{"id":"f1b6040f-29d6-4d53-92ec-f1e349fe5d0b","arxiv_id":"2607.25324","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A soft mixture-of-experts model with a load-balancing loss improves multimodal glaucoma detection AUC by 0.5–1.5 points over strong baselines on three datasets.","lead":"A deep-learning model using three 'expert' networks—two for single eye-imaging types and one for both—plus a balancing loss achieves slightly higher glaucoma-detection AUC than prior methods on three public datasets. The gains are small and not shown to be statistically significant, but the approach is a plausible incremental improvement.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-balancing loss is never ablated, and its weight λ_dyn=0.001 may make the term numerically negligible; the central 'balanced' contribution is therefore unverified.","rationale":"The reader's weakest-assumption analysis is correct and identifies the same load-bearing concern: the paper's central causal claim—that the CV-based load-balancing loss improves balanced multi-modal learning—is never tested by ablation. I read the paper in good faith: the raw AUC numbers are plausible, and the proposed architecture may well work. The issue is attribution. The abstract explicitly frames the method as 'balanced soft mixture-experts ... with load balancing loss' and the discussion attributes the consistent gains to the balancing mechanism. But without an ablation, the observed improvements could be caused entirely by the three-expert ensemble or the learned gating network. I strengthen the reader's concern by noting the hyperparameter scale: λ_dyn=0.001 is likely too small to influence training. The load-balancing loss in Eq. 1 is a squared coefficient of variation; even if the CV is large, the weighted term is orders of magnitude smaller than the cross-entropy loss. Therefore the 'balanced' component may be inert. This concern is internal to the paper's argument, not a matter of consensus. It is also more load-bearing than the secondary issues (missing significance tests, underspecified SOTA comparison), because those would weaken the strength of the performance claim, whereas the missing ablation undermines the claimed mechanism itself. A single controlled comparison of λ_dyn=0 versus λ_dyn=0.001 would settle the matter. The reader's conditional verdict remains appropriate: the paper should be accepted only after this ablation is provided.","tokens_in":12257,"tokens_out":5612,"duration_ms":59165,"concrete_test":"On the same data splits and seeds, train SMoE with λ_dyn=0 (Eq. 1 removed) and λ_dyn=0.001, and optionally λ_dyn=1, and record (i) test AUC on all three datasets and (ii) the batch-averaged gating weight per expert and its CV during training. If the λ=0 run matches the λ=0.001 run in AUC and gating distribution, the load-balancing loss is not load-bearing; if λ=0 collapses to one expert or changes AUC, the balancing mechanism matters. Also report the relative magnitudes of classification loss and L_balance to show whether 0.001 is in an effective regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central contribution is the 'balanced' SMoE: a three-expert soft mixture with a coefficient-of-variation load-balancing loss (Eq. 1). The abstract claims this balanced design surpasses uni-modal and balanced multi-modal baselines, but no experiment isolates Eq. 1. There is no ablation with λ_dyn=0, no analysis of expert gating loads, and no loss-scale analysis. Section 4.3 sets λ_dyn=0.001. For typical cross-entropy losses around 0.5–1, a squared CV (usually ≤1) weighted by 0.001 contributes at most ~0.001 to the total loss, several orders of magnitude below the classification objective. Unless the gating network is trained with a separate optimizer or loss scaling (not described), the balancing term is likely too weak to change expert utilization. Thus the observed AUC gains could be entirely from the three-expert architecture and the learned gating network, with the 'balanced' mechanism being a no-op. The Limitations section discusses model complexity and missing modalities but never flags this missing ablation. To support the abstract's claim that the load-balancing loss is what makes the method work, the authors must show that disabling Eq. 1 changes the outcome.