{"id":"c8404312-b6d5-48f1-b9bb-1041981b4719","arxiv_id":"2501.15393","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A diffusion-based hierarchical negative sampling framework with hardness-adaptive training improves multimodal knowledge graph completion on three benchmarks.","lead":"This paper proposes a diffusion-model-based negative sampling method for multimodal knowledge graph completion, generating candidate negative triples at different hardness levels and training the graph model with adaptive margins. The authors report improved link prediction scores on three benchmark datasets, but key details of the adaptive margin are not specified.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Off-manifold generated negative embeddings: DHNS's gains may not transfer to real-entity ranking unless generated vectors are shown to lie near the entity embedding manifold.","rationale":"The most load-bearing condition for the abstract's claim is that the synthetic embeddings generated by DiffHEG are usable as negative triples. Sec. 3.2 defines hardness by time step and Eqs. 12-13 feed the generated vectors into the KGE loss, yet no experiment checks whether those vectors are close to real entity embeddings or whether the small-t negatives that are described as closer to positives are ever actual positives. If this condition fails, the training loss minimizes against off-manifold points, so improved MRR on the three benchmarks would not demonstrate better negative sampling. The reader's conditional verdict already identifies this as the weakest assumption, and I agree. I considered the copied baseline tables and the unspecified forms of gamma_t and w(negt); these are reproducibility threats, but they are less central than whether the core generative signal is valid. The proposed projection ablation gives a decisive, low-cost check. If the ablation preserves performance, the concern is answered; if not, the central claim would need to be revised. I therefore see no reason to change the verdict from conditional.","tokens_in":13925,"tokens_out":10658,"duration_ms":101112,"concrete_test":"On DB15K with the released code, add one post-processing step before Eq. 12: replace each generated negative embedding with the embedding of its nearest real entity under cosine similarity, using the current entity embedding matrix, and keep all other hyperparameters and losses unchanged. If MRR and H10 remain close to the reported 34.36 and 53.72, the generated negatives are effectively on-manifold and the concern is resolved; if performance drops substantially, the reported gains rely on off-manifold vectors that do not correspond to real entities. Also record how often the nearest entity is the true tail of the positive triple, which directly measures the false-negative rate of the small-t negatives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the reverse-diffusion outputs from Eqs. 5-6, scored in Eq. 12 and optimized in Eq. 13, are valid negative triples. The paper never establishes that these vectors lie on the manifold of real entity embeddings or correspond to any entity in E; conversely, at small t they are explicitly described as closer to positives (Sec. 3.2, hardness inversely proportional to t), and no false-negative or nearest-entity check is reported. Because the hardness-adaptive loss (Eq. 13) pushes down the score of a synthetic vector with no entity identity, the KGE model is trained to assign low scores to arbitrary points in embedding space, while inference ranks only real entity embeddings. If the generated vectors are off-manifold, this training signal may not transfer, and the gains in Tables 2-4 could be an artifact of the generative module rather than evidence of better negative sampling. This is the critical link in the central claim because the method's stated mechanism depends on the generated negatives being semantically meaningful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes DHNS, a diffusion-based negative sampling scheme for multimodal knowledge graph completion. DHNS trains a conditional diffusion model (DiffHEG) to generate tail (or head) entity embeddings from Gaussian noise conditioned on the counterpart entity, relation, and multimodal features; hardness of the generated negative triples is controlled by stopping the reverse diffusion at different time steps. The generated negatives are then used in a Negative Triple-Adaptive Training (NTAT) loss with hardness-adaptive margins. Experiments on DB15K, MKG-W, and MKG-Y report improvements over existing unimodal, multimodal, and negative-sampling baselines, with ablations showing the contribution of each module.","tokens_in":14157,"tokens_out":9612,"duration_ms":75467,"significance":"If the claims are substantiated, DHNS would be a novel and useful contribution to MMKGC: it is, to the best of the paper's claim, the first diffusion-based negative embedding generation approach specific to multimodal knowledge graph completion, and the idea of controlling hardness via diffusion time steps is intuitive. The paper also proposes a promising adaptive-margin training scheme. However, the current manuscript leaves key definitions underspecified and lacks empirical validation under controlled comparisons, so the significance of the work is not yet demonstrated.","major_comments":[{"comment":"The reverse diffusion update rule in Eqs. (4)-(6) appears to use β_t in the denominator, but β_t is defined as 1 - α_t in Eq. (1). For small t, α_t is typically near 1, so β_t is near 