RADD: Retrieval-Augmented Discrete Diffusion for Multi-Modal Knowledge Graph Completion Guanglin Niu and Bo Li School of Artificial Intelligence, Beihang University {beihangngl, boli}@buaa.edu.cn
Conventional Single-Scorer Pipeline
Abstract Most multi-modal knowledge graph completion (MMKGC) models use one embedding scorer to do both retrieval over the full entity set and final decision making. We argue that this coupling is a core bottleneck: global high-recall search and local fine-grained disambiguation require different inductive biases. Therefore, we propose a Retrieval-Augmented Discrete Diffusion (RADD) framework to decouple retrieve and reranking for MMKGC. A relationaware multimodal KGE retriever serves as both global retriever and distillation teacher, while a conditional discrete denoiser performs shortlistlevel entity-identity generation for reranking. Training combines KGE supervision, denoising cross-entropy, and temperature-scaled distillation from the retriever to the denoiser. At inference, the designed Diff-Rerank first forms a top-K shortlist with the retriever and then reranks it with the denoiser, ensuring that recall is a strict prerequisite for precision. Experiments on three MMKGC benchmarks show that RADD achieves the best performance and consistent gains over strong unimodal, multimodal, and LLM-based baselines, while ablations further verify the contribution of each component.
1
Introduction
Knowledge graph completion (KGC) predicts missing head or tail entities in incomplete triples. Translation-based (Bordes et al., 2013), bilinear (Yang et al., 2015; Trouillon et al., 2016), and rotational (Sun et al., 2019; Chao et al., 2021) models provide strong foundations for structural KGC. Multi-modal knowledge graphs (MMKGs) further augment entities with image and text (Liu et al., 2019; Wang et al., 2023; Chen et al., 2024), offering richer evidence but causing uneven modality quality and relation-dependent signal reliability. We argue that the central bottleneck in MMKGC is decision allocation, not representation fusion. A single embedding scorer has to both preserve high
Query Triple
(h, r, ?)
Text
Image
Coupled Objective
Full-Entity Ranking 1 0.95 𝑒!
Objective 1: Global retrieval requires high recall over large candidate space → KGE model Objective 2: Prediction requires fine-grained local sharpness → Generative model
2
𝑒"
0.91
3
𝑒#
0.88
4
𝑒$
0.83
5
𝑒%&'(
0.81
6
…
…
Coupled objective issue: One model cannot process two incompatible objectives Decoupling Retrieval and Prediction
Assign each objective to its naturally suited module for operating at their full potential
RADD: Retrieval-then-Denoiser Pipeline Conditional Discrete Denoiser
KGE Retrieval
Query Triple
(h, r, ?)
Top-K Objective: Local prediction Objective: Global recall Shortlist via entity-identity recovery over entity set over the Top-K shortlist
𝑒
[Mask]
𝑒%&'(
…
arXiv:2604.25693v1 [cs.AI] 28 Apr 2026
Structure
… 𝑡=T
𝑡=1 Condition
𝑡=0
Ranking
✓ 1 𝑒%&'( 2
𝑒!
3
𝑒"
4
𝑒#
5
𝑒$
6
…
Advantage: Both objectives can be achieved exactly via the specialized modules
Figure 1: Conventional MMKGC uses one scorer for both high-recall search and fine-grained reranking, two objectives with conflicting inductive biases. RADD decouples them: a KGE retriever handles global search, while a discrete denoiser handles shortlist reranking.
recall over the full entity set and resolve delicate ambiguity within a small set of near-tied candidates, yet these subproblems favor different inductive biases: geometric smoothness for global search and conditional sharpness for local discrimination. Optimising one scorer for both therefore yields a compromise that serves neither role well. KGEonly models maintain strong recall but often struggle with shortlist-level disambiguation, whereas generative-only approaches (Saxena et al., 2022; Huang et al., 2025) model richer conditional distributions but suffer from severe vocabulary-scale sparsity without a prior retrieval stage. Our ablations make this trade-off explicit: on DB15K and MKG-W, retriever-only and denoiser-only variants lag behind the full model by 8–20 points. Figure 1 illustrates this mismatch. The dominant response in prior MMKGC work has been
to improve representation fusion: relation-aware combinations of structural, visual, and textual embeddings (Xie et al., 2017; Sergieh et al., 2018; Pezeshkpour et al., 2018; Wang et al., 2019; Lu et al., 2022; Wang et al., 2021; Cao et al., 2022; Li et al., 2023; Lee et al., 2023; Zhang et al., 2024b,a, 2025), and more recently LLM-based entity representations (Guo et al., 2024, 2025). These advances improve the input space, but still rely on a single embedding-based scorer at inference time, leaving the dual-objective coupling unresolved. To address this issue, we propose a RetrievalAugmented Discrete Diffusion (RADD) framework that explicitly decouples the two objectives and assigns each to to its naturally suited module as in Figure 1. Considering that the KGC task is fundamentally a conditional recovery task: given one observed entity and a relation, the model must recover the missing entity that completes the triple. A relation-aware multimodal KGE retriever handles search over the full entity set, where geometric efficiency and recall breadth are essential. A conditional discrete denoiser then operates only after the retriever has pruned the candidate space, recovering entity identities directly in the discrete vocabulary without confronting full-vocabulary sparsity. Conditioning the denoiser on the observed entity and relation semantics (including multimodal signals) provides a principled and effective way to control generation toward triple-consistent candidates. The two components are coupled during training through temperature-scaled teacher-student distillation. At inference, the developed Diff-Rerank enforces the decomposition: the retriever commits to a hard top-K shortlist, and the denoiser reranks exclusively within it. Our contributions are: • We identify dual-objective coupling as a structural bottleneck in MMKGC and recast the task as retrieval-augmented discrete diffusion over entity vocabularies, adapting D3PM (Austin et al., 2021) to MMKGC. • We propose a RADD framework comprising a relation-aware multimodal retriever, a conditional discrete denoiser, teacher-student distillation, and a Diff-Rerank inference mechanism that explicitly enforces retrieve-thenrerank decision allocation. • Experiments on three benchmarks show the better performance of our model than existing
baselines. The ablation studies further confirm the effectiveness of each component.