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a balanced soft mixture-of-experts (SMoE) model for glaucoma detection. The architecture uses three EfficientNet-B1 experts: one for SLO fundus images, one for OCT images, and one that jointly processes both modalities. A lightweight CNN gating network assigns per-sample soft weights to the three experts, and a coefficient-of-variation load-balancing loss (Eq. 1) is added to encourage uniform expert utilization. The model is evaluated on three public datasets (FairVision, FairDomain, HarvardGF) and compared against uni-modal CNN/ViT baselines and balanced multimodal baselines (OPM, OGM, CGGM, OPM+OGM). The authors report the highest overall AUC for SMoE on all three datasets (83.29%, 84.31%, 86.33%, respectively) and claim that the proposed method surpasses all uni-modal, conventional multimodal, and state-of-the-art balanced multimodal baselines.","tokens_in":12621,"tokens_out":4021,"duration_ms":42038,"significance":"If the reported results are statistically robust and the balancing mechanism is actually responsible for the gains, the paper would provide a modest but potentially useful architecture for balanced multimodal medical-image classification, evaluated on three public benchmarks with standard deviations reported. The comparison against OPM/OGM/CGGM is a strength, as is the use of externally published datasets. However, the central contribution—the load-balancing loss—is never isolated experimentally, and the claimed advantage over the strongest baseline is numerically tiny on FairDomain (0.01 AUC) without significance testing. The significance of the paper therefore depends on additional ablation and statistical validation that the current manuscript does not provide.","major_comments":[{"comment":"The load-balancing loss is the paper's central contribution, but it is never ablated. There is no experiment with λ_dyn=0, no varying λ_dyn, and no analysis of the actual gating-load distribution. Given λ_dyn=0.001 (Section 4.3) and typical cross-entropy loss magnitudes of order 0.5–1, the squared-CV term weighted by 0.001 is likely several orders of magnitude smaller than the classification objective; unless the gating network is trained with separate scaling (which is not described), the balancing term may be numerically negligible. The observed AUC improvements could therefore come entirely from the three-expert architecture and the learned gating network. The Limitations section discusses complexity, missing modalities, and fairness but does not flag this missing ablation. To support the abstract's claim that the balanced design is what makes the method work, the authors must show th","section":null},{"comment":"No statistical significance tests are reported. The FairDomain result is particularly problematic: SMoE achieves 84.31±0.0029 AUC while OPM+OGM achieves 84.30±0.0048; the difference is 0.01 percentage points, well within one standard deviation of both runs. Without paired significance tests (e.g., DeLong test or bootstrapped confidence intervals) or multiple-seed comparisons, the claim that SMoE 'surpasses' all balanced multimodal baselines is unsupported on this dataset. The abstract's unconditional superiority claim requires statistical support, especially where the margin is so small.","section":null},{"comment":"The architecture is under-specified. The 'Normal' multimodal baseline in Table 2 is not described at all: what fusion does it use, what is the exact architecture, and is it a simple concatenation plus classification head? Similarly, the multi-modal expert is said to 'jointly process both modalities,' but the fusion mechanism inside that expert is not defined (e.g., early concatenation, attention, cross-modal interactions). The gating network is described as processing 'fused multi-modal features,' but the fusion method for the gating input is also omitted. These details are necessary for reproducibility and for interpreting the comparison against conventional multimodal baselines.","section":null},{"comment":"The claim that SMoE surpasses 'current state-of-the-art balanced multi-modal models' is broader than the evidence. Table 2 compares against only four balanced strategies (OPM, OGM, CGGM, OPM+OGM). The paper does not compare with other recent balanced multimodal methods or with DynMM, which is cited as inspiration. Unless a more comprehensive comparison is provided, the 'state-of-the-art' claim should be qualified to 'compared methods.'","section":null}],"minor_comments":[{"comment":"The HarvardGF percentages in Figure 1 sum to 101% (53.4% glaucoma + 47.6% non-glaucoma). Also, the FairVision description says the dataset has 10,000 subjects each for AMD, DR, and glaucoma, but then states 48.7% are diagnosed with glaucoma; these numbers are inconsistent and need clarification.","section":null},{"comment":"Many typographical and formatting issues: 'T able 1' and 'T able 2' in table captions, 'F airvision'/'F airdomain' in table rows, 'Comparitive methods' in §4.1, 'In additon' in §3, 'figure [5]' instead of 'Fig. 5', and missing spaces in figure captions.","section":null},{"comment":"The sentence 'CGGM performs less poorly than the normal model' should be 'CGGM performs worse than the normal model' (or similar).","section":null},{"comment":"Some references are incompletely formatted (e.g., [8] 'Hwang, D.K., al.', [25] 'Cai, Y., et al.'); full author lists or consistent et al. formatting should be used.