0, making the coefficient 1/√β_t in Eq. (5) extremely large and the update numerically unstable. In the standard DDPM posterior, the coefficient is 1/√α_t or 1/√(1-β_t), where the noise variance is small for small t. Please re-derive the reverse process from the forward process in Eqs. (1)-(2) and correct the equations, or clarify if β_t is intended to denote the noise variance rather than 1 - α_t.","section":"3.2, Eqs. (4)-(6)"},{"comment":"The hardness-adaptive margin γ_t and the negative weights w(negt) in Eq. (13) are never explicitly defined. The text states that γ_t is adaptive to the hardness level HL(t) and that weights are higher for embeddings closer to the halfway point of the diffusion process, but no functional form, schedule, or update rule is provided. This makes the proposed NTAT loss non-reproducible and prevents independent verification of the adaptive-margin mechanism. Please specify γ_t and w(negt) in closed form or with an exact algorithmic description.","section":"3.3, Eq. (13)"},{"comment":"The framework never checks whether the generated negative embeddings lie near the manifold of real entity embeddings or whether the composed triples (h, r, neg_t) are already true triples. The generated negatives are arbitrary points in the embedding space, and at small t they are explicitly described as being closer to positives, increasing the risk of false negatives. Consequently, the loss in Eq. (13) may train the KGE model to assign low scores to off-manifold vectors, while inference only ranks real entities, so the improvements reported in Tables 2-4 may not transfer to the actual ranking task. Please provide quantitative evidence about the relation of generated embeddings to real entities (e.g., nearest-neighbor statistics) and report the false-negative rate among generated negatives.","section":"3.2-3.3, Eqs. (5)-(6), (12)-(13)"},{"comment":"Baseline results are copied from prior publications ([38] for Tables 2-3 and [34] for Table 4) rather than re-run in the same experimental framework, and no error bars or significance tests are reported anywhere. The claimed consistent improvements over baselines therefore rest on potentially incompatible setups and are not statistically validated. Please re-run the comparisons under identical conditions, or at least report multiple-run statistics (mean and standard deviation) and significance tests for the reported metrics.","section":"4.1, Tables 2-4"}],"minor_comments":[{"comment":"The symbol t is used both for the tail entity in a triple and for the diffusion time step, causing ambiguity in the loss definition; using e_t or a distinct subscript for the tail entity would clarify the equation.","section":"Eq. (13)"},{"comment":"In the opening of Section 3.3, \"HTAT\" appears to be a typo for \"NTAT\", the negative triple-adaptive training mechanism.","section":"3.3"},{"comment":"The word \"Hardmard\" should be \"Hadamard\" in the description of the multiplication operation in Eq. (7).","section":"3.2, Eq. (7)"},{"comment":"The phrase \"As we can be concerned\" is awkward and should be rephrased, for example as \"To the best of our knowledge\".","section":"Contributions, Section 1"},{"comment":"The claim that DHNS \"consistently and significantly outperforms\" other NS strategies is not fully supported by Table 4: for TransE on MKG-W, the H10 of DHNS (46.85) is lower than that of MMRNS (47.11), and several other differences are small. Please either soften the claim or discuss the discrepancies in the text.","section":"4.2, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper tackles a relevant and timely problem, and the diffusion-based generation of negative embeddings with hardness control is a promising idea. However, the technical formulation needs to be corrected and completed (especially the reverse-diffusion equations and the definitions of γ_t and w(negt)), and the empirical evaluation should be strengthened with re-run baselines or error bars. The off-manifold concern is central to the validity of the method and must be addressed with concrete experiments. The manuscript also contains several typos and notational inconsistencies that should be cleaned up. I recommend major revision rather than rejection, since the core idea is plausible and the issues appear addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: DHNS is a plausible and genuinely new combination—DDPM-based generation of negative entity embeddings, conditioned on multimodal features, with hardness controlled by the diffusion time step and a margin that adapts to that hardness. It is the first diffusion-based negative sampler aimed specifically at multimodal KGC, it is pluggable across KGE scorers, and the ablation study suggests both the generation module and the adaptive loss contribute. Code is released, which is real evidence of good faith. If I were working on negative sampling, I would read it and likely cite it.\n\nThe soft spots are mostly about reporting, not the core idea. First, the two most important quantities in the hardness-adaptive loss, gamma_t and w(negt), are never given functional forms. The paper says they are adaptive and that halfway timesteps get higher weights, but equations 13 and 15 are not reproducible without those definitions. That is the largest single issue. Second, the stress-test concern about off-manifold generated negatives is legitimate: the paper never checks whether the reverse-diffusion outputs lie near actual entity embeddings, and at small t they are explicitly closer to positives, raising false-negative risk. The reported gains could partly come from training the scorer to assign low scores to arbitrary points in embedding space. A nearest-entity check or a false-negative rate would address this. Third, baseline numbers are copied from prior papers rather than re-run, and there are no error bars or significance tests, so the consistent HR@10 and MRR gains are plausible but not rigorously established. Fourth, DMNS and KGDM, the two diffusion-based negative samplers the authors themselves cite, are not included in the comparisons.