2
Related Work
KGE-based KGC. Translation-based (Bordes et al., 2013), bilinear (Yang et al., 2015; Trouillon et al., 2016; Balazevic et al., 2019), and rotational (Sun et al., 2019) KGE models score triples via geometric energies over learned entity and relation embeddings. Graph-neural (Zhu et al., 2021) and language-model scorers (Yao et al., 2019) show that the scoring mechanism design is at least as consequential as the representation. Common-senseenhanced frameworks (Niu et al., 2025a) further improve KGC by jointly leveraging facts and common sense. MMKGC adds uneven modality coverage and relation-dependent feature dominance (Chen et al., 2024; Niu et al., 2026, 2025b), rendering a single scorer increasingly inadequate. Multi-modal KGC. Early MMKGC models inject visual or textual features directly into KGE objectives (Xie et al., 2017; Sergieh et al., 2018; Pezeshkpour et al., 2018). Subsequent work improves fusion quality via optimal transport (Cao et al., 2022), cross-modal attention (Li et al., 2023), hybrid transformers (Lee et al., 2023), modalityaware negative sampling (Xu et al., 2022; Zhang et al., 2023), adaptive weighting under incomplete observations (Zhang et al., 2024b,a), and relationconditioned modality routing (Zhang et al., 2025). LLM-augmented variants (Guo et al., 2024, 2025; Su et al., 2026) further narrow the performance gap by enriching entity representations with language model knowledge. All retain a single scoring function at inference. RADD instead separates retrieval from reranking, making the retrieve-then-rerank boundary an explicit architectural constraint. Diffusion models for structured prediction. Continuous DDPM (Ho et al., 2020) and discrete D3PM (Austin et al., 2021) established the theoretical foundations for denoising-based generation. D3PM’s absorbing and uniform forward transitions over discrete state spaces are directly relevant to entity-identity recovery. DiffusionCom (Huang et al., 2025) brings a related perspective to MMKGC by multimodal diffusion. RADD differs in two respects: it denoises discrete entity identities rather than continuous embeddings, and confines the denoiser to a KGE-supplied shortlist, avoiding full-vocabulary sparsity.
Structure 𝐞(")
Text
KGE Retrieval Relation Gate 𝛼 (") 𝛼 ($) 𝛼 (%)
𝐞($) ('()*$)
𝐞& Image
𝐞(%)
KGE Scorer ℒ23/
Relation 𝑟
𝐫 Scores over 𝐸 entities
Context 𝐜
Conditional Discrete Denoiser 𝑞(𝐱$ | 𝐱+ , 𝑡) Forward Process
Teacher 𝐩 #
𝑡=1
𝑡=T-1 𝐱$
Denoiser Tail Prediction 𝐫 ? 𝐡 MLP
Top-K Shortlist 𝐶"
ℒ$,)-
[
[Mask]
… 𝑡=0
Context
]
1 2 3 4 5 𝑒+ 𝑒* 𝑒) 𝑒( 𝑒'
Shortlist 𝒞" (𝑞)
𝑡=T
… …
Ranked Entities
ℓ (𝑒|𝑞), 𝑒 ∈ 𝒞𝐾 (𝑞) 𝑆78 𝑒|𝑞 = @ diff −𝑀, 𝑒 ∉ 𝒞= (𝑞)
Head Prediction ? 𝐫 𝐭
Bidirectional Training
𝒄. ; 𝐫 ; 𝐱$ ; 𝜑(𝑡) ; 𝐝
Diff-Rerank Inference Query 𝑞 ℎ, 𝑟, ?
Denoiser Score ℓ$%&& (𝑒|𝑞)
Training Objectives
MLP ℒ./,0
ℒ0)44 = ℒ$,)- + 𝜆. ℒ./,0 Student 𝐩!
MLP 𝑝1 (𝑥+ = 𝑒|𝑥$ , 𝒄. , 𝑟, 𝑡, 𝐝)
ℒ0)"$)-- = 𝐾𝐿(𝐩 5 ||𝐩6 )
ℒ23/
Figure 2: Overview of the RADD framework. The KGE Retriever (left) fuses structural, visual, and textual embeddings via a relation gate and scores all |E| candidates to form a top-K shortlist (recall stage). The Conditional Discrete Denoiser (centre) takes the retriever context ch , a noisy entity token xt , a timestep embedding ϕ(t), and a direction embedding d as input, producing logits over entity identities (precision stage). During training, Ldistill aligns the denoiser distribution pD (student) with the frozen retriever’s soft rankings pT (teacher). At inference, Diff-Rerank hard-gates diffusion scores to the shortlist, enforcing recall as a strict prerequisite for reranking.
3
Task Definition
Let G = (E, R, T ) be a multi-modal KG where each entity carries a structural embedding e(s) , a projected visual embedding e(v) , and a projected textual embedding e(t) ∈ Rd . Given an incomplete triple (h, r, ?) or (?, r, t), the task is to rank all |E| candidates. RADD decomposes this into two coupled stages: a KGE-based retriever scoring every candidate and producing a top-K shortlist, and a conditional discrete denoiser reranking only those K candidates. KGE supervision and knowledge distillation connect the two stages, ensuring joint optimization rather than post-hoc assembly.
4
Approach
Figure 2 shows an overview of the RADD framework. The KGE retriever fuses structural, visual, and textual signals to construct a top-K shortlist, and the conditional discrete denoiser reranks that shortlist by recovering entity identities from a corrupted token. The two modules are coupled during training via temperature-scaled distillation. 4.1
Relation-Aware Multimodal Retriever
The retriever serves two tightly coupled roles: it performs search over the full entity set to construct the shortlist seen by the denoiser, and it supplies the multimodal conditioning context used to resolve shortlist-level ambiguity. Both roles are served by the same relation-gated representation. For each entity, the retriever constructs a relation(joint) conditioned joint representation er by fusing three modality-specific embeddings (structural e(s) ,
visual e(v) , textual e(t) ) through a relation gate: e(joint) = r
X
(r) (m) αm e ,
(1)
m∈{s,v,t} (r)
where the relation-specific weights αm allow each relation to up-weight its most informative modality. A RotatE-style scoring function on e(joint) in complex space serves as the retrieval energy, as complex-valued rotation natively models symmetric, antisymmetric, inverse, and compositional relation patterns (Sun et al., 2019), while admitting closed-form scoring over all |E| candidates. The (joint) fused representation er is then reused as the conditioning component ch in the denoiser input (Eq. 3), directly linking retrieval quality to denoising quality through a shared feature space. 4.2
Conditional Discrete Diffusion
The denoiser’s key novelty lies in its target space: it predicts clean entity identities as discrete indices into E, rather than continuous embedding vectors. This design choice directly aligns the denoising objective with the KGC objective itself: recovering the missing entity in an incomplete triple conditioned on the observed entity and relation. In our formulation, this conditioning signal is explicitly multimodal, since the denoiser receives relationaware context built from structural, visual, and textual features of the observed side. It also avoids the quantization step required when continuous generation outputs must be post-hoc mapped back to entity indices, a mapping that can introduce nonnegligible ranking errors when multiple entities
share similar embedding neighborhoods. Finally, cross-entropy over the discrete vocabulary E provides well-defined gradients without continuous relaxations, making joint training with the KGE objective straightforward. Following D3PM (Austin et al., 2021), we define a forward process over E ∪ {[MASK]}. Let x0 ∈ E be the clean target entity. The forward process corrupts x0 over T timesteps via a schedule-defined keep probability pkeep (t): q(xt | x0 , t) = pkeep (t) δx0 + pmask (t) δ[MASK] + prep (t) U(E \ {x0 }), (2) where pmask (t) + prep (t) = 1 − pkeep (t). Setting prep (t) = 0 recovers mask-only corruption, and we additionally use random replacement with base probability 0.3 and a linear schedule in experiments. The denoiser fθ is parameterized as an MLP that takes as input the concatenation x̃ = ch ; r; xt ; ϕ(t); d , (3) (joint)