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper applies a soft mixture-of-experts design with three experts and a gating network to multimodal glaucoma detection, reporting AUC gains on three public benchmarks. The gains over existing balanced multimodal baselines are real-looking and consistent in direction, though small. That part mostly holds up. But the central claim of the paper—that the coefficient-of-variation load-balancing loss makes the model 'balanced'—is never tested. There is no ablation with the loss disabled, no expert utilization plots, no loss-scale analysis. With λ_dyn set to 0.001 and a typical cross-entropy loss around 0.5–1, the balancing term is several orders of magnitude smaller than the classification objective. Unless the gating network is trained with a separate optimizer or loss scaling, the mechanism is likely a no-op. The authors do flag limitations about complexity and missing modalities, but they do not flag this missing ablation.\n\nWhat is new: soft MoE with three modality-specific experts (two unimodal, one multimodal) plus a CV-based balancing penalty, applied to SLO/OCT glaucoma detection. That is a legitimate extension of DynMM-style ideas to a clinically relevant task, and the study is run on three standard datasets with a reasonable baseline set (OPM, OGM, CGGM, OPM+OGM). The writing is clear and the limitations section is honest as far as it goes.\n\nSoft spots beyond the missing ablation: no statistical significance tests; the FairDomain margin over OPM+OGM is 0.01 AUC, which is noise without error bars on the difference. The 'Normal' multimodal baseline and the fusion mechanism inside the multimodal expert are underspecified—we never learn how the modalities are combined. The abstract claims to surpass 'state-of-the-art balanced multi-modal models,' but the comparison is limited to four modulation-based methods with the same backbone. The author overlap with the datasets is worth a note but not disqualifying, since the benchmarks are public and labels are independent.\n\nWho this is for: researchers working on multimodal medical imaging, especially those interested in MoE-style architectures for glaucoma screening. It is a workmanlike empirical paper, not a methodological breakthrough.\n\nMy recommendation: send it to peer review, but require the load-balancing ablation and basic significance testing before acceptance. Without those, the 'balanced' contribution is unverified, and the paper reads as a straightforward MoE extension with an unsubstantiated novelty claim.","headline":"Solid but incremental SMoE application to multimodal glaucoma detection; the main claimed contribution—the CV-based load-balancing loss—is never ablated and its weight is set so low that the term may be a no-op.","tokens_in":13019,"tokens_out":2081,"would_cite":false,"duration_ms":23925,"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":"A balanced soft mixture-of-experts model surpasses all baselines for glaucoma detection across three datasets.","keywords":["glaucoma detection","multi-modal learning","mixture of experts","load balancing loss","coefficient of variation","OCT","fundus photography","AUC"],"falsifier":"Train the same SMoE architecture with the load-balancing loss removed (set the regularization coefficient to 0) and compare AUC on the three datasets. If removing the loss leaves AUC essentially unchanged in all three, the paper's stated contribution is not supported. Alternatively, if the loss is necessary, the AUC should drop measurably when it is disabled.","tokens_in":12211,"feed_emoji":"👁️","tokens_out":3546,"duration_ms":31103,"temperature":0.7,"pith_summary":"Glaucoma detection from retinal images usually relies on either fundus or OCT scans alone, or naively fuses them. The paper argues that when two modalities are combined, one modality tends to dominate training, leaving the other under-optimized. It