\n\nNone of this is fatal. The central mechanism is coherent, the ablation is consistent with the claims, and the open problems are addressable in revision. The paper deserves a serious referee, but the version in front of me should not be accepted as-is. I would send it to review with a request for the missing definitions, a sanity check on generated negatives, and at least a few re-run baselines. The math is standard DDPM machinery, no obvious circularity, and the self-citation pattern is clean. It is a solid within-subfield contribution, not a breakthrough, and that is fine.","headline":"A useful new combination—diffusion-based negative sampling for multimodal KGC—but the experimental reporting is too thin to fully support the central claim.","tokens_in":14651,"tokens_out":1706,"would_cite":true,"duration_ms":17909,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes DHNS, a diffusion-based scheme that generates negative triples for multimodal knowledge graph completion, and reports it outperforms state-of-the-art baselines on three benchmarks.","keywords":["multimodal knowledge graph completion","negative sampling","diffusion model","denoising diffusion probabilistic model","hierarchical embedding generation","hardness-adaptive loss","knowledge graph embedding","MMKGC"],"falsifier":"One decisive check is to score every generated negative triple against the KG: if a large share of generated negatives at small time steps are either already true triples or lie far outside the convex hull of real entity embeddings, the training signal is corrupted and the claimed advantage should shrink or vanish.","tokens_in":13707,"feed_emoji":"🌀","tokens_out":10818,"duration_ms":78159,"temperature":0.7,"pith_summary":"The paper proposes DHNS, a negative-sampling scheme for multimodal knowledge graph completion (MMKGC) that replaces the usual step of sampling a corrupted entity with a diffusion model that generates new entity embeddings conditioned on the head entity, relation, and visual, textual, and structural features. The generated negatives are organized by hardness level, controlled by the diffusion time step, and the training loss adapts its margin to that hardness. The authors claim that this yields higher-quality and more diverse negative triples than random, adversarial, and prior multimodal negative-sampling strategies, and report consistent improvements over state-of-the-art baselines on DB15K, MKG-W, and MKG-Y. If true, DHNS is a pluggable training component that could improve any MMKGC model without changing its score function.","feed_headline":"Diffusion model generates hard negatives for multimodal KGC","feed_subtitle":"Pluggable diffusion negative sampling lifts link prediction across three MMKG benchmarks.","key_machinery":"The load-bearing machinery is DiffHEG (Diffusion-based Hierarchical Embedding Generation), a denoising diffusion probabilistic model that reverses a forward noising Markov chain to produce synthetic entity embeddings. Conditioning is multimodal, combining structural, visual, and textual features of the head entity and the relation through one of three interaction mechanisms (Hadamard, bilinear, or additive) matched to the KGE model. Hardness is controlled by the reverse-diffusion time step $t$, with hardness level $\\mathrm{HL}(\\hat{x}_t) \\propto 1/t$, so early steps give near-positive \"hard\" negatives. These generated negatives are consumed by NTAT (Negative Triple-Adaptive Training), whose Hardness-Adaptive Loss uses a margin $\\gamma_t$ that shrinks as the negative gets harder, and whose multimodal joint scoring averages the KGE score over the three modality-specific generated embeddings.","core_discovery":"The central claim is that negative triples for MMKGC should be generated, not sampled: a conditional denoising diffusion model (DiffHEG) starts from pure Gaussian noise and, guided by the head entity and relation embeddings plus structural, textual, and visual features, produces synthetic tail-entity embeddings. Time steps $T/20$, $T/10$, $T/5$, and $T/2$ yield negatives of increasing easiness, with hardness inversely proportional to the time step. These synthesized negatives are scored by the KGE model's own score function through multimodal joint scoring, and a Hardness-Adaptive Loss assigns smaller margins to harder negatives. The paper argues that direct generation combined with adaptive margins gives a more robust training signal than sampling-based approaches, and reports that DHNS integrated with RotatE achieves the best or second-best MRR, Hits@1, Hits@3, and Hits@10 across DB15K, MKG-W, and MKG-Y.","pith_inferences":["A testable extension would be to filter generated negatives that already exist as true triples in the KG, or to project generated embeddings onto the nearest real entity embedding; the paper does not report whether such false negatives occur and how often.","Because hardness is tied to the diffusion time step, the same generator could