where ch = er is the conditioning entity’s relation-aware representation, xt is the corrupted token xt (or the learned mask embedding when xt = [MASK]), ϕ(t) ∈ Rdt is a sinusoidal timestep embedding, and d ∈ Rddir is a learned direction embedding distinguishing head and tail prediction. The denoiser outputs logits over all |E| entities, reflecting the reverse model: pθ (x0 = e | xt , ch , r, t, d) = softmax(fθ (x̃)). (4) Since the target entity is unknown at inference, we set xT = [MASK] and query the denoiser directly at t = T , obtaining a full entity distribution in a single forward pass without iterative sampling. 4.3
Bidirectional Diffusion Training
KGC requires completing triples in two directions: tail prediction (h, r, ?) and head prediction (?, r, t). These two subtasks are structurally asymmetric: in most knowledge graphs, each relation places stronger type constraints on the tail than on the head, and head entities tend to follow a more skewed frequency distribution, making head prediction consistently harder. A unidirectional denoiser trained only on tail queries would underserve half of all inference cases, and training both directions jointly but weighting them equally would similarly underestimate the head-prediction difficulty. The
framework therefore trains both directions with an asymmetric loss, upweighting the head loss to compensate for this structural imbalance. The combined denoising loss is: Ldiff = Ltail + λh Lhead ,
(5)
where each directional term is a cross-entropy loss over E. The head-weighting coefficient λh > 1 compensates for the harder head-prediction task. The tail-only training variant serves as the corresponding ablation baseline. 4.4
Joint Optimization with Knowledge Distillation
The full training objective is: L = Lkge + Ldiff + λd Ldistill + λr Lrank .
(6)
Lkge is the standard negative-sampling KGE loss. The distillation term transfers the retriever’s soft predictions into the denoiser via temperature-scaled KL divergence over a candidate subset C containing the positive entity, random negatives, and KGE hard negatives: Ldistill = KL pT (C) pD (C) , exp(skge (e)/τ ) pT (e) = P , ′ e′ ∈C exp(skge (e )/τ )
(7)
where pD (e) is the denoiser’s softmax distribution over C at the same temperature τ . By using the retriever’s soft rankings as supervision targets, distillation transfers the retriever’s candidate-ordering knowledge into the denoiser before shortlist-level discrimination begins, so the denoiser starts from a probability mass already concentrated on plausible candidates rather than a uniform prior. An optional margin-based ranking term Lrank = max(0, m − (s+ − s− )) provides an additional margin signal. The retriever is frozen after a warm-up phase so that the teacher distribution is stationary when distillation begins, and sensitivity analysis (Section 5.5.4) confirms that earlier freezing strictly improves denoiser convergence. 4.5
Retrieval-Augmented Inference
We use a single inference rule, Diff-Rerank, because it most directly instantiates the proposed division of labor: the retriever is responsible for global retrieval, and the denoiser is responsible for local decision refinement. For a query q, the retriever
Table 1: Dataset statistics. Img and Txt count entities with at least one visual or textual feature. Dataset
#Ent
#Rel
Train
Valid
Test
Img
Txt
DB15K MKG-W MKG-Y
12,842 15,000 15,000
279 169 28
79,222 34,196 21,310
9,904 4,276 2,665
9,902 4,274 2,663
12,818 14,463 14,244
9,078 14,123 12,305
first computes KGE scores over the full entity set and forms a shortlist CK (q) = TopKe∈E skge (e | q).
(8)
This step preserves the retriever’s strongest property: stable high-recall search over the entire entity set under noisy or incomplete multimodal evidence. The denoiser is then queried with xT = [MASK] and the same query context to produce denoiser log-scores ℓdiff (e | q) = log pθ (x0 = e | xT , q). The final Diff-Rerank score is ( ℓdiff (e | q), e ∈ CK (q), Sdr (e | q) = (9) −M, e∈ / CK (q), where M is a large constant that hard-gates any entity outside the shortlist to last place. This constraint makes recall a strict prerequisite for final ranking: an entity can only be promoted by the denoiser if the retriever has already shortlisted it, eliminating the failure mode of soft score fusion where a high-confidence denoiser score could compensate for a retrieval miss. The result is a clean separation: the retriever is solely responsible for coverage, and the denoiser is solely responsible for precision within the retrieved set.
5
Experiments
5.1
Datasets
We evaluate on three standard MMKGC benchmarks, DB15K, MKG-W, and MKG-Y, all providing pre-extracted visual and textual embeddings alongside structural triples (Table 1) (Zhang et al., 2025). Following the standard filtered KGC evaluation protocol, we report MRR, H@1, H@3, and H@10 averaged over head and tail prediction queries. To test whether the gain depends on multimodal evidence, we additionally evaluate RADD in structure-only mode on two traditional KGC benchmarks, FB15K-237 and NELL995. 5.2
Implementation Details
Modality embeddings are projected into the entity embedding space via a learned linear layer. Retriever. Embedding dimension 250, batch size
1024, 128 negatives per positive, KGE learning rate 10−4 , and the KGE encoder is frozen after epoch 1000 to serve as a stable distillation teacher. Denoiser. Learning rate 10−4 , T =100 diffusion steps, cosine noise schedule, replacement corruption probability 0.3 with linear annealing, bidirectional training with head-loss weight λh =2.0, EMA decay 0.9999. Distillation and inference. Temperature τ =0.7 over a 64-candidate training pool, and Diff-Rerank applies a top-K=256 shortlist at inference. All hyperparameter choices are confirmed by the sensitivity analysis in Section 5.5.4. 5.3
Baselines
We compare against 27 baselines in four groups. Unimodal KGC: TransE (Bordes et al., 2013), TransD (Ji et al., 2015), DistMult (Yang et al., 2015), ComplEx (Trouillon et al., 2016), RotatE (Sun et al., 2019), PairRE (Chao et al., 2021), GC-OTE (Tang et al., 2020), plus RADDS (structure-only RADD) to isolate architectural from multimodal gains. Multimodal KGC: IKRL (Xie et al., 2017), TBKGC (Sergieh et al., 2018), TransAE (Wang et al., 2019), MMKRL (Lu et al., 2022), RSME (Wang et al., 2021), VBKGC (Zhang and Zhang, 2022), OTKGE (Cao et al., 2022), MoSE (Zhao et al., 2022), IMF (Li et al., 2023), QEB (Wang et al., 2023), VISTA (Lee et al., 2023), AdaMF-MAT (Zhang et al., 2024b), and MoMoK (Zhang et al., 2025), all of which use a single scoring function at inference. Negative sampling: KBGAN (Cai and Wang, 2018), MANS (Zhang et al., 2023), MMRNS (Xu et al., 2022), and DHNS (Niu and Zhang, 2026) improve training-time sample quality. LLM-augmented: MKGL (Guo et al., 2024) and K-ON (Guo et al., 2025) enrich entity representations with LLM but retain a single-scorer inference pipeline. 5.4
Research Questions
We structure the analysis around four questions: (RQ1) Does RADD outperforms the existing strong baselines? (RQ2) Does RADD reduce the structural asymmetry between head and tail prediction, and what design choices are responsible? (RQ3) Which components drive the gains and are they consistent across datasets? (RQ4) How sensitive is RADD to its key hyperparameters, and which design choices require careful tuning?
