{"id":"5c2678fa-8179-445e-9d94-afe1373dace5","arxiv_id":"2507.20620","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"MoCME combines expert-network fusion weighted by estimated mutual information and entropy-based negative sampling, and reports state-of-the-art multi-modal knowledge graph completion on five benchmarks.","lead":"This paper proposes MoCME, a method for predicting missing facts in knowledge graphs by combining images, text, numbers, audio and video, weighting each source by how much unique information it adds. The authors report improved accuracy over existing methods on five benchmarks, but do not release code and omit a direct comparison to the method that inspired the design.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not testable as reported because MoMoK, the direct predecessor whose expert architecture MoCME extends, is omitted from every baseline table; adding MoMoK to Table 2 is the decisive check.","rationale":"The reader's weakest_assumption focuses on MINE and unspecified encoders; those are real reproducibility gaps, but they attack the mechanism rather than the claimed outcome. The most load-bearing issue is the evaluation design: a SOTA claim is only meaningful if the relevant existing methods are in the table. The paper explicitly credits MoMoK as the inspiration for its expert module, so MoMoK cannot be treated as an incidental omission; it is the strongest available control for whether the complementarity weighting adds anything. A test that adds MoMoK to Table 2 directly decides the abstract's central assertion. My recommendation is unchanged from the reader's REJECT: the omission, together with no code/data and no variance reporting, leaves the central empirical claim unverified. I partially agree with the reader: the stated weakest assumption (MINE stability) is not the one I would lead with, but the reader's rationale already calls out the missing MoMoK baseline.","tokens_in":20153,"tokens_out":5186,"duration_ms":56312,"concrete_test":"Obtain the official MoMoK (arXiv:2405.16869) implementation, or reimplement it faithfully, and run it under the paper's stated protocol: RotatE scoring, embedding dimension 256, Adam lr 1e-4, batch size 1024, 1000 epochs with early stopping on validation MRR (Section 4.1), on MKG-W, MKG-Y, DB15K, KVC16K and TIVA. Insert the resulting MRR/Hit@1/Hit@10 rows into Table 2. If MoMoK is within run-to-run variance, or outperforms MoCME on any dataset, then 'surpassing existing approaches' is not supported; if MoCME still leads by the reported margins, the SOTA claim survives this objection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim is empirical: MoCME 'consistently achieves the best overall performance' on five benchmarks (Section 4.4). The paper states in Section 3.2 that its CMoE design is 'inspired by MoMoK [54]', and MoMoK is cited in the references, but Section 4.3 and Table 2 list 19 baselines without MoMoK. MoMoK is the closest existing system: it also uses per-modality mixtures of knowledge experts for MMKGC and is the natural null hypothesis for a model whose contribution is a complementarity-weighting layer on top of that design. Without MoMoK in the comparison, the reported gains over NATIVE cannot establish SOTA relative to the most relevant prior work. The omission is not merely cosmetic: if MoMoK matches or exceeds MoCME, the central abstract claim fails. This is compounded by the absence of released code and the lack of error bars despite 'three independent runs', but those are secondary to the missing comparator.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MoCME, a multi-modal knowledge graph completion framework combining a Complementarity-guided Modality Knowledge Fusion (CMKF) module and an Entropy-guided Negative Sampling (EGNS) mechanism. CMKF uses per-modality expert networks and mutual information estimates to weight multi-view and multi-modal embeddings, while EGNS reweights negative samples by their binary entropy. The paper reports experiments on MKG-W, MKG-Y, DB15K, KVC16K, and TIVA and claims consistent state-of-the-art results over 19 baselines.","tokens_in":20339,"tokens_out":4636,"duration_ms":48546,"significance":"If the claimed results hold, the paper would make a useful empirical contribution to MMKGC by showing that complementarity-based weighting and entropy-guided negative sampling improve over plain fusion. Strengths include the clear problem framing, the inclusion of five datasets with varying modality coverage, and ablations of components, fusion strategies, backbones, expert counts, and thresholds. However, the paper's central SOTA claim is currently not sufficiently supported: the most relevant prior model is omitted from the comparisons, key components are underspecified, and no variability information is reported. The contribution is therefore better assessed as promising but requiring substantial revision rather than as an established result.","major_comments":[{"comment":"The paper states in §3.2 that the expert design is 