{"id":"97a76a00-38e5-4f54-8eaf-223a4b99f05f","arxiv_id":"2504.19458","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CDMEA improves multi-modal entity alignment by subtracting a tuned multiple of the visual-only alignment score from the fused alignment score, with reported Hits@1 gains over 14 baselines on 9 datasets.","lead":"CDMEA is a counterfactual-debiasing method for multi-modal entity alignment that ranks candidates with the fused graph-plus-image score minus a tuned fraction of the image-only score. On 9 benchmark datasets it reports gains over 14 baselines, largest when training data is scarce, images are noisy, or corresponding images look dissimilar.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The counterfactual subtraction in Eq. 10/23 is algebraically just a linear reweighting of the three modality scores, so the causal-NDE interpretation is unsupported and the reported gains do not demonstrate causal debiasing.","rationale":"The reader's verdict is CONDITIONAL, and the reader's weakest assumption already points to the zeroing operation in Eq. 18. My analysis sharpens that concern: even granting the causal graph in Fig. 4(a), the implemented TIE is exactly a linear reweighting of the three unimodal scores, because the counterfactual branch reduces to alpha_v Y_v. Consequently, the causal terminology carries no identifying power in the reported numbers; the empirical superiority may be genuine, but it demonstrates a tuned reweighting heuristic, not estimation of a Natural Direct Effect or Total Indirect Effect. The proposed check is inexpensive given the public code and directly settles whether the causal apparatus is load-bearing. I therefore leave the reader's CONDITIONAL verdict unchanged: the method is plausible as a heuristic and the experimental comparison is broad, but the central causal interpretation should be withdrawn or re-supported by an actual intervention that changes the visual input rather than zeroing scores.","tokens_in":20540,"tokens_out":5660,"duration_ms":59473,"concrete_test":"Train CDMEA on FB-DB15K (20%) with the released code, then compare two inference procedures: (A) the published TIE of Eq. 23 with the zeroed counterfactual branch; (B) a plain weighted sum with fixed weights w_v = (1-beta) alpha_v, w_g = alpha_g, w_m = alpha_m over the same trained encoders. If the H@1, H@10, MRR, and rankings agree up to floating point, the causal subtraction is redundant and the NDE interpretation is refuted. As a stronger check, replace the visual branch inputs with shuffled images at inference and compare H@1, which tests whether zeroing similarity scores matches a true intervention on the visual modality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"With the attentive fusion of Eq. 20, Y_{v,g,m} = alpha_v Y_v + alpha_g Y_g + alpha_m Y_m, where alpha_k = exp(phi_k) / sum exp(phi_k') are fixed learnable weights. Eq. 18 sets Y_{g*} = Y_{m*} = 0, so the 'counterfactual' branch collapses to Y_{v,g*,m*} = alpha_v Y_v; it is not M(V=v*, G=g*) as defined in Eq. 5. Substituting into Eq. 23 gives TIE = (1-beta) alpha_v Y_v + alpha_g Y_g + alpha_m Y_m. This is a weighted linear combination of the same three similarity scores, obtainable by simply changing the fusion weights. The causal graph, the no-confounder assumption, and the TE/NDE/TIE vocabulary impose no constraint beyond one scalar beta. Setting a similarity score to zero is not Pearl's do-operator; G is not a descendant of V yet it is also zeroed; and with beta=0.2 the final score retains 0.8 alpha_v Y_v, so the 'blocked' direct effect is only damped, not blocked. The central causal claim therefore does not follow from the implementation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CDMEA, a framework for multi-modal entity alignment (MMEA) that aims to reduce visual modality bias. The authors set up a causal graph with visual (V), graph (G), fused (M), and prediction (Y) nodes, define TE, NDE, and TIE, and propose a debiasing inference that computes TIE = Y_{v,g,m} - beta * Y_{v,g*,m*}. The implementation consists of three encoders (visual, graph RRGAT, fused RRGAT), a weighted fusion of per-modality similarity scores, InfoNCE training, and a counterfactual inference that zeroes the graph and fused scores. The paper reports state-of-the-art results on nine benchmarks, with the largest gains under low-resource, noisy, and low-similarity-image settings.","tokens_in":20816,"tokens_out":7300,"duration_ms":66925,"significance":"If the causal interpretation held, the paper would contribute a principled and general debiasing principle for MMEA, with a reusable implementation and extensive empirical validation. The strengths are the breadth of the evaluation (9 benchmarks, 14 baselines), the module-ablation and robustness studies, the code release, and the observation that the visual modality can hurt performance. However, the central causal quantity reduces algebraically to a linear reweighting of the same three scores, so the reported gains currently demonstrate a tuned weighted fusion rather than a causal debiasing effect. The empirical contribution may still be useful, but the title, abstract, and framing over-claim what is established.","major_comments":[{"comment":"The counterfactual branch