Table 2: The comparison results (%) of MMKGC on three datasets. The best results are bold, and the second-best ones are underlined. RADD-S denotes the structure-only variant of RADD. DB15K
Model
MKG-W
MKG-Y
MRR
H@1
H@3
H@10
MRR
H@1
H@3
H@10
MRR
H@1
H@3
H@10
Unimodal KGC
TransE TransD DistMult ComplEx RotatE PairRE GC-OTE RADD-S
24.86 21.52 23.03 27.48 29.28 31.13 31.85 41.17
12.78 8.34 14.78 18.13 17.87 21.62 22.11 33.08
31.48 29.93 26.98 31.57 36.12 35.91 36.52 45.08
47.07 44.24 40.59 45.37 49.66 49.30 51.18 57.17
29.19 26.84 20.95 28.09 33.67 34.40 33.92 36.48
21.06 19.65 15.89 21.45 26.80 28.24 26.55 30.80
33.20 31.48 22.88 30.89 36.68 36.71 35.96 38.59
44.23 42.68 36.80 44.77 46.73 46.04 46.05 47.34
30.73 26.39 25.04 28.94 34.95 32.01 32.95 38.96
23.45 17.01 19.32 23.11 29.10 25.53 26.77 35.32
35.18 33.05 27.80 31.07 38.35 35.83 36.44 40.56
43.37 40.41 39.95 43.48 45.30 43.89 44.08 45.40
MMKGC
IKRL TBKGC TransAE MMKRL RSME VBKGC OTKGE MoSE IMF QEB VISTA AdaMF AdaMF-MAT MoMoK
26.82 28.40 28.09 26.81 29.76 30.61 23.86 28.38 32.25 28.18 30.42 32.51 35.14 39.57
14.09 15.61 21.25 13.85 24.15 19.75 18.45 21.56 24.20 14.82 22.49 21.31 25.30 32.38
34.93 37.03 31.17 35.07 32.12 37.18 25.89 30.91 36.00 36.67 33.56 39.67 41.11 43.45
49.09 49.86 41.17 49.39 40.29 49.44 34.23 41.67 48.19 51.55 45.94 51.68 52.92 54.14
32.36 31.48 30.00 30.10 29.23 30.61 34.36 33.34 34.50 32.38 32.91 34.27 35.85 35.89
26.11 25.31 21.23 22.16 23.36 24.91 28.85 27.78 28.77 25.47 26.12 27.21 29.04 30.38
34.75 33.98 34.91 34.09 31.97 33.01 36.25 33.94 36.62 35.06 35.38 37.86 39.01 37.54
44.07 43.24 44.72 44.69 40.43 40.88 44.88 41.06 45.44 45.32 45.61 47.21 48.42 46.13
33.22 33.99 28.10 36.81 34.44 37.04 35.51 36.28 35.79 34.37 30.45 38.06 38.57 37.91
30.37 30.47 25.31 31.66 31.78 33.76 31.97 33.64 32.95 29.49 24.87 33.49 34.34 35.09
34.28 35.27 29.10 39.79 36.07 38.75 37.18 37.47 37.14 36.95 32.39 40.44 40.59 39.20
38.26 40.07 33.03 45.31 39.09 42.30 41.38 40.81 40.63 42.32 41.53 45.48 45.76 43.20
Negative Sampling
KBGAN MANS MMRNS DHNS
24.83 28.82 32.68 34.59
19.48 16.87 23.01 26.80
27.15 36.58 37.86 39.20
36.49 49.26 51.01 51.20
25.78 30.88 35.03 36.41
19.05 24.89 28.59 29.73
28.71 33.63 37.49 39.75
37.89 41.78 47.47 49.22
24.26 29.03 35.93 36.49
20.68 25.25 30.53 32.88
25.68 31.35 39.07 38.89
29.16 34.49 45.47 42.95
LLM-based MMKGC
MKGL K-ON
32.86 36.64
26.54 30.05
35.57 38.72
44.76 48.26
29.11 -
24.30 -
31.60 -
38.28 -
27.14 38.10
18.68 30.13
30.39 42.77
43.87 53.59
Ours
RADD
45.02
37.16
48.94
60.26
45.45
39.88
47.58
55.83
39.40
35.75
40.93
45.87
5.5
Results and Analysis
5.5.1
Main Results (RQ1)
Table 2 summarises competitive results across all three benchmarks. RADD ranks first in MRR on all three datasets, achieving 45.02 on DB15K, 45.45 on MKG-W, and 39.40 on MKG-Y, the only model in the comparison to lead simultaneously across all three. Against the strongest single-scorers (MoMoK at 39.57/35.89 on DB15K/MKG-W, and AdaMF-MAT at 38.57 on MKG-Y), the absolute MRR gains are +5.5, +9.6, and +0.8. RADD also exceeds the best LLM-augmented baseline K-ON (MRR 36.64 on DB15K) by +8.3 points, confirming that the retrieve-then-rerank decomposition is more effective than enriching representations alone. Notably, RADD achieves this with approximately 22 M trainable parameters, on par with MKGL (20 M) and over an order of magnitude fewer than K-ON (215 M), demonstrating that architectural role specialization delivers superior accuracy more efficiently than scaling up pretrained language models. Gains are most pronounced in
H@3 and H@10, the metrics most directly sensitive to shortlist reranking quality. The structure-only variant RADD-S already reaches 41.17, 36.48, and 38.96 MRR on DB15K, MKG-W, and MKG-Y, surpassing the strongest unimodal baselines on each dataset (GC-OTE, PairRE, and RotatE, respectively) by +9.3, +2.1, and +4.0 MRR. Strikingly, RADD-S’s score of 41.17 on DB15K also exceeds every multimodal and LLMaugmented baseline in Table 2, establishing that the retrieve-then-rerank architecture alone, without any visual or textual input, already sets a new state of the art on that benchmark. Full RADD further improves over RADD-S by +3.9 on DB15K, +9.0 on MKG-W, and +0.4 on MKG-Y, confirming that multimodal conditioning provides incremental gains on datasets with higher modality coverage and relational diversity. The gains follow a consistent gradient across datasets. On DB15K and MKG-W, where relational diversity, triple density, and modality coverage are all high, RADD’s improvements are substantial. On MKG-Y (28 relation types, lower
Table 3: Head and tail prediction results (%) on DB15K and MKG-W. Best in bold, second-best underlined. DB15K Model
MKG-W
Head
Tail
Head
Tail
MRR H@1 H@3 H@10 MRR H@1 H@3 H@10 MRR H@1 H@3 H@10 MRR H@1 H@3 H@10 AdaMF-MAT 23.94 16.67 27.59 26.23 19.69 28.94 MoMoK RADD 39.71 33.18 42.35
36.85 38.29 52.02
47.94 35.51 56.14 51.68 43.18 56.29 50.44 41.15 55.54
