RADD: Retrieval-Anchored Discrete Diffusion for Multi-Modal Knowledge Graph Completion

arXiv:2604.25693 · cs.AI · Submitted 2026-04-28 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "RADD: Retrieval-Anchored Discrete Diffusion for Multi-Modal Knowledge Graph Completion".

Jane: The gist: RADD proposes a Retrieval-Augmented Discrete Diffusion framework to decouple retrieval and reranking for multi-modal knowledge graph completion,

Tom: First, who's behind it and why it matters.

Title and authors: Jane: Let’s get into what they actually propose in this Retrieval-Anchored Discrete Diffusion for Multi-Modal Knowledge Graph Completion paper. They introduce this RADD framework which is specifically designed to tackle that decoupling problem we talked about before >

Lu: Essentially, the paper argues that the central bottleneck in MMKGC models is decision allocation because it requires two incompatible inductive biases, and RADD assigns each of those demands to its naturally suited module >

Tom: That means one bias wants global high recall over a large candidate space, and another wants fine-grained local sharpness for prediction, and these are two fundamentally different demands that a single scorer struggles with >

Jane: And RADD proposes assigning each of those objectives to the retriever and the denoiser so they can operate at their full potential without fighting each other >

Lalam: The retrieval-aware multimodal knowledge graph embedding KGE retriever serves as both global retriever and distillation teacher, while the conditional discrete denoiser performs shortlistlevel entity-identity generation for reranking >

Meng: So, in simple terms, the retrieval component finds the candidates and provides context for the denoiser to make its final choice over those candidates before it commits to a prediction >

Tom: Right. And then during training they combine KGE supervision, denoising cross-entropy loss, and temperature-scaled distillation from the retriever to the denoiser >

Jane: Which means they’re using a teacher-student setup to transfer information from the retriever's soft rankings into the denoiser’s learning process so it learns what a good candidate looks like >

The paper's summary: Lu: One major improvement is that they use teacher-student distillation to align the denoiser distribution with the frozen retriever’s soft rankings via KL divergence over a candidate subset C >

Meng: This is clever because it seeds the denoiser with a concentrated prior over plausible candidates before shortlist-level discrimination even begins, which should stabilize training >

Lalam: It’s like giving the denoiser a head start by showing it what's likely to be correct based on what the retriever already found, making its job of picking the best candidate much easier >

Tom: And inference is controlled by Diff-Rerank, which enforces that the retriever commits to a hard top-K shortlist, and then the denoiser reranks exclusively within it >

Jane: That means recall is a strict requirement for precision, ensuring that you get recall before you get local precision in every single prediction made at test time >

The paper's improvements: Lu: This method also incorporates explicit head-direction supervision in the bidirectional training, which is important because head entities follow a more skewed frequency distribution, making head prediction consistently harder >

Meng: That asymmetry is something we see often in these tasks, so having an asymmetric loss to compensate for that imbalance seems like a solid way to tackle the difficulty of predicting the main entity first >

Lalam: The paper also uses bidirectional training for both head and tail recovery with this asymmetric loss, and it shows that RADD reduces the structural asymmetry between head and tail prediction compared to other methods >

Tom: That’s interesting because it suggests that simply training on the tail isn't enough; you need to actively supervise the head prediction direction too >

Conclusion: Jane: So, to wrap up this Retrieval-Anchored Discrete Diffusion for Multi-Modal Knowledge Graph Completion paper: it’s a framework that decouples retrieval and reranking using a relation-aware multimodal KGE retriever for search and a conditional discrete denoiser for reranking >

Lu: The main implication is that this separation allows these two components to operate independently, solving the core bottleneck of decision allocation in MMKGC models by letting them work at their full potential >

Meng: For us in production, it means we have a clearer view of what parts are doing what and how to scale them independently for better reliability when dealing with both high recall and precision simultaneously >

Lalam: The paper is called RADD: Retrieval-Anchored Discrete Diffusion for Multi-Modal Knowledge Graph Completion, and the results show it ranks first in ten out of twelve dataset-metric combinations across DB15K, MKG-W, and MKG-Y >

Tom: It’s a lot of performance across those benchmarks where it improves MRR over the strongest external baselines by nine point five eight, nine point zero four, and zero point eight three points respectively >

Jane: The overall message is that this architecture works well because it effectively separates the search from the prediction step to get better results >

Lu: RADD is a strong framework for how we approach complex tasks like MMKGC by combining these two different ways of thinking about how to structure those components >

Meng: It gives us a clear path on how to build systems that are reliable when you have to manage both high recall and precision simultaneously without one component completely dictating the other >

Lalam: So, RADD is a strong framework for how we approach complex tasks like MMKGC by combining these two different ways of thinking.