as implemented collapses to alpha_v * Y_v. With the attentive fusion in Eq. (20), Y_{v,g,m} = alpha_v Y_v + alpha_g Y_g + alpha_m Y_m, and Eq. (18) sets Y_{g*}=Y_{m*}=0, so Y_{v,g*,m*} = alpha_v Y_v. Substituting into Eq. (23) gives TIE = (1-beta)*alpha_v*Y_v + alpha_g*Y_g + alpha_m*Y_m. This is a one-parameter reweighting of the existing scores; the NDE/TIE vocabulary and the causal graph impose no additional constraint. The claim that the model predicts based on the Total Indirect Effect is therefore not supported by the implementation. Either implement a genuine counterfactual (for example, recomputing M(V=v*,G=g*) and the fusion weights in the blocked world) or reframe the method as a heuristic reweighted fusion.","section":"3.2 / 3.3.4 / 3.4.2 (Eqs. 20, 18, 23)"},{"comment":"Setting a blocked modality score to zero is not Pearl's do-operator and does not correspond to a no-treatment counterfactual in the stated graph. In the graph of Fig. 4(a), G is not a descendant of V, yet the NDE computation zeroes both G and M; and the 'blocked' direct effect is only damped by beta=0.2, not blocked. The paper should justify the zero-imputation as an intervention or remove the causal-effect terminology.","section":"3.2 (Eqs. 7-10)"},{"comment":"The hyperparameter beta is grid-searched from 0.0 to 0.9 on H@1 (Section 4.1.3, Fig. 9), so the final debiasing strength is selected on the target metric. This makes the TIE prediction a calibrated interpolation rather than a parameter-free causal estimate. At minimum, report a validation-set selection or a sensitivity analysis with beta chosen without test feedback.","section":"4.1.3 / 4.4.5 (beta tuning)"},{"comment":"All reported numbers are single runs without error bars. On the bilingual benchmarks the gains over IBMEA are 0.6-1.3% H@1, which is within the typical run-to-run variance of such models. Without multiple seeds or significance testing, the claim of consistent state-of-the-art performance on all nine benchmarks is not firmly established.","section":"4.2 (Tables 1-2)"}],"minor_comments":[{"comment":"'generalCasual Debiasing framework' should be 'general Causal Debiasing framework', and the Section 2.2.2 heading 'Casual Effects' should be 'Causal Effects'.","section":"Section 1 and Section 2.2.2"},{"comment":"The attention denominator is malformed ('exp(q^T h_rk) ... exp(q^T h_rk'))); please fix the formula and the summation notation.","section":"Eq. (13)"},{"comment":"The orthogonality proof should explicitly use the stated normalization ||h_rk||=1 inside the displayed equation; as written, the second equality hides this assumption.","section":"Eq. (14)"},{"comment":"The Table 4 header 'FB-YG15K ((50%)' has a double parenthesis, and the improvement percentages in Section 4.2 should state whether they are absolute or relative.","section":"Table 4 and Section 4.2"},{"comment":"Entities without images are assigned random visual vectors; given the paper's focus on visual reliability, please ablate this choice or discuss its effect on the low-similarity improvements. Also, 'the results improve as image similarity decreases' in Section 4.4.4 is ambiguous: presumably the relative advantage over baselines grows as similarity decreases, not the absolute H@1.","section":"Section 4.1.3 and Section 4.4.4"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is valid and lands exactly on the paper's central claim: Eq. (23) reduces to a linear reweighting of the three similarity scores. I would not reject, because the empirical study is solid enough that a reframed 'weighted fusion with tuned visual suppression' could be a publishable contribution; however, the causal framing cannot stand as written. The authors should also address the single-run results before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper is worth reading for the empirical result—a simple inference-time correction that consistently helps in low-resource, noisy, and low-similarity settings—but the causal framing is largely decorative. The implementation reduces to a weighted linear combination of the same three modality scores with one tuned scalar.\n\nWhat's genuinely new: this is the first MMEA paper to apply TE/NDE/TIE-style counterfactual subtraction, and the causal graph with V, G, M, and Y is specific to the task. The experiments are broad: nine benchmarks, fourteen baselines, ablations, low-resource curves, noise robustness, and they show that applying the same correction to existing baselines (MCLEA, DESAlign, MEAformer) gives consistent gains. That is a real, useful finding. The low-resource and low-similarity improvements are the most credible; the effect sizes are modest but not negligible (e.g., +4.3% H@1 at 20% seeds on FB-DB15K).\n\nThe load-bearing problem is that the counterfactual branch doesn't match the causal graph. Eq. 18 sets Y_g and Y_m to zero while keeping Y_v, and because fusion is a softmax-weighted sum, Y_{v,g*,m*} collapses to alpha_v Y_v. So TIE becomes (1-beta)*alpha_v Y_v + alpha_g Y_g + alpha_m Y_m. That is just a reweighting of the same three scores—you could get it by changing the fusion weights. Setting scores to zero is not Pearl's do-operator; G is zeroed even though it is not a descendant of V; and with beta=0.2 the 'blocked' direct effect is only damped, not blocked. The TE/NDE/TIE vocabulary is not doing any work here. The causal claim should be dropped or the implementation should actually estimate the counterfactual. Also: beta is grid-searched on the test H@1, there are no error bars anywhere, and missing images are replaced with random vectors, which the paper reports but does not interrogate. These are real issues, but they are fixable and do not change the empirical pattern.