70.42 67.89 68.50
18.56 13.10 20.57 18.50 13.59 19.82 35.67 30.86 37.48
28.57 27.61 44.06
52.80 44.85 56.79 51.66 45.95 54.14 55.24 48.90 57.67
67.69 62.70 67.62
Table 4: Ablation results (%). Each ablated model removes or replaces one component of RADD. Best results in bold, and the most degraded variant per dataset is marked with superscript ∗ . DB15K
Model Full model
MKG-W
MKG-Y
MRR
H@1
H@3
H@10
MRR
H@1
H@3
H@10
MRR
H@1
H@3
H@10
45.02
37.16
48.94
60.26
45.45
39.88
47.58
55.83
39.40
35.75
40.93
45.87
33.08 27.77
45.08 39.69
57.17 50.53
36.48 35.23
30.80 29.97
38.59 37.44
47.34 45.33
38.96 38.44
35.32 35.05
40.56 40.07
45.40 44.61
Retriever
w/o Multimodality w/o Retrieval
41.17 35.56
Denoiser
w/o Denoising Tail-only Training
32.83 24.27 36.91 49.65 35.05 27.99 38.29 47.96 38.74 34.57 40.59 45.85 25.40∗ 19.31∗ 28.24∗ 37.19∗ 30.02∗ 25.90∗ 31.76∗ 38.11∗ 31.07∗ 27.62∗ 32.86∗ 37.83∗
Training
w/o Distillation
39.93
32.00
43.56
55.30
modality diversity), the gain narrows to +0.8 MRR, as the simpler relational structure leaves less residual ambiguity for the denoiser to resolve. Rather than being a limitation, this graduated behavior validates the central hypothesis: RADD’s margin scales predictably with shortlist disambiguation difficulty, delivering its largest gains precisely where the retrieve-then-rerank design is most needed. 5.5.2 Head vs. Tail Entity Prediction (RQ2) Table 3 separates head and tail prediction to locate the source of RADD’s overall gains. Head prediction. Competing models score head MRR of 23.94–26.23 on DB15K and 18.50–18.56 on MKG-W, while their tail MRR reaches 47.94– 51.68 and 51.66–52.80, yielding tail-to-head ratios of approximately 2.0× and 2.8×. RADD uniquely closes this structural gap, reducing the ratios to 1.27× and 1.55× through dramatic head-side improvements: +13.5 MRR (+51.4% relative) over MoMoK on DB15K and +17.1 MRR (+92.2% relative) over AdaMF-MAT on MKG-W. The H@1 gains are even more pronounced, at +68.5% relative on DB15K and +127% relative on MKG-W, and no other model comes close to these margins in either prediction direction. This recovery is driven by bidirectional diffusion training, which together counteract the many-to-one relational bias that systematically suppresses head-side recall in standard single-direction training. Tail prediction. On DB15K, RADD’s tail MRR (50.44) falls slightly below MoMoK (51.68), and
35.60
29.35
38.39
47.09
39.08
35.35
40.69
45.49
MoMoK also leads in tail H@3. On MKG-W, RADD leads in tail MRR, H@1, and H@3, while AdaMF-MAT retains a narrow advantage in H@10. This shows that competitive baselines already capture most of the available tail-side signal, and RADD does not sacrifice tail performance while substantially recovering head-side recall. Overall. Balanced bidirectional performance is RADD’s most distinctive structural advantage among the compared models. Rather than further improving the already-strong tail side, the retrieve-then-rerank design with asymmetric bidirectional training substantially recovers head-side recall without sacrificing tail performance, and it is precisely this head-side recovery that drives the largest share of RADD’s aggregate MRR improvements in Table 2. 5.5.3
Ablation Study (RQ3)
Table 4 isolates the contribution of each component by ablating one module at a time. Three findings hold consistently across all three benchmarks. Both the retriever and denoiser are necessary: neither suffices alone. The full model achieves 45.02/45.45/39.40 MRR on DB15K/MKG-W/MKG-Y. Using the retriever alone reduces MRR to 32.83/35.05/38.74 (gaps of 12.2/10.4/0.7 points), while the denoiser alone reaches only 35.56/35.23/38.44. Both singlemodule variants lag substantially behind the full system on the two richer datasets, confirming that recall-focused retrieval and shortlist-level de-
H@10
50.0
52.0
50.0
50.0
47.5 45.0 42.5
Performance (%)
54.0
52.5 47.5 45.0 42.5
48.0 46.0 44.0
40.0
40.0
42.0
37.5
37.5
40.0
256
Batch Size
512
1024
55.0
55.0
52.5
52.5
50.0
50.0
47.5 45.0 42.5 40.0
128
256
512
Top-K Shortlist
768
38.0
1024
47.5 45.0 42.5 40.0
3
4
Margin
5
6
1e-4
2e-4
5e-4
KGE Learning Rate
1e-3
50.0
Performance (%)
50.0
Performance (%)
52.0
50.0
48.0 46.0 44.0
42.0
40.0
40.0
1e-3
0.3
0.5
0.7
0.9
44.0
40.0 5e-4
0.1
46.0
42.0
2e-4
20
50
75
Diffusion Steps T
100
150
KGE Freeze Epoch
48.0
42.0
Diffusion Learning Rate
2000
37.5 5e-5
54.0
1e-4
1500
42.5
52.0
5e-5
1000
45.0
54.0
44.0
300
47.5
52.0
46.0
500
250
50.0
54.0
48.0
100
200
Embedding Dimension
40.0
35.0 2
150
52.5
37.5
37.5
100
55.0
Performance (%)
128
35.0
Performance (%)
Performance (%)
H@3
55.0
56
Performance (%)
H@1
52.5 Performance (%)
Performance (%)
MRR 55.0
Distillation Weight
Figure 3: Hyperparameter sensitivity on MKG-W. Each subplot plots MRR, H@1, H@3, and H@10 (%) against one swept parameter. The dashed vertical line indicates the primary configuration used in the main experiments.