Beihang University

cs.AI

Submitted: 2026-04-28

Updated: 2026-10-19

Comments: 16 pages, 3 figures, 8 tables

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 82/100

The gist: The gist: RADD proposes a Retrieval-Augmented Discrete Diffusion framework to decouple retrieval and reranking for multi-modal knowledge graph completion, achieving state-of-the-art performance by

Key concepts

Relation-Aware Multimodal Retriever: This component searches the entire entity set using fused embeddings from structural, visual, and textual modalities. It generates a shortlist of candidates and provides context to help the denoiser resolve ambiguities at the shortlist level, ensuring high global recall.
Conditional Discrete Denoiser: The denoiser takes a concatenated vector of context and query information to output logits over every possible entity in the vocabulary. It functions as a local scorer that recovers specific entity identities by modeling the reverse process of diffusion within the retrieved candidate set.
Diff-Rerank Inference Strategy: During inference, RADD enforces a strict two-stage process: first, the retriever selects a fixed top-K shortlist. Second, the denoiser only scores entities within this shortlist. This decomposition ensures that retrieval focuses on broad coverage and denoising focuses on fine-grained ranking within plausible candidates.

Terminology

Knowledge Graph — 3× in this paper · explained in 8 episodes
Distillation — 2× in this paper · explained in 2 episodes
Ablation — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Core — 1× in this paper · explained in 2 episodes
Retrieval-Augmented Discrete Diffusion — 1× in this paper

Summary

The gist: RADD proposes a Retrieval-Augmented Discrete Diffusion framework to decouple retrieval and reranking for multi-modal knowledge graph completion, achieving state-of-the-art performance by assigning global recall to a KGE retriever and local precision to a conditional discrete denoiser.

Decoupling Objectives

The paper argues that the central bottleneck in MMKGC is decision allocation, which requires two incompatible inductive biases: Global recall requires high recall over large candidate space <ref:2604.25693#pg2> and Prediction requires fine-grained local sharpness <ref:2604.25693#pg2>. Conventional single-scorer pipelines struggle to handle both objectives simultaneously because one model cannot process two incompatible objectives <ref:2604.25693#pg2>. RADD addresses this by assigning each objective to its naturally suited module for operating at their full potential <ref:2604.25693#pg2>.

Framework Architecture

RADD integrates a relation-aware multimodal KGE retriever and a conditional discrete denoiser, structured as follows:

Relation-Aware Multimodal Retriever: This component 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 <ref:2604.25693#pg5>. It constructs a relationaware joint embedding e(joint)r by fusing three modality-specific embeddings (structural e(s), visual e(v), textual e(t)) through a relation gate <ref:2604.25693#pg8>.
Conditional Discrete Denoiser: This component recovers entity identities over the full vocabulary, and it takes as input the concatenated vector x˜ = ch; r; xt; ϕ(t); d <ref:2604.25693#pg9>. The denoiser outputs logits over all E entities, reflecting the reverse model: pθden (x0 = e x˜) = softmax(fθden(x˜))e <ref:2604.25693#pg9>.

Training and Distillation

The two components are integrated during training through a teacher-student distillation mechanism <ref:2604.25693#pg2>. The training objectives combine KGE supervision, denoising cross-entropy, and temperature-scaled distillation from the retriever to the denoiser <ref:2604.25693#pg2>. Specifically, Ldistill aligns the denoiser distribution pD (student) with the frozen retriever’s soft rankings pT (teacher) via KL divergence over a candidate subset C <ref:2604.25693#pg2>. The full training objective is L = Lkge + Ldiff + λdLdistill, where 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 <ref:2604.25693#pg2>.

Inference Strategy

At inference, RADD employs a single strategy called Diff-Rerank which enforces the decomposition: the retriever commits to a hard topK shortlist and the denoiser reranks exclusively within it <ref:2604.25693#pg2>. The process involves two stages:

Global candidate retrieval: The retriever first computes KGE scores over the full entity set and forms a shortlist CK(q) = TopKe∈E skge(e q) <ref:2604.25693#pg9>.
One-pass diffusion scoring: The denoiser is then queried with xT = [MASK] and the same query context to produce denoiser log-scores ldiff(e q) = log pθden (x0 = e xT, q) <ref:2604.25693#pg9>.

The final Diff-Rerank score is Sdr(e q) = (ldiff(e q), e ∈ CK(q), −M, e /∈ CK(q)> <ref:2604.25693#pg9>.

Performance and Contributions

Experiments on three MMKGC benchmarks show that RADD achieves the best performance and consistent gains over strong unimodal, multimodal, and LLM-based baselines <ref:2604.25693#pg2>. RADD exceeds its retrieveronly/denoiser-only variants by 16.32/13.59 MRR on DB15K and 10.40/10.22 on MKG-W <ref:2604.25693#pg9>. Furthermore, RADD leads all baselines across three benchmarks, demonstrating strong overall performance <ref:2604.25693#pg2>. The framework's bidirectional training covers head and tail recovery, while retriever-to-denoiser distillation transfers candidate priors that guide diffusion toward plausible entities <ref:2604.25693#pg2>.