\n\nWho this is for: anyone working on MMEA or modality fusion who wants a cheap robustness trick; also a good teaching example of how causal language can outrun implementation. I would send it to a serious referee, but the referee should insist on either a reframing as a heuristic weighting or a genuinely causal estimator. The reader's conditional verdict is about right; my own take is slightly more skeptical of the causal claims but still positive on the empirical contribution.","headline":"A useful inference-time reweighting trick for MMEA, wrapped in a causal story the implementation doesn't actually support; engage with it for the empirical results, not for the causal claims.","tokens_in":21362,"tokens_out":2931,"would_cite":true,"duration_ms":29400,"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":"The paper argues that a counterfactual score—total effect minus the visual modality's direct effect—removes image-matching shortcut bias in multi-modal entity alignment and beats 14 prior methods on nine benchmarks.","keywords":["multi-modal entity alignment","multi-modal knowledge graph","counterfactual debiasing","modality bias","total indirect effect","causal inference","shortcut learning"],"falsifier":"Compare CDMEA's debiased score with the counterfactual graph and fused scores set to zero against a version that sets them to random baseline values. If the ranking gains depend on which baseline is used, the improvement is tied to the zero convention rather than to a causal estimate. In the opposite direction, on a test set where equivalent entities already share highly similar images, the causal account predicts little or no gain from debiasing; a large gain there would indicate the method is adjusting score weights rather than removing a shortcut.","tokens_in":20344,"feed_emoji":"🔗","tokens_out":13102,"duration_ms":119436,"temperature":0.7,"pith_summary":"The paper targets multi-modal entity alignment—matching the same real-world entity across two knowledge graphs that carry both relational structure and images. It documents that most equivalent entities have low-similarity images (86.3% below cosine 0.5 in FB-DB15K) and that existing models perform worst on exactly those pairs, arguing that they have learned a shortcut image-matching bias instead of using graph evidence. To remove that bias, the paper models the alignment prediction as a causal graph with visual, graph, and fused-modality inputs, then scores candidates by the Total Indirect Effect: the factual prediction minus a scaled counterfactual prediction in which the graph and fused branches are blocked. Experiments across nine benchmarks and fourteen prior methods report consistent gains, concentrated in the settings where visual cues are least trustworthy: low image similarity, high image noise, and scarce training alignments.","feed_headline":"Counterfactual subtraction lifts entity alignment past 14 rivals","feed_subtitle":"Removing visual modality's direct effect pays off most when images are noisy, dissimilar, or training data is scarce.","key_machinery":"The load-bearing object is the causal graph over $V$ (visual input), $G$ (graph input), $M$ (fused modality, a mediator), and $Y$ (prediction), together with the counterfactual subtraction $TIE = Y_{v,g,m} - \\beta Y_{v,g^*,m^*}$. Here $Y_{v,g^*,m^*}$ is obtained by setting the graph and fused similarity scores to zero while keeping the visual branch's learned weights, implementing the counterfactual world in which the visual direct path is the only contributor; subtracting a fraction $\\beta=0.2$ of it from the factual prediction is meant to cancel the visual shortcut while preserving the indirect path $V \\to M \\to Y$. The framework is implemented with per-modality encoders (VGG/ResNet for images, a relational reflection graph attention network for graph structure and for fusion) and an attentive weighted sum over the three score branches.","core_discovery":"On the paper's own terms, the discovery is that visual-modality bias in multi-modal entity alignment can be diagnosed and removed causally. The authors define the factual prediction $Y_{v,g,m}$ from visual, graph, and fused scores, and a counterfactual prediction $Y_{v,g^*,m^*}$ in which the graph and fused similarity scores are set to zero while the visual branch stays active. The debiased score is $TIE = Y_{v,g,m} - \\beta Y_{v,g^*,m^*}$, which they interpret as the Total Effect minus the Natural Direct Effect of the visual modality; the optimal $\\beta$ is $0.2$. With this score, CDMEA reports the best Hits@1, Hits@10, and MRR on all nine benchmark settings, with the largest margins at low training ratios (e.g., FB-DB15K 20% H@1 $0.674$ vs. runner-up $0.631$).","pith_inferences":["Beyond the paper, the same total-effect-minus-direct-effect recipe could be ported to