noising are complementary rather than redundant. Removing the multimodal signals while keeping the dual-module structure (RADD-S) still reduces MRR by −3.9/−9.0/−0.4, indicating that visual and textual modalities provide essential evidence beyond structural signals alone, with the effect amplified on datasets with higher modality coverage. Bidirectional training is indispensable for head-side disambiguation. Tail-only training incurs the largest overall degradation (∆MRR: −19.6/−15.4/−8.3 on DB15K/MKG-W/MKG-Y). Per-direction inspection (Table 3) reveals that tail performance remains competitive under tail-only training, while head MRR collapses disproportionately, indicating that the bulk of the degradation is head-side. This confirms that inverse-relation augmentation and the asymmetric head-loss weight λh are both necessary to counteract the structural many-to-one bias that suppresses head-side recall in standard training. Distillation gain scales with retriever quality. Removing distillation reduces MRR by −5.1 on
DB15K and −9.9 on MKG-W, but only −0.3 on MKG-Y. The pattern tracks dataset richness: on DB15K and MKG-W, the KGE shortlist carries informative soft rankings that transfer actionable reranking guidance to the denoiser, while on MKGY, where fewer relation types and lower modality diversity reduce shortlist confidence, the distillation channel conveys correspondingly less signal. 5.5.4
Hyperparameter Sensitivity (RQ4)
Figure 3 sweeps nine parameters and reveals a practically favorable tuning asymmetry: retrievalside settings drive nearly all performance variance, while diffusion-side settings are broadly stable. General parameters (Row 1). Top-K shortlist size is the most sensitive factor (MRR range across tested values: 4.4%): K = 128 focuses denoising on the most confident KGE candidates, while K > 256 admits marginal candidates the retriever cannot separate. Batch size improves MRR monotonically (41.68→44.58), and embedding dimension is stable (MRR range: 0.8%).
KGE-specific parameters (Row 2). KGE learning rate shows the sharpest sensitivity (MRR range: 3.4%): raising it to 5×10−4 degrades H@1 by 3.6 pp with little effect on H@10, reflecting deteriorated top-rank geometry. Earlier retriever freezing is consistently better (epoch 100: 45.14% vs. epoch 2000: 42.93%), as it provides a stationary distillation target. Margin peaks at 4, with 2.0 pp loss at both extremes. Diffusion-specific parameters (Row 3). All diffusion-side settings are nearly flat (MRR range <0.5% across all tested values): distillation weight shifts MRR by less than 0.1 pp over a 9× range, and step count T varies by only 0.2 pp over T ∈ {20, . . . , 150}. Tuning effort can therefore be concentrated entirely on the retriever, while the denoiser performs near-optimally at its default configuration. Instance-level rank traces on representative queries from DB15K and MKG-W are presented in Appendix C, providing per-query evidence for the role specialization hypothesis.
6
Limitations
The most fundamental limitation of RADD is the hard recall. Diff-Rerank permanently excludes any entity not in the top-K shortlist, so retrieval misses are unrecoverable and overall performance is bounded by retriever recall. In terms of computational complexity, each query requires one KGE forward pass and one denoiser logit projection procedures, roughly doubling per-query latency relative to a single scorer. The retriever must warm up and stabilise before being frozen as the distillation teacher, introducing a serial dependency that increases total training time. Besides, embeddings are pre-extracted and fixed, and end-to-end finetuning of the encoders is a natural extension that may yield further gains on modality-rich datasets.
7
Conclusion
RADD enforces an explicit retrieve-then-rerank decomposition for MMKGC: a relation-aware multimodal retriever handles recall over the full entity set, and a conditional discrete denoiser resolves residual shortlist ambiguity. Across three benchmarks, role-specialized composition consistently outperforms enriched single-scorers, with gains scaling with retriever shortlist quality and withinshortlist ambiguity. Additional structure-only experiments on FB15K-237 and NELL995 further
show that the same decomposition remains effective even without multimodal inputs. Ablations confirm that the multimodal retriever, bidirectional training, and Diff-Rerank hard gating are each necessary, and that distillation gains are proportional to teacher quality. Discrete denoising contributes most when its scope is bounded by prior retrieval, complementing rather than replacing advances in fusion or representation.
References Jacob Austin, Daniel D. Johnson, Jonathan Ho, Daniel Tarlow, and Rianne van den Berg. 2021. Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems. Ivana Balazevic, Carl Allen, and Timothy M. Hospedales. 2019. Tucker: Tensor factorization for knowledge graph completion. In Conference on Empirical Methods in Natural Language Processing and International Joint Conference on Natural Language Processing, pages 5184–5193. Antoine Bordes, Nicolas Usunier, Alberto GarcíaDurán, Jason Weston, and Oksana Yakhnenko. 2013. Translating embeddings for modeling multirelational data. In Advances in Neural Information Processing Systems. Liwei Cai and William Yang Wang. 2018. Kbgan: Adversarial learning for knowledge graph embeddings. In North American Chapter of the Association for Computational Linguistics. Zongsheng Cao, Qianqian Xu, Zhiyong Yang, Yuan He, Xiaochun Cao, and Qingming Huang. 2022. Otkge: Multi-modal knowledge graph embeddings via optimal transport. In Advances in Neural Information Processing Systems. Linlin Chao, Jianshan He, Taifeng Wang, and Wei Chu. 2021. Pairre: Knowledge graph embeddings via paired relation vectors. In Annual Meeting of the Association for Computational Linguistics. Zhuo Chen, Yichi Zhang, Yin Fang, Yuxia Geng, Lingbing Guo, Xiang Chen, Qian Li, Wen Zhang, Jiaoyan Chen, Yushan Zhu, Jiaqi Li, Xiaoze Liu, Jeff Z. Pan, Ningyu Zhang, and Huajun Chen. 2024. Knowledge graphs meet multi-modal learning: A comprehensive survey. Preprint, arXiv:2402.05391. Lingbing Guo, Zhongpu Bo, Zhuo Chen, Yichi Zhang, Jiaoyan Chen, Lan Yarong, Mengshu Sun, Zhiqiang Zhang, Yangyifei Luo, Qian Li, and 1 others. 2024. MKGL: Mastery of a three-word language. Advances in Neural Information Processing Systems, 37:140509–140534. Lingbing Guo, Yichi Zhang, Zhongpu Bo, Zhuo Chen, Mengshu Sun, Zhiqiang Zhang, Wen Zhang, and
