{"id":"7548fefa-8fbe-4d96-8473-2a697c716d9a","arxiv_id":"2502.18481","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"MDE claims to enhance multi-modal recommendation by amplifying modality differences, but the published difference loss minimizes the difference it is supposed to maximize.","lead":"This paper proposes MDE, a graph-based method for multi-modal recommendation that tries to amplify modality-specific signals while aligning shared ones. It reports modest top-5 gains on two of three Amazon datasets, but its central equation appears to do the opposite of what the text claims.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (6) defines a loss whose minimization shrinks the cross-modal distance, directly contradicting the stated MDA goal of amplifying modality differences; the central mechanism is inverted as written.","rationale":"Reading the paper in good faith, the central claim is that MDE extracts modality-specific information by amplifying cross-modal differences while preserving shared features through alignment, and that this yields superior recommendation accuracy. The load-bearing condition is that the MDA loss actually amplifies differences. Equation (6) fails this condition: minimizing || -|x-y| || is equivalent to minimizing ||x-y||, so the loss shrinks differences. This is an internal inconsistency in the objective, not a disagreement with external consensus. The reader's stated weakest assumption was the node-level trade-off monotonicity, which I also find under-motivated, but the Eq. (6) problem is more fundamental: even if the trade-off hypothesis were true, the MDA component as written does not do what it claims. The ablation study in Table II attributes gains to MDA, but if the implemented loss is the one in Eq. (6), the improvement cannot be attributed to difference amplification; if the implementation differs from the text, the paper is misleading about its own method. I credit the paper for using public datasets, standard evaluation protocols, and for attempting to address a real limitation of alignment-only multimodal recommenders. However, the central mechanism must be internally consistent for the empirical results to support the stated cause. The NLT assumption, lack of error bars, unreported hyperparameters, and missing code are additional concerns, but the Eq. (6) inversion is the single decisive issue. A corrected loss, code release, and uncertainty reporting could make a revision credible; as written, rejection is appropriate.","tokens_in":7193,"tokens_out":3068,"duration_ms":32542,"concrete_test":"Analytically or numerically verify the gradient of Eq. (6). Let x = H*_{i,v}, y = H*_{i,t}, and w = W_diff ≥ 0. Then L_diff = σ_diff · || w ⊙ |x-y| ||_2 = σ_diff · || w ⊙ (x-y) ||_2 when x-y is elementwise nonnegative, and in general the norm of the absolute difference equals the norm of the difference. The gradient ∂L_diff/∂x = σ_diff · (w⊙(x-y)) / ||w⊙(x-y)||_2 points from y to x, and a gradient-descent update with the given positive coefficient in Eq. (11) decreases ||x-y||. Run one optimization step on two randomly initialized vectors and record the Euclidean distance before and after; it will shrink, proving the loss aligns modalities. If the authors instead intend to amplify differences, the correct loss must be -σ_diff·||x-y||_2 or an equivalent sign flip in Eq. (11); check the code or the updated manuscript for such a correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core of the proposed Modality Difference Amplification (MDA) is Equation (6). It defines H_diff = -|H*_{i,v} - H*_{i,t}| ⊙ W_diff and L_diff = σ_diff · ||H_diff||_2. Since the L2 norm is invariant to the elementwise absolute value, L_diff = σ_diff · || (H*_{i,v} - H*_{i,t}) ⊙ W_diff ||_2 (for nonnegative W_diff, which it is because W_diff derives from a softmax-based preference difference). Minimizing L_diff therefore drives the visual and textual embeddings closer together, i.e., it performs alignment, not difference amplification. The text explicitly states 'By minimizing L_diff, we can maximize the differences between modality features,' which is mathematically false. This is not a matter of tuning or evaluation noise: the loss as written removes the very signal the method claims to enhance. The entire rationale for MDA, and the claimed gains on Baby and Sports, rest on a mechanism that the mathematical formulation contradicts. Although the authors may have intended a sign flip or a different objective (e.g., maximizing a distance term or applying a negative coefficient), no such correction appears in Equations (6) or (11). In the final objective (11), L_diff is added with a positive coefficient, so optimizing L cannot amplify differences if Eq. (6) is the operative definition. This is a load-bearing internal inconsistency that invalidates the central contribution as described.