Key Findings on Directionality

Head prediction shows a clear directional advantage, raising head MRR from the strongest reported baseline of 26.23 to 48.39 on DB15K, a gain of 22.16 points <ref:2604.25693#pg9>. RADD reduces the structural asymmetry between head and tail prediction, narrowing the head-to-tail MRR ratios from MoMoK's 1.97 and 2.79 on DB15K/MKG-W to 1.03 and 1.55, respectively <ref:2604.25693#pg9>. Explicit head-direction supervision is established as the most influential component beyond tail-only diffusion training, causing the largest degradation in every metric on every dataset <ref:2604.25693#pg9>.

Ablation Study Insights

Ablation studies confirm that The retriever and denoiser make complementary contributions <ref:2604.25693#pg2>. Removing the denoiser/retriever lowers MRR by 16.32/13.59 points on DB15K, 10.40/10.22 points on MKG-W, and 0.66/0.96 points on MKG-Y <ref:2604.25693#pg9>. Removing the denoiser w/o Denoising lowers MRR by 9.85/10.53 points on DB15K and MKG-W, confirming their complementary roles <ref:2604.25693#pg9>.

Hyperparameter Sensitivity

The model is broadly robust across tested ranges, with noticeable sensitivity confined to shortlist construction and KGE optimization <ref:2604.25693#pg9>. The Top-K parameter has the largest effect on performance, with an MRR range of 4.4 pp: K = 128 focuses denoising on high-confidence candidates, whereas K > 256 introduces weaker candidates <ref:2604.25693#pg9>. Diffusion-specific parameters are particularly robust, with MRR varying by less than 0.1 pp across positive weights λd ∈ [0.1, 0.9] and by only 0.2 pp over T ∈ [20,..., 150] <ref:2604.25693#pg9>.

Limitations

RADD has three main limitations: First, DiffRerank imposes a hard shortlist constraint: if the target entity is absent from the retriever’s top-K candidates, the denoiser cannot recover it <ref:2604.25693#pg2>. Second, inference requires both full-entity KGE scoring and a full-vocabulary denoiser projection, introducing additional computation relative to a single scorer <ref:2604.25693#pg2>. Finally, the multimodal embeddings are pre-extracted and fixed, which limits task-specific adaptation of the visual and textual encoders <ref:2604.25693#pg2>.

Conclusion

To address the conflicting demands of global highrecall search and local fine-grained discrimination, RADD proposes a retrieval-augmented discrete diffusion framework for MMKGC tasks <ref:2604.25693#pg2>. The retriever supplies candidates and distillation priors, while the denoiser models head and tail recovery <ref:2604.25693#pg2>. Across DB15K, MKG-W, and MKG-Y, RADD ranks first in 10 of 12 dataset-metric combinations and improves MRR over the strongest external baselines by 9.58, 9.04, and 0.83 points <ref:2604.25693#pg2>. The results validate the architecture’s effectiveness and the contribution of each module <ref:2604.25693#pg2>.

Algorithm Summary

The training algorithm involves two stages: Stage 1 is full-vocabulary retrieval where modalities are fused via Eq.

Improvements for AI systems

Retrieval-Augmented Discrete Diffusion (RADD) framework for Multi-Modal Knowledge Graph Completion (MMKGC). This system decouples retrieval and reranking, allowing a relation-aware multimodal KGE retriever serves as both global retriever and distillation teacher while a conditional discrete denoiser performs shortlistlevel entity-identity generation for reranking.
Explicit head-direction supervision in Bidirectional Diffusion Training. The framework uses an asymmetric loss, upweighting the head loss to compensate for structural imbalance, which is crucial because head entities follow a more skewed frequency distribution, making head prediction consistently harder.
Hardgated candidate support during inference via Diff-Rerank. The system enforces the decomposition by ensuring the retriever commits to a hard topK shortlist and that the denoiser reranks exclusively within it, which is enforced by the score formula: Sdr(e q) = (ldiff(e q), e ∈ CK(q), −M, e /∈ CK(q).
Bidirectional training for head and tail recovery. The model trains both directions with an asymmetric loss, as the paper notes that relations typically impose stronger type constraints on tails, and head entities follow a more skewed frequency distribution, leading to RADD reduces these ratios to 1.03 and 1.55 compared to MoMoK on MKG-W.
Knowledge distillation for prior transfer during training. The system uses a distillation term where Ldistill aligns the denoiser distribution pD (student) with the frozen retriever’s soft rankings pT (teacher) via temperature-scaled KL divergence over a candidate subset C, which seeds the denoiser with a concentrated prior over plausible candidates before shortlist-level discrimination begins.