Huajun Chen. 2025. K-ON: Stacking knowledge on the head layer of large language model. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 11745–11753. Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems. Wei Huang, Meiyu Liang, Peining Li, Xu Hou, Yawen Li, Junping Du, Zhe Xue, and Zeli Guan. 2025. Diffusioncom: Structure-aware multimodal diffusion model for multimodal knowledge graph completion. Preprint, arXiv:2504.06543. Guoliang Ji, Shizhu He, Liheng Xu, Kang Liu, and Jun Zhao. 2015. Knowledge graph embedding via dynamic mapping matrix. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics, pages 687–696. Jaejun Lee, Chanyoung Chung, Hochang Lee, Sungho Jo, and Joyce Jiyoung Whang. 2023. Vista: Visualtextual knowledge graph representation learning. In Findings of the Association for Computational Linguistics: EMNLP, pages 7314–7328. Xinhang Li, Xiangyu Zhao, Jiaxing Xu, Yong Zhang, and Chunxiao Xing. 2023. Imf: Interactive multimodal fusion model for link prediction. In The Web Conference, pages 2572–2580. Ye Liu, Hui Li, Alberto García-Durán, Mathias Niepert, Daniel Oñoro-Rubio, and David S. Rosenblum. 2019. Mmkg: Multi-modal knowledge graphs. In The Semantic Web, pages 459–474. Xinyu Lu, Lifang Wang, Zejun Jiang, Shichang He, and Shizhong Liu. 2022. Mmkrl: A robust embedding approach for multi-modal knowledge graph representation learning. Applied Intelligence, 52(7):7480– 7497. Guanglin Niu, Bo Li, and Siling Feng. 2025a. A pluggable common sense-enhanced framework for knowledge graph completion. IEEE Transactions on Big Data, 11(6):3282–3299. Guanglin Niu, Bo Li, and Yangguang Lin. 2025b. A survey of task-oriented knowledge graph reasoning: Status, applications, and prospects. Preprint, arXiv:2506.11012. Guanglin Niu, Bo Li, and Yangguang Lin. 2026. A comprehensive survey of knowledge graph reasoning: Approaches and applications. IEEE Transactions on Big Data, pages 1–20. Guanglin Niu and Xiaowei Zhang. 2026. Diffusionbased hierarchical negative sampling for multimodal knowledge graph completion. In Database Systems for Advanced Applications, pages 479–495, Singapore. Springer Nature Singapore.
Pouya Pezeshkpour, Liyan Chen, and Sameer Singh. 2018. Embedding multimodal relational data for knowledge base completion. In Conference on Empirical Methods in Natural Language Processing, pages 3208–3218. Apoorv Saxena, Adrian Kochsiek, and Rainer Gemulla. 2022. Sequence-to-sequence knowledge graph completion and question answering. In Annual Meeting of the Association for Computational Linguistics, pages 2814–2828. Hatem Mousselly Sergieh, Teresa Botschen, Iryna Gurevych, and Stefan Roth. 2018. A multimodal translation-based approach for knowledge graph representation learning. In *SEM @ NAACL-HLT, pages 225–234. Siyue Su, Jian Yang, Bo Li, and Guanglin Niu. 2026. Tokenization, fusion and decoupling: Bridging the granularity mismatch between large language models and knowledge graphs. Preprint, arXiv:2602.22698. Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. 2019. Rotate: Knowledge graph embedding by relational rotation in complex space. In International Conference on Learning Representations. Yun Tang, Jing Huang, Guangtao Wang, Xiaodong He, and Bowen Zhou. 2020. Orthogonal relation transforms with graph context modeling for knowledge graph embedding. In Annual Meeting of the Association for Computational Linguistics. Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In International Conference on Machine Learning, pages 2071– 2080. 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. In ACM Multimedia, pages 2735–2743. Xin Wang, Benyuan Meng, Hong Chen, Yuan Meng, Ke Lv, and Wenwu Zhu. 2023. Tiva-kg: A multimodal knowledge graph with text, image, video and audio. In ACM Multimedia, pages 2391–2399. Zikang Wang, Linjing Li, Qiudan Li, and Daniel Zeng. 2019. Multimodal data enhanced representation learning for knowledge graphs. In International Joint Conference on Neural Networks, pages 1–8. Ruobing Xie, Zhiyuan Liu, Huanbo Luan, and Maosong Sun. 2017. Image-embodied knowledge representation learning. In International Joint Conference on Artificial Intelligence, pages 3140–3146. Derong Xu, Tong Xu, Shiwei Wu, Jingbo Zhou, and Enhong Chen. 2022. Relation-enhanced negative sampling for multimodal knowledge graph completion. In ACM Multimedia, pages 3857–3866.
Bishan Yang, Wen tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. Embedding entities and relations for learning and inference in knowledge bases. In International Conference on Learning Representations. Liang Yao, Chengsheng Mao, and Yuan Luo. 2019. Kgbert: Bert for knowledge graph completion. Preprint, arXiv:1909.03193. Yichi Zhang, Mingyang Chen, and Wen Zhang. 2023. Modality-aware negative sampling for multi-modal knowledge graph embedding. In International Joint Conference on Neural Networks, pages 1–8. Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Binbin Hu, Ziqi Liu, Wen Zhang, and Huajun Chen. 2024a. Native: Multi-modal knowledge graph completion in the wild. In International ACM SIGIR Conference on Research and Development in Information Retrieval. Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Binbin Hu, Ziqi Liu, Wen Zhang, and Huajun Chen. 2025. Multiple heads are better than one: Mixture of modality knowledge experts for entity representation learning. In International Conference on Learning Representations. Yichi Zhang, Zhuo Chen, Lei Liang, Huajun Chen, and Wen Zhang. 2024b. Unleashing the power of imbalanced modality information for multi-modal knowledge graph completion. In Joint International Conference on Computational Linguistics, Language Resources and Evaluation, pages 17120–17130. Yichi Zhang and Wen Zhang. 2022. Knowledge graph completion with pre-trained multimodal transformer and twins negative sampling. Preprint, arXiv:2209.07084. Yu Zhao, Xiangrui Cai, Yike Wu, Haiwei Zhang, Ying Zhang, Guoqing Zhao, and Ning Jiang. 2022. Mose: Modality split and ensemble for multimodal knowledge graph completion. In Conference on Empirical Methods in Natural Language Processing, pages 10527–10536. Zhaocheng Zhu, Zuobai Zhang, Louis-Pascal A. C. Xhonneux, and Jian Tang. 2021. Neural bellman-ford networks: A general graph neural network framework for link prediction. In Advances in Neural Information Processing Systems, pages 29476–29490.