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MDE, a multimodal recommendation framework that learns user and item representations from heterogeneous and homogeneous graphs, fuses visual and textual features using learnable node-level modality preferences, and introduces two auxiliary losses: Modality Difference Amplification (MDA) and Modality Similarity Alignment (MSA), balanced by a Node-Level Trade-off (NLT). The claimed contribution is that amplifying cross-modal differences while preserving shared semantics improves Top-5 recommendation accuracy. Experiments on Baby, Sports, and Clothing report gains over several baselines, with ablations for each module. The central technical problem is that the MDA loss in Eq. (6) is written so that its minimization reduces cross-modal differences, contradicting the paper's central claim.","tokens_in":7464,"tokens_out":6859,"duration_ms":67637,"significance":"If the mechanism worked as described, the node-level preference-based trade-off would be a simple and potentially useful addition to graph-based multimodal recommendation, and the paper's use of three Amazon benchmarks and MMRec baselines is appropriate. The paper also includes ablation and trade-off analyses, which are strengths. However, no code or reproducibility artifact is provided, and the mathematical formulation of the core MDA term is internally inconsistent with the stated purpose. Because the main empirical gains are attributed to that term, the results cannot be interpreted as evidence for the proposed mechanism in its current written form. The Clothing results also weaken the broad 'significantly outperforms' claim.","major_comments":[{"comment":"The loss L_diff is defined as σ_diff·||−|H*_{i,v}−H*_{i,t}|⊙W_diff||_2. Because the L2 norm of a vector equals the norm of its negative, minimizing L_diff minimizes the weighted absolute difference between the visual and textual embeddings. The sentence 'By minimizing L_diff, we can maximize the differences between modality features' is therefore mathematically false. Since Eq. (11) adds L_diff with a positive coefficient, the final objective as written can only shrink modality differences, not amplify them. This invalidates the central MDA mechanism and makes the 'w/o MDA' ablation in Table II uninterpretable as evidence for difference amplification.","section":"II-C, Eq. (6)"},{"comment":"The final alignment loss is written as L_cl = σ_cl·(L_t→v_cl + L_t→v_cl), so the previously defined L_v→t_cl term never appears in the objective. This is an internal inconsistency: if the implementation uses both directional contrastive terms, the text should be corrected; if it does not, the MSA loss is not the one described. In either case, the optimization objective in the paper is ambiguous.","section":"II-C, Eq. (7)"},{"comment":"The abstract and conclusion state that MDE 'significantly outperforms' state-of-the-art methods, but on Clothing MDE's Recall@5 is 0.0402 versus 0.0401 for DRAGON, while its MAP (0.0219 vs 0.0225) and NDCG (0.0267 vs 0.0270) are lower. The paper's explanation of 'instability of user modality preferences' is post hoc and not tested. The superiority claim should be restricted to the datasets and metrics where it holds, with significance tests or error bars.","section":"Table I"},{"comment":"The monotone relationship in Eq. (8), namely that larger |P_i,v − P_i,t| implies more difference amplification and smaller difference implies more alignment, is a stated hypothesis with no independent support. Because NLT is a central novel component, it needs a targeted diagnostic (e.g., comparing preference-difference magnitudes with per-node performance or embedding distances) rather than only the aggregate ablation in Table II.","section":"II-C, Node-Level Trade-off"}],"minor_comments":[{"comment":"The text says 'L_t is omitted due to space constraints'; please provide the complete loss in an appendix for reproducibility.","section":"II-D, Eq. (10)"},{"comment":"The phrase 'recommend system' should read 'recommendation systems'.","section":"II-B"},{"comment":"The broadcast function fb is not defined precisely; it should specify how a per-node scalar is expanded to the feature dimension and whether the same weights are reused for user nodes.","section":"II-C, Eq. (8)"},{"comment":"The text refers to bar charts and a blue dotted line, but the experimental protocol for variants A and B is not fully specified; include dataset, hyperparameters, and how the fixed preference weights were chosen.","section":"III-D, Fig. 2"},{"comment":"No variance or significance measures are reported; given the small differences between top methods, error bars over repeated runs are necessary.","section":"III-B, Table I"}],"recommendation":"reject","confidential_remarks":"The core issue is a sign error in the central loss: Eq. (6) as written implements alignment, not difference amplification, and this contradicts the paper's main claim. Without code, it is impossible to know whether the experiments used a corrected objective. A revised submission would need to fix the objective, re-run the experiments, clarify Eq. (7), and provide significance tests. The paper also builds heavily on the authors' own ECAI 2023 work [16]; the incremental contribution of the trade-off mechanism should be stated more precisely."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core mechanism is mathematically backwards. Equation (6) defines L_diff as a positive norm over the absolute difference between modality embeddings, so minimizing L_diff pushes those embeddings together. The text says the opposite. This is a load-bearing inconsistency, not a tuning issue, and it undercuts the entire MDA contribution.