'inspired by MoMoK [54]', yet MoMoK is not included among the 19 baselines in Table 2. Because MoMoK is the direct predecessor using per-modality mixtures of knowledge experts for MMKGC, the central claim in §4.4 that MoCME 'consistently achieves the best overall performance' cannot be evaluated without this comparison. Please add MoMoK to the baseline tables and report its MRR, Hit@1, and Hit@10 on all five datasets.","section":"§3.2, §4.3, Table 2"},{"comment":"The complementarity weights in Eqs. (3) and (4) depend on mutual information estimates obtained with MINE, but the manuscript never specifies the MINE architecture, its training procedure, or how estimates are aggregated across entities and batches. The citation [22] also does not appear to be the MINE paper. Since these weights are the core novelty of CMKF, the current description is not reproducible and the reported weights could be arbitrary. Please provide full details or replace the estimator with an explicitly defined, computable quantity.","section":"§3.2, Eqs. (2)-(4)"},{"comment":"For KVC16K and TIVA, which include audio and video modalities, the paper reports large gains over NativE, but no audio or video encoder is named anywhere in §4.1; only VGG16 and BERT are mentioned. Please specify the encoders, input representations, and preprocessing for audio and video, and confirm whether the same frozen-encoder pipeline applies to these modalities.","section":"§4.1, Table 1"},{"comment":"The description of the structural modality is internally contradictory: the text first says structure is 'encoded into token sequences and encode them using BERT', and then states that 'the final structural embedding eS is not directly extracted from a frozen encoder but rather initialized as a learnable parameter'. Please clarify which mechanism is actually used, because the two formulations lead to different models and affect the interpretation of the structural-modality results.","section":"§3.2"},{"comment":"All reported numbers are averages over three independent runs, but no standard deviations or significance tests are given. The headline gains over NativE on MKG-W and MKG-Y are only +1.21 and +1.33 MRR, which may be within run-to-run noise. Please report standard deviations and, where feasible, statistical significance of the differences.","section":"§4.1, Table 2"}],"minor_comments":[{"comment":"The text states that removing image features causes the largest MRR decrease (-0.57), but Table 3 shows text removal gives the largest decrease (-1.21). Please correct this contradiction.","section":"§4.5.1"},{"comment":"The threshold grid search is described as being reported in Figure 2, but the corresponding figure is Figure 3; the expert-count ablation is Figure 2.","section":"§4.5.5"},{"comment":"The claim that the optimal expert count tracks the modality count 'across all three datasets' is not supported, since only DB15K is described in that paragraph.","section":"§4.5.4"},{"comment":"The notation ωa_m · ve,m = Σ_i ... is confusing because ωa_m is a vector of weights; please rewrite with explicit index notation.","section":"Eq. (3)"},{"comment":"The reference list contains incomplete entries (e.g., [22] has no venue or year, [55] is listed only as '(No Title)'); please check all references.","section":"References"},{"comment":"Please state whether code will be released, as the method currently cannot be re-implemented from the text alone.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The omission of MoMoK is the decisive issue; if the authors cannot include that comparison, the SOTA claim should be withdrawn or substantially weakened. I would also check whether the mutual information estimator is actually the MINE method or a different estimator, since Eq. (2) writes a discrete sum while MINE is a neural lower bound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the empirical claim in the abstract is the whole ballgame, and the missing MoMoK baseline makes it unverifiable. The paper says in Section 3.2 that the CMoE design is “inspired by MoMoK [54],” and MoMoK is in the references, but it is absent from Table 2 and all comparisons. That is not a cosmetic gap; MoMoK is the natural null hypothesis for a model whose main contribution is a complementarity-weighting layer on top of that exact architecture. If MoMoK matches or beats MoCME, the SOTA statement collapses. The stress-test note is right on this point.\n\nWhat the paper does well: the core idea is clean. Using negative mutual information between expert views and between modalities to weight fusion is a sensible, well-motivated approach, and the entropy-based negative sampling is reasonable even if the ingredients are known. The ablation suite is genuinely more thorough than most MMKGC papers—components, fusion strategies, backbones, expert count, and threshold sensitivity are all covered. The reported gains over NativE are modest but consistent across five datasets, which is the right kind of claim to probe.