A
Training and Inference Algorithms
Algorithm 1 and Algorithm 2 describe the RADD training procedure and Diff-Rerank inference, respectively. The training procedure jointly optimises the retriever and denoiser via three losses: KGE margin loss Lkge , bidirectional denoising crossentropy Ldiff , and temperature-scaled distillation Ldistill from the retriever’s soft rankings to the denoiser. The retriever is frozen after epoch τf to provide a stationary distillation teacher. At inference,
Algorithm 1 RADD Training (one epoch, one batch) Require: Mini-batch B = {(hi , ri , ti )}, current epoch e, freeze threshold τf , head-loss weight λh , distillation temperature τ , weight λd 1: // Step 1: Retriever forward 2: for each (h, r, t) ∈ B do Compute relation-gated representation 3: (joint) e via Eq. (1) 4: Score all entities and sample adversarial negatives 5: Accumulate Lkge (margin + adversarial) 6: end for 7: if e ≤ τf then 8: Update θret ← θret − η ∇Lkge ▷ retriever frozen after epoch τf 9: end if 10: // Step 2: Denoiser forward (bidirectional) 11: for each (h, r, t) ∈ B do 12: Sample t′ ∼ Uniform{1, . . . , T } 13: Tail prediction: corrupt x0 =t to xt′ ; set (joint) (emb) ch = [eh ; r; ext′ ; ϕ(t′ ); dtail ] 14: Ltail ← CE pθ (x0 | ch ), t Head prediction: corrupt x0 =h to xt′ ; set 15: (joint) (emb) ct = [et ; r; ext′ ; ϕ(t′ ); dhead ] 16: Lhead ← CE pθ (x0 | ct ), h 17: Ldiff ← Ltail + λh Lhead (Eq. 5) 18: Ldistill ← KL pT (C) ∥ pD (C) (Eq. 7) 19: end for 20: Update θden ← θden − η ∇(Ldiff + λd Ldistill )
Diff-Rerank enforces a strict two-stage decomposition: the retriever commits to a hard top-K shortlist, and the denoiser reranks exclusively within it.
B
Traditional KGC Results
To assess whether RADD’s advantage is rooted in its architecture or merely in its access to multimodal signals, Table 5 evaluates RADD in structure-only mode on two conventional KGC benchmarks where no visual or textual features are available. On FB15K-237, RADD reaches 34.80 MRR, surpassing the strongest classical baseline RotatE by +1.2, with consistent improvements across all metrics. On NELL995, the margin is substantially wider: RADD achieves 48.20 MRR, exceeding the best baseline ComplEx by +11.3 and raising H@1 by +14.4. These results confirm that RADD’s core benefit is architectural and modality-agnostic.
Algorithm 2 Diff-Rerank Inference Require: Query (h, r, ?) or (?, r, t), shortlist size K, diffusion steps T Ensure: Ranked candidate list 1: // Stage 1: Retrieval 2: Compute retriever score s(e) for all e ∈ E 3: S ← top-K {s(e) : e ∈ E} ▷ hard shortlist 4: // Stage 2: Denoising rerank 5: Set conditioning context c from query 6: Initialise xT = [MASK] 7: for t = T, T −1, . . . , 1 do 8: Compute posterior pθ (x0 | xt , c) restricted to S 9: Sample xt−1 ∼ pθ (xt−1 | xt , c) 10: end for 11: Score each ei ∈ S by pθ (ei | x1 , c) 12: return S reranked by denoiser scores Table 5: Results on two traditional unimodal KGC datasets. RADD operates in structure-only mode. Best in bold, second-best underlined. FB15K-237
Model
NELL995
MRR H@1 H@3 H@10 MRR H@1 H@3 H@10 TransE DistMult ComplEx RotatE
33.10 28.40 30.80 33.60
23.40 22.30 22.00 24.00
36.90 31.00 33.70 37.40
52.60 44.70 48.60 53.00
23.90 7.90 34.80 34.80 26.80 38.50 36.90 28.90 40.50 35.80 26.90 40.40
50.80 50.00 51.80 52.60
RADD
34.80 24.80 38.00
56.20
48.20 43.30 51.10
55.90
The contrast between the two datasets is informative. The smaller but uniform gain on FB15K-237 indicates that retrieve-then-rerank remains useful even when strong geometric baselines already provide competitive candidate ordering. The much larger gain on NELL995 suggests that shortlistlevel denoising becomes more valuable when structural retrieval leaves greater residual ambiguity. Together with the RADD-S results in Table 2, these experiments confirm that the retrieve-then-rerank decomposition is the primary source of RADD’s gains: it delivers consistent improvements even without multimodal evidence, and multimodal signals provide a further additive boost on top of that architectural foundation.
C
Case Study
Table 6 reports the rank of the correct entity e∗ for eight representative queries across DB15K and MKG-W, covering both head and tail prediction. RADD achieves rank 1 in every case, while neither single-module baseline does so consistently.
Table 6: Filtered rank of the correct entity for representative DB15K and MKG-W queries. Ret. = retrieval-only, Den. = denoiser-only, RADD = full model. Query
Correct Entity
Ret. Den. RADD
DB15K (?, type, Island) (?, country, United_States) (Larry Klein, spouse, ?) (Stuyvesant HS, grades, ?)
K. Balachander Huntington WV Joni_Mitchell Ninth_grade
(?, P495, Q142) (?, P31, Q11424) (Q1235040, P17, ?) (Q207176, P131, ?)
Q15961525 Q897112 Q145 Q25
506 1066 462 387 23 2 58 2
1 1 1 1
241 236 146 5
1 1 1 1
MKG-W 360 747 3 2
Head prediction: both modules are individually necessary. In all four head queries, retrievalonly and denoiser-only both produce low ranks (236–1066). Retrieval-only fails because KGE scoring over the full entity vocabulary cannot separate e∗ from hundreds of structurally similar head candidates; for instance, DB15K’s rdf:type/Island relation has 7638 confusable entities, creating a dense embedding neighbourhood that global similarity cannot resolve. Denoiser-only fails for a complementary reason: without a retrieval shortlist to bound its scope, the correct head entity is too weakly signalled to emerge at rank 1. RADD resolves the impasse: the retriever elevates e∗ into the top-K shortlist, and the denoiser exploits shortlistconditioned context to distinguish e∗ from the remaining candidates. Tail prediction: each module contributes complementary precision. Retrieval-only already places e∗ at moderate rank (2–58 on DB15K; 5– 146 on MKG-W), and denoiser-only reaches nearrank-1 in several cases (rank 2–3). Yet neither consistently achieves rank 1. RADD does so in all four cases: the retriever ensures e∗ is included in the shortlist, and the denoiser elevates it above the retriever’s residual errors, confirming that finegrained shortlist conditioning, not denoising power alone, is the decisive factor. Implication. These cases provide direct instancelevel evidence for RADD’s core hypothesis: effective KGC requires recall-oriented retrieval and precision-oriented shortlist denoising. Consistent rank-1 outcomes across diverse queries, two datasets, and both prediction directions confirm that role specialization is the structural source of RADD’s aggregate advantages.