\n\nWhat's worth keeping: the framing is honest and the problem is real — preserving modality-specific information in multimodal recommendation is a legitimate gap. The node-level trade-off (NLT) based on learned preference weights is a thoughtful way to balance alignment and differentiation, and the ablation design is clean. The paper also uses standard benchmarks and MMRec, which makes reproduction easier if code were included. On Clothing, the authors openly acknowledge the weaker result and try to explain it, which is more candid than most.\n\nBut the soft spots are serious beyond the sign error. There are no error bars or significance tests, and the reported gains are modest — around 10% relative on Baby, 7% on Sports, and effectively zero on Clothing, where MDE's MAP and NDCG are lower than DRAGON's. The NLT hypothesis is asserted without independent evidence; it may be a reasonable heuristic, but it's not validated. The paper also doesn't release code or hyperparameter details.\n\nGiven Eq. (6), the ablation results are difficult to interpret. If MDA as written aligns modalities, then the \"w/o MDA\" variant removes an alignment loss, and the comparison doesn't test the claimed difference amplification. This makes the internal evidence for the core idea even weaker.\n\nVerdict: reject. The central claim is unsupported by the paper's own equations. With a corrected objective (e.g., a negative distance term or maximizing a distance), proper uncertainty reporting, and code, the NLT idea might be worth revisiting. As it stands, I wouldn't cite it or send it to review.","headline":"Equation (6) is inverted: minimizing L_diff reduces modality differences instead of amplifying them, so the paper's central claim is unsupported as written.","tokens_in":7990,"tokens_out":2742,"would_cite":false,"duration_ms":27706,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MDE claims that actively widening visual-textual differences, balanced node-by-node against alignment, improves multi-modal recommendation.","keywords":["multi-modal recommendation","modality-specific features","modality-shared features","modality difference amplification","node-level trade-off","contrastive learning","graph neural networks","Bayesian personalized ranking"],"falsifier":"Measure the per-node preference gap $|P_{i,v} - P_{i,t}|$ on a validation set, then compare each node's recommendation-quality change when the difference-amplification loss is removed; if there is no positive correlation between gap size and gain from amplification, the Node-Level Trade-off hypothesis is false. A simpler check is to train MDE on a dataset where modality preferences are identical across users by construction and see whether it still beats a static-weight variant; if it does, the stated trade-off mechanism is not what drives the gains.","tokens_in":6942,"feed_emoji":"📈","tokens_out":6589,"duration_ms":59224,"temperature":0.7,"pith_summary":"Multi-modal recommenders typically align the visual and textual representations of an item, which preserves what the modalities share but washes out what makes each modality informative on its own. This paper argues that recommendation accuracy improves when the model also does the opposite—actively amplifies differences between modality representations—and decides per user and per item how much differentiation versus alignment to apply. The proposed framework, MDE, pairs a modality difference amplification loss with a contrastive similarity alignment loss and couples them through a node-level trade-off driven by learned visual/textual preference weights. On the Baby and Sports benchmarks it reports gains over prior state-of-the-art models across recall, precision, MAP, and NDCG, and near-parity on Clothing. The paper's central claim is that jointly capturing modality-shared and modality-specific features, rather than alignment alone, is what drives the improvement.","feed_headline":"Modality differences, not just alignment, lift recommendation recall","feed_subtitle":"MDE sharpens visual-vs-textual signals per node, beating top baselines on Baby and Sports","key_machinery":"The load-bearing mechanism is the Node-Level Trade-off (NLT), a per-node pair of weight matrices built from learned modality preferences. For each item (and similarly for each user), the fusion module learns a scalar preference $P_{i,v}$ for the visual modality and $P_{i,t}$ for the textual modality; the trade-off then sets $W_{\\mathrm{diff}} = f_b(|P_{i,v} - P_{i,t}|)$ and $W_{\\mathrm{cl}} = f_b(1 - |P_{i,v} - P_{i,t}|)$, where $f_b$ broadcasts the scalar difference to the feature dimension. $W_{\\mathrm{diff}}$ scales a loss that pushes visual and textual representations apart, while $W_{\\mathrm{cl}}$ scales a contrastive loss that aligns them. Because one weight is the complement of the other, the mechanism automatically shifts each node's objective between difference amplification and alignment according to how strongly that node's modalities are predicted to disagree.","core_discovery":"The paper's central discovery is that explicit modality discrimination—encouraging visual and textual embeddings to differ—does not destroy the shared semantics needed for recommendation, provided the conflict is managed node by node. MDE learns a preference weight for each modality on each user and item, fuses features accordingly, and then uses the absolute preference gap to set complementary weights on a difference-amplification loss and an alignment loss. The result, on the paper's evidence, is that the model captures modality-specific cues such as shape and color versus function and price that pure