support curriculum learning, starting training with easy negatives and gradually increasing difficulty, which the paper does not explore.","The conditioning mechanisms are matched to the algebraic structure of the base KGE model, suggesting a general recipe: the interaction used for conditioning should align with the score function's composition rule, a principle the paper illustrates but does not state explicitly."],"forward_implications":["DHNS can be plugged into existing KGE models such as TransE, DistMult, and RotatE to improve their MMKGC accuracy without altering their score functions.","Because the diffusion process can generate negatives of any hardness by choosing the time step, training can be balanced across easy and hard negatives instead of relying on a single fixed margin.","The margin adaptation in the Hardness-Adaptive Loss implies that the same KGE model can be trained more effectively on a mixture of negative difficulties, reducing sensitivity to the global margin hyperparameter.","The ablation results indicate that the DiffHEG generation module, rather than the training strategy, is the main source of gain; removing it causes the largest performance drops across all three datasets."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model (forward noise and reverse denoising, Eqs. 1-6) that DiffHEG adapts to generate entity embeddings.","marker":"[9]"},{"why":"MMRNS; the strongest multimodal negative-sampling baseline and the source of the same BEiT/SBERT feature preprocessing used for DHNS.","marker":"[34]"},{"why":"AdaMF; state-of-the-art MMKGC baseline and the source of the baseline numbers in the main tables and the feature-extraction pipeline.","marker":"[38]"},{"why":"RotatE; the KGE model DHNS is integrated with for the main results and the rotation interaction that determines the Hadamard conditioning in Eq. 7.","marker":"[25]"},{"why":"MMKG; provides the DB15K dataset and its image/text collection used in evaluation.","marker":"[15]"},{"why":"MMKRL; provides the MKG-W and MKG-Y datasets used in evaluation.","marker":"[17]"},{"why":"KBGAN; representative adversarial negative-sampling baseline that motivates the need to directly generate hard negatives rather than select pre-sampled ones.","marker":"[4]"},{"why":"MANS; modality-aware negative-sampling baseline that DHNS must beat to support the claim that direct diffusion generation outperforms sampling-based multimodal NS.","marker":"[37]"}],"fun_headline_variants":["Diffusion-based negative sampling lifts multimodal KGC","Synthetic hard negatives via diffusion for MMKG completion","Adaptive diffusion negatives improve MMKG link prediction","Hierarchical diffusion negatives outdo sampling in MMKGC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme rests on the assumption that a random vector denoised under the condition of a head entity and relation lands on a plausible entity embedding that the scoring model can treat as a sensible negative, even though nothing explicitly checks that the generated vector resembles a real entity or is not a true triple.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion-based negative sampling lifts multimodal KGC","Synthetic hard negatives via diffusion for MMKG completion","Adaptive diffusion negatives improve MMKG link prediction","Hierarchical diffusion negatives outdo sampling in MMKGC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1266,"prompt_tokens":966,"completion_tokens":300,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":582,"tokens_out":300,"duration_ms":3416,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:19:37.234892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive check is to score every generated negative triple against the KG: if a large share of generated negatives at small time steps are either already true triples or lie far outside the convex hull of real entity embeddings, the training signal is corrupted and the claimed advantage should shrink or vanish.","supporting_citations":[{"cited_title":"In: ACM MM","cited_arxiv_id":null,"evidence_quote":"MMRNS; the strongest multimodal negative-sampling baseline and the source of the same BEiT/SBERT feature preprocessing used for DHNS."},{"cited_title":"In: LREC-COLING","cited_arxiv_id":null,"evidence_quote":"AdaMF; state-of-the-art MMKGC baseline and the source of the baseline numbers in the main tables and the feature-extraction pipeline."},{"cited_title":"In: ICLR (2019)","cited_arxiv_id":null,"evidence_quote":"RotatE; the KGE model DHNS is integrated with for the main results and the rotation interaction that determines the Hadamard conditioning in Eq. 7."},{"cited_title":"In: Hitzler, P., Fernández, M., Janowicz, K., Zaveri, A., Gray, A.J., Lopez, V., Haller, A., Hammar, K","cited_arxiv_id":null,"evidence_quote":"MMKG; provides the DB15K dataset and its image/text collection used in evaluation."},{"cited_title":"Applied Intelligence52, 7480–7497 (2021)","cited_arxiv_id":null,"evidence_quote":"MMKRL; provides the MKG-W and MKG-Y datasets used in evaluation."},{"cited_title":"In: NAACL","cited_arxiv_id":null,"evidence_quote":"KBGAN; representative adversarial negative-sampling baseline that motivates the need to directly generate hard negatives rather than select pre-sampled ones."},{"cited_title":"In: IJCNN","cited_arxiv_id":null,"evidence_quote":"MANS; modality-aware negative-sampling baseline that DHNS must beat to support the claim that direct diffusion generation outperforms sampling-based multimodal NS."}],"review_version":1}