\n\nWhere it goes soft, in proportion: the missing MoMoK is the load-bearing defect. Secondary but real: no code or data, no error bars despite “three independent runs,” MINE training details are unspecified, audio/video encoders are never named, and the structural modality description wavers between BERT-token sequences and a learnable embedding. There is also a small internal inconsistency—Section 4.5.5 says the threshold grid is in Figure 2, but Figure 2 is the expert-count ablation and Figure 3 is the heatmap. That is minor, but it adds friction when the authors are asking for trust.\n\nFinal judgment: this is a plausible incremental method with an untested central comparison. It deserves a serious referee and a clear request for revision: add MoMoK to Table 2, report standard deviations, specify MINE and the encoders, and release code. Do not desk-reject; the idea is testable and the ablations are useful. But the current version should not be accepted as-is.","headline":"The central SOTA claim is unverifiable as written because MoMoK—the direct predecessor MoCME is built on—is missing from every baseline table, so the paper needs major revision before its empirical headline can be trusted.","tokens_in":20891,"tokens_out":1594,"would_cite":false,"duration_ms":19823,"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":"MoCME claims that weighting each expert view and each modality by negative mutual information, then training with entropy-weighted hard negatives, produces state-of-the-art multi-modal knowledge graph completion on five benchmarks.","keywords":["multi-modal knowledge graph completion","modality complementarity","mixture of experts","mutual information","entropy-guided negative sampling","RotatE","link prediction"],"falsifier":"On a fixed dataset such as DB15K, rerun MoCME with the complementarity weights replaced by uniform random weights while keeping everything else identical; if MRR does not drop materially below 39.62, the complementarity signal is not what carries the reported result.","tokens_in":19889,"feed_emoji":"🧩","tokens_out":5828,"duration_ms":62392,"temperature":0.7,"pith_summary":"The paper tries to establish that multimodal knowledge graph completion improves when fusion weights reflect complementarity rather than attention or gating: each modality and each expert view is weighted by how little information it shares with its peers. It packages this as MoCME, built from a complementarity-guided fusion module and an entropy-guided negative-sampling loss. A careful reader should care because existing methods treat modalities as additive evidence, while this work treats redundancy itself as the signal to down-weight, which is a different principle. The paper reports that MoCME reaches the best MRR, Hit@1, and Hit@10 among all compared methods on five benchmark datasets, with the largest margins on datasets carrying four modalities.","feed_headline":"Complementarity-based fusion tops five multimodal KG benchmarks","feed_subtitle":"MoCME weights each modality by how little it overlaps with the others, then trains on entropy-selected hard negatives.","key_machinery":"The load-bearing object is the complementarity weight: for a set of views, the weight of view $i$ is $\\exp(-\\sum_{j\\neq i} I(v^{(i)}; v^{(j)})) / \\sum_{i'} \\exp(-\\sum_{j\\neq i'} I(v^{(i')}; v^{(j)}))$, using the mutual information neural estimator (MINE) to measure overlap; the same operation is applied again across modalities. It converts an information-theoretic notion of redundancy into concrete fusion coefficients, and the paper's argument is that this conversion is what distinguishes MoCME from concatenation, gating, and attention. The second mechanism is entropy-based negative sampling, which partitions negative triples by the binary entropy $\\mathcal{H} = -p\\log p - (1-p)\\log(1-p)$ of their predicted plausibility into easy, ambiguous, and hard buckets with weights satisfying $\\lambda_{\\text{easy}} < \\lambda_{\\text{amb}} < \\lambda_{\\text{hard}}$.","core_discovery":"The central claim is that entity representations become better for link prediction when each view and each modality is weighted by its non-redundancy, estimated as negative mutual information between distributions of expert outputs. Intra-modality, the framework fuses K expert views of one modality with softmax weights over negative pairwise mutual information; inter-modality, it fuses the resulting modality embeddings the same way. The fused embedding is scored with RotatE, and training uses a negative-sampling loss in which easy, ambiguous, and hard negatives receive three different weights derived from the binary entropy of each negative's score. The paper argues this hierarchy of complementarity weighting is what lets the model keep informative signals when modalities are missing or noisy, and the experiments are offered as evidence that the design pays off, especially on datasets with audio, video, text, and image modalities.","pith_inferences":["An implication the paper leaves implicit: the complementarity-weighting