any multimodal retrieval task with a suspected shortcut modality, with a per-modality $\\beta$ tuned on validation data.","The reported $\\beta$ sensitivity suggests $\\beta$ could be made adaptive per entity or per image-similarity bin, subtracting more visual effect precisely where image similarity is low.","The framework could be extended to more than two modalities by building a multi-level causal graph and subtracting each suspect modality's direct effect in turn."],"forward_implications":["Existing MMEA models can be debiased at inference time simply by subtracting a scaled counterfactual visual score; the paper shows this improves MCLEA, DESAlign, and MEAformer without retraining them.","The benefit should be largest when image similarity is low, image noise is high, or alignment seeds are scarce; the paper's experiments report gains concentrated in those regimes.","Partial subtraction is necessary: $\\beta=0.2$ beats both no subtraction ($\\beta=0$) and full subtraction ($\\beta=1$), so the visual modality still contributes useful signal through the fused path.","The debiasing also speeds up training: CDMEA converges faster and reports lower training time than four strong baselines on the same hardware."],"supporting_citations":[{"why":"Supplies the causal definitions (Total Effect, Natural Direct Effect, Total Indirect Effect) and the counterfactual-world formalism the framework builds on.","marker":"[18]"},{"why":"Provides the counterfactual debiasing inference recipe, including the $\\beta$ control factor used to scale the subtracted direct effect.","marker":"[39]"},{"why":"Introduces the MMKG datasets (FB-DB15K, FB-YG15K) and the task setting that all benchmark comparisons use.","marker":"[26]"},{"why":"Supplies the cross-modal graph attention idea that the fused-modality encoder is built on.","marker":"[56]"},{"why":"IBMEA; the strongest prior baseline and the information-bottleneck approach that motivates handling misleading visual cues.","marker":"[37]"},{"why":"MEAformer; a strong baseline and the model that gains most when the paper's counterfactual debiasing inference is grafted onto it.","marker":"[8]"},{"why":"DESAlign; a strong baseline used in the generality experiments showing the debiasing inference transfers across models.","marker":"[51]"}],"fun_headline_variants":["Causal debiasing of visual cues improves entity alignment","Neutralizing visual shortcut lifts MMEA past 14 methods","Counterfactual visual debiasing beats 14 rivals in alignment","Removing visual direct effect boosts entity matching","Causal visual debiasing wins entity alignment benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that making the graph and fused-modality similarity scores zero in the imagined alternative world is a faithful way to switch those pathways off, and that nothing unobserved drives both the images and the predictions; if either assumption fails, the subtracted quantity is not really the visual modality's direct effect.","fun_headline_variants_meta":{"raw":{"variants":["Causal debiasing of visual cues improves entity alignment","Neutralizing visual shortcut lifts MMEA past 14 methods","Counterfactual visual debiasing beats 14 rivals in alignment","Removing visual direct effect boosts entity matching","Causal visual debiasing wins entity alignment benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1452,"prompt_tokens":978,"completion_tokens":474,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":394}},"tokens_in":594,"tokens_out":474,"duration_ms":4527,"temperature":1.0,"reasoning_tokens":394,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:51:57.624169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare CDMEA's debiased score with the counterfactual graph and fused scores set to zero against a version that sets them to random baseline values. If the ranking gains depend on which baseline is used, the improvement is tied to the zero convention rather than to a causal estimate. In the opposite direction, on a test set where equivalent entities already share highly similar images, the causal account predicts little or no gain from debiasing; a large gain there would indicate the method is adjusting score weights rather than removing a shortcut.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the counterfactual debiasing inference recipe, including the $\\beta$ control factor used to scale the subtracted direct effect."},{"cited_title":"Rosenblum","cited_arxiv_id":null,"evidence_quote":"Introduces the MMKG datasets (FB-DB15K, FB-YG15K) and the task setting that all benchmark comparisons use."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the cross-modal graph attention idea that the fused-modality encoder is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"IBMEA; the strongest prior baseline and the information-bottleneck approach that motivates handling misleading visual cues."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MEAformer; a strong baseline and the model that gains most when the paper's counterfactual debiasing inference is grafted onto it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DESAlign; a strong baseline used in the generality experiments showing the debiasing inference transfers across models."}],"review_version":1}