alignment discards, yielding top-5 Recall@5 of 0.0414 on Baby and 0.0479 on Sports, up from 0.0374 and 0.0449 for DRAGON, the previous best in its tables, with matching gains on MAP and NDCG. On Clothing the recall is essentially tied (0.0402 vs 0.0401) while MAP and NDCG are slightly lower, which the paper attributes to unstable user modality preferences in that domain.","pith_inferences":["If the node-level trade-off hypothesis is right, the per-node gain from difference amplification should be concentrated on nodes with larger learned preference gaps; the paper does not report this, but it is a direct testable corollary.","On datasets where one modality dominates almost every user, the learned preference gaps become uniformly small or large, and the node-level trade-off should collapse toward a static weighting; this predicts where MDE's margin will shrink beyond the reported Clothing case.","The same amplify-differences-then-align recipe could transfer to other multi-modal tasks such as retrieval or classification that need modality-specific cues, but that transfer is not part of the paper's claims."],"forward_implications":["Because each user and item carries its own learned visual/textual preference, MDE treats modality importance as heterogeneous across nodes rather than as a single global weight.","Because difference amplification and alignment are combined in one objective, preserving modality-specific cues does not require abandoning shared semantic structure; the two losses coexist under the trade-off.","The ablation results imply that removing any of the three components—difference amplification, similarity alignment, or the node-level trade-off—reduces Recall@5, so each is load-bearing for the reported gain.","The Clothing results imply that MDE's advantage depends on the stability of user modality preferences; when those preferences are noisy, the trade-off helps less."],"supporting_citations":[{"why":"Supplies the LightGCN encoder used for message propagation on the heterogeneous graph (Eq. 2).","marker":"[17]"},{"why":"Provides the three Amazon benchmark datasets (Baby, Sports, Clothing) used for evaluation.","marker":"[20]"},{"why":"Defines the experimental settings and protocol that MDE and all baselines follow.","marker":"[21]"},{"why":"Supplies the Bayesian Personalized Ranking objective used as the main recommendation loss (Eq. 9).","marker":"[19]"},{"why":"Underpins the contrastive modality similarity alignment loss (Eq. 7).","marker":"[9]"},{"why":"The paper cites this as the inspiration for combining heterogeneous and homogeneous graphs in the multi-modal feature learning module.","marker":"[16]"}],"fun_headline_variants":["Modality clash boosts recommendations: MDE shows how","Forget alignment: MDE amplifies modality gaps for better recall","MDE: teach modalities to disagree for sharper recommendations","Node-level modality trade-off lifts recommendation recall","Modality differences, not just alignment, lift recommendation recall"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a larger learned gap between a node's visual and textual preference weights really means differentiation is more useful for that node than alignment; if the true relationship is not monotone, the trade-off weights lose their meaning and the reported gains lose their stated cause.","fun_headline_variants_meta":{"raw":{"variants":["Modality clash boosts recommendations: MDE shows how","Forget alignment: MDE amplifies modality gaps for better recall","MDE: teach modalities to disagree for sharper recommendations","Node-level modality trade-off lifts recommendation recall","Modality differences, not just alignment, lift recommendation recall"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000163,"raw_usage":{"total_tokens":1229,"prompt_tokens":917,"completion_tokens":312,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":234}},"tokens_in":533,"tokens_out":312,"duration_ms":3533,"temperature":1.0,"reasoning_tokens":234,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:21:36.928019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the per-node preference gap $|P_{i,v} - P_{i,t}|$ on a validation set, then compare each node's recommendation-quality change when the difference-amplification loss is removed; if there is no positive correlation between gap size and gain from amplification, the Node-Level Trade-off hypothesis is false. A simpler check is to train MDE on a dataset where modality preferences are identical across users by construction and see whether it still beats a static-weight variant; if it does, the stated trade-off mechanism is not what drives the gains.","supporting_citations":[{"cited_title":"Lightgcn: Simplifying and powering graph convolution network for recommenda- tion,","cited_arxiv_id":null,"evidence_quote":"Supplies the LightGCN encoder used for message propagation on the heterogeneous graph (Eq. 2)."},{"cited_title":"Bootstrap latent representations for multi-modal recommen- dation,","cited_arxiv_id":null,"evidence_quote":"Defines the experimental settings and protocol that MDE and all baselines follow."},{"cited_title":"Enhancing dyadic relations with homogeneous graphs for multimodal recommendation,","cited_arxiv_id":null,"evidence_quote":"The paper cites this as the inspiration for combining heterogeneous and homogeneous graphs in the multi-modal feature learning module."}],"review_version":1}