recipe is not tied to RotatE, so the same two-level mutual-information fusion could be grafted onto other knowledge graph scoring functions and tested as a general plug-in.","A testable extension the paper does not run: freeze the trained expert views and replace MINE with a cheaper closed-form correlation estimate; if MRR survives, the specific estimator is not the active ingredient, and the method would scale to larger graphs.","The paper's framing suggests a curriculum reading of the entropy sampler: thresholds could be annealed during training rather than fixed, which the paper does not explore but its own sensitivity analysis makes plausible."],"forward_implications":["If the central claim holds, state-of-the-art scores on all five benchmarks shift to a method whose fusion weights come from information overlap rather than learned attention, with the largest reported gains on the four-modality datasets TIVA and KVC16K.","A direct corollary of the ablations is that each modality contributes, and that removing the inter-modality complementarity layer costs more than removing any single modality, so cross-modal redundancy is the main carrier of the improvement.","The entropy-guided negative sampler implies that the same model trains better when easy negatives are down-weighted relative to ambiguous and hard ones, making negative-sample difficulty a tunable component of multimodal knowledge graph completion.","On datasets with only image and text, the gains over the previous best are smaller, so the claimed advantage is conditional on modality richness."],"supporting_citations":[{"why":"Supplies the RotatE scoring function that scores candidate triples and generates the negative-sample probabilities used by the entropy weighting.","marker":"[3]"},{"why":"Provides the MKG-Y and MKG-W benchmark datasets and the relation-enhanced negative-sampling baseline.","marker":"[13]"},{"why":"Sets the hyperparameter and baseline configuration the paper follows, and is the strongest comparison result it must beat.","marker":"[19]"},{"why":"Cited as the source of the mutual information neural estimator and of the principle that complementarity aids multimodal robustness.","marker":"[22]"},{"why":"Supplies the mixture-of-modality-experts design that MoCME adapts into per-modality expert views.","marker":"[54]"},{"why":"Defines the DB15K multimodal knowledge graph used for the three-modality experiments and ablations.","marker":"[56]"},{"why":"Defines the TIVA dataset, the four-modality benchmark where the largest gains are reported.","marker":"[57]"},{"why":"Defines the KVC16K dataset with dense audio and video coverage, used for the imbalance experiments.","marker":"[58]"}],"fun_headline_variants":["MoCME weights modalities by non-redundancy for KG completion","Non-redundancy fusion and entropy negatives top KG bench","Complementarity-driven fusion beats five KG benchmarks","MoCME: fuse by mutual information, sample by entropy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes that the mutual-information estimates between expert views and between modalities are stable and meaningful, so the fusion weights reward genuinely non-redundant information rather than noise; if those estimates are biased, the weights are effectively arbitrary.","fun_headline_variants_meta":{"raw":{"variants":["MoCME weights modalities by non-redundancy for KG completion","Non-redundancy fusion and entropy negatives top KG bench","Complementarity-driven fusion beats five KG benchmarks","MoCME: fuse by mutual information, sample by entropy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000142,"raw_usage":{"total_tokens":1141,"prompt_tokens":889,"completion_tokens":252,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":183}},"tokens_in":505,"tokens_out":252,"duration_ms":4114,"temperature":1.0,"reasoning_tokens":183,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:26:16.421067+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a fixed dataset such as DB15K, rerun MoCME with the complementarity weights replaced by uniform random weights while keeping everything else identical; if MRR does not drop materially below 39.62, the complementarity signal is not what carries the reported result.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MKG-Y and MKG-W benchmark datasets and the relation-enhanced negative-sampling baseline."},{"cited_title":"Zhang, Z","cited_arxiv_id":null,"evidence_quote":"Sets the hyperparameter and baseline configuration the paper follows, and is the strongest comparison result it must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited as the source of the mutual information neural estimator and of the principle that complementarity aids multimodal robustness."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the DB15K multimodal knowledge graph used for the three-modality experiments and ablations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the TIVA dataset, the four-modality benchmark where the largest gains are reported."}],"review_version":1}