple MMKGs with different scales and learning difficulties\. Uniform graph sampling may allocate insufficient training effort to difficult graphs, whereas sampling directly according to instantaneous losses may produce unstable distributions or overemphasize a small subset of graphs\. We therefore adopt a loss\-guided dynamic sampling strategy that adaptively allocates training steps while preserving a minimum sampling probability for every graph\.

For each pre\-training graph𝒢i\\mathcal\{G\}\_\{i\}, we maintain an exponential moving average \(EMA\) of its training loss\. After a training step produces a lossℓi\\ell\_\{i\}on𝒢i\\mathcal\{G\}\_\{i\}, the corresponding estimate is updated as

ℒ¯i←α​ℒ¯i\+\(1−α\)​ℓi,\\bar\{\\mathcal\{L\}\}\_\{i\}\\leftarrow\\alpha\\bar\{\\mathcal\{L\}\}\_\{i\}\+\(1\-\\alpha\)\\ell\_\{i\},\(11\)whereα∈\[0,1\)\\alpha\\in\[0,1\)denotes the EMA smoothing coefficient\. Consequently, graphs that remain more difficult to optimize receive higher estimated losses\. The preliminary graph sampling probability is obtained by applying a temperature\-scaled softmax to the EMA loss estimates:

p~i=exp⁡\(ℒ¯∗i/τ\)∑∗j=1G​exp⁡\(ℒ¯j/τ\),\\tilde\{p\}\_\{i\}=\\frac\{\\exp\\left\(\\bar\{\\mathcal\{L\}\}\*i/\\tau\\right\)\}\{\\sum\*\{j=1\}^\{G\}\\exp\\left\(\\bar\{\\mathcal\{L\}\}\_\{j\}/\\tau\\right\)\},\(12\)whereGGis the number of pre\-training graphs andτ\>0\\tau\>0controls the concentration of the sampling distribution\. A smallerτ\\tauassigns a larger proportion of training steps to graphs with higher estimated losses\. To prevent low\-loss graphs from being under\-sampled, we impose a minimum sampling probability:

pi=\(1−G​pmin\)​p~∗i\+p∗min,p\_\{i\}=\\left\(1\-Gp\_\{\\min\}\\right\)\\tilde\{p\}\*i\+p\*\{\\min\},\(13\)wherepminp\_\{\\min\}denotes the minimum sampling probability assigned to each graph\. This formulation guarantees thatpi≥pminp\_\{i\}\\geq p\_\{\\min\}and∑i=1Gpi=1\\sum\_\{i=1\}^\{G\}p\_\{i\}=1, provided thatpmin≤1/Gp\_\{\\min\}\\leq 1/G\. During the initial training steps, the EMA estimates may not yet provide a reliable measure of graph difficulty\. We therefore adopt uniform graph sampling during a warm\-up phase and activate the loss\-guided sampling distribution once the estimates become sufficiently stable\.

We pre\-train DuPLeR on three MMKGs and setτ=0\.6\\tau=0\.6,α=0\.8\\alpha=0\.8, andpmin=0\.2p\_\{\\min\}=0\.2in all experiments\.

## Appendix CEffect\-Size Analysis

We conduct exact one\-sided paired Wilcoxon signed\-rank tests using results obtained with the same three random seeds to assess whether DuPLeR consistently improves over ProLINK\. Because only three paired runs are available, the tests have limited statistical power; with three nonzero paired differences, the smallest attainable exact one\-sidedpp\-value is 0\.125\. We therefore report the tests together with descriptive effect\-size estimates\. We report additional multi\-seed results for two 1\-shot settings, OpenBG\-IMG\-R V2 and V4, as shown in Table[4](https://arxiv.org/html/2607.26909#A3.T4)\.

Table 4:Per\-seed Hits@10 results \(%\) of DuPLeR and ProLINK with Qwen3\-8B on OpenBG\-IMG\-R under the 1\-shot setting\. Mean and sample standard deviation are computed over three random seeds\.Across OpenBG\-IMG\-R V2 and V4 under the 1\-shot setting, DuPLeR yields absolute improvements of 6\.77 and 3\.37 Hits@10 percentage points over ProLINK, respectively\. The exact one\-sided paired Wilcoxon signed\-rank tests yieldW\+=6\.0W^\{\+\}=6\.0andp=0\.125p=0\.125in both settings and therefore do not establish statistical significance atα=0\.05\\alpha=0\.05\. Nevertheless, all three paired differences are positive in both settings\. The observed paired effect sizes are large, with Cohen’sdz=4\.71d\_\{z\}=4\.71for V2 anddz=10\.47d\_\{z\}=10\.47for V4\.

## Appendix DPrompt Settings and Calibration Analysis

![Refer to caption](https://arxiv.org/html/2607.26909v1/x4.png)Figure 4:Prompt\-configuration and calibration analysis for DuPLeR with Qwen3\-VL\-8B\-Instruct on OpenBG\-IMG\-R V4 under the 1\-shot setting\. The prompt abbreviations follow the definitions provided in the main paper\.To analyze the impact of different prompt formulations and filtering thresholds on performance, we evaluated DuPLeR with the Qwen3\-VL\-8B\-Instruct model on the OpenBG\-IMG\-R V4 dataset under the 1\-shot setting using various combinations of prompt formulations and filtering thresholds\. Specifically, there are nine prompt configurations formed by combining three relation\-information formats \(Relation\-only, Triple\-only, and Fusion\) with three entity\-type constraints \(Mandatory, Guided, and Unrestricted\)\.

As shown in Figure[4](https://arxiv.org/html/2607.26909#A4.F4), the Avg calibration strategy yields consistently strong performance across all prompt configurations, with Hits@10 ranging from 17\.33 to 17\.37\. Although Relation\-only with the Mandatory constraint obtains the highest individual result of 17\.37, its margin over the other configurations is only 0\.04 points\. More importantly, the Fusion prompt exhibits the strongest robustness under fixed\-threshold filtering: all three type\-constraint variants achieve 17\.33 forτ=1\\tau=1,33, and55\. In comparison, Relation\-only prompts remain between 16\.00 and 16\.33, while Triple\-only prompts reach 17\.33 only under the Unrestricted constraint\. These results suggest that combining a relation description with an example triple provides the VLM with complementary semantic and structural cues, enabling it to generate richer and more discriminative relation context\.

The results also show that performance is largely insensitive to the exact value of the fixed threshold once filtering is applied, as most configurations produce identical results acrossτ∈\{1,3,5\}\\tau\\in\\\{1,3,5\\\}\. Moreover, the differences among Mandatory, Guided, and Unrestricted constraints become negligible when the Fusion prompt is used, indicating that sufficiently informative relation context reduces sensitivity to the type\-constraint formulation\. By contrast, stricter constraints may limit the usefulness of the generated context when only relation descriptions or example triples are provided\. Overall, the results show that both prompt composition and calibration affect downstream KGC performance when VLMs are used for relation\-context enrichment\. Among the evaluated configurations, Fusion provides the most stable performance across different filtering thresholds and type constraints\.