proposes a soft mixture-of-experts model with three branches—one for each modality and one for both—whose predictions are blended by a learned gating network. A coefficient-of-variation load-balancing loss keeps the three experts roughly equally used. On three public benchmarks the model reports higher AUC than all uni-modal and multi-modal baselines compared.","feed_headline":"Three-expert model tops glaucoma detection on all three datasets","feed_subtitle":"Soft routing plus a balance loss outdoes uni-modal and balanced multi-modal baselines, hitting 86.33% AUC.","key_machinery":"The load-bearing mechanism is the soft mixture-of-experts (SMoE) module: a lightweight CNN gating network computes a weight per expert (two uni-modal, one multi-modal) that sums to one, and the final prediction is the weighted average of expert outputs. The balance regularizer is the coefficient of variation (CV) of per-expert gating load across a mini-batch, squared and scaled by a hyperparameter; minimizing CV pushes the model toward uniform expert usage, preventing gating collapse.","core_discovery":"The central claim is that a soft mixture-of-experts architecture, with separate uni-modal experts and a joint multi-modal expert plus a gating network trained with a coefficient-of-variation load-balancing loss, achieves the highest glaucoma-detection AUC on FairVision (83.29%), FairDomain (84.31%), and HarvardGF (86.33%), outperforming uni-modal EfficientNet, DenseNet, ResNet, ViT-B, and balanced multi-modal methods OPM, OGM, CGGM, and OPM+OGM. The authors interpret this as evidence that adaptive per-sample expert weighting, guided by a balance penalty, mitigates modality imbalance more effectively than forward-pass or gradient-based modulation.","pith_inferences":["The paper does not ablate the CV loss; the reported gains might come entirely from the three-expert soft routing architecture, with the balancing loss making little difference. A comparison of SMoE with and without the CV term would isolate the contribution.","The gating network itself may be learning to ignore noisy samples from one modality, which would make SMoE's robustness a property of weighting rather than balance; this could be tested by checking per-sample gate distributions on misclassified cases.","If the CV loss matters, it implies that equal expert usage is a proxy for balanced learning; an alternative test would be to replace CV with entropy of the gate distribution to see if any 'balance' penalty suffices.","Because all experts share the same backbone and are trained end-to-end, the multi-modal expert may simply be a fine-tuned copy of the uni-modal experts; inspecting the expert weights on held-out data would reveal whether it actually learns a distinct fusion representation."],"forward_implications":["If the reported AUCs hold, SMoE provides a drop-in fusion head for two-modality retinal imaging that beats existing balancing tricks without changing the backbone.","The method's gain over OPM/OGM/CGGM suggests that per-sample soft weighting, rather than gradient or prediction modulation, is a more effective way to counteract modality dominance in this domain.","The approach is modality-agnostic and can be applied to other paired-image diagnostics, such as diabetic retinopathy as the authors note.","Stable standard deviations across runs indicate the results are reproducible under the authors' training protocol."],"fun_headline_variants":["Balanced soft MoE tops glaucoma detection on 3 datasets","Three-expert model with load balancing wins glaucoma AUC","Soft gating plus balance loss lifts glaucoma detection","Load-balanced mixture of experts improves glaucoma screening","Mixture-of-experts with balance loss beats uni-modal baselines"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's central contribution—the coefficient-of-variation load-balancing loss—is never ablated, so the reported AUC gains could be caused by the gating network and three-expert architecture alone, with the balancing loss contributing little or nothing.","fun_headline_variants_meta":{"raw":{"variants":["Balanced soft MoE tops glaucoma detection on 3 datasets","Three-expert model with load balancing wins glaucoma AUC","Soft gating plus balance loss lifts glaucoma detection","Load-balanced mixture of experts improves glaucoma screening","Mixture-of-experts with balance loss beats uni-modal baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1025,"prompt_tokens":737,"completion_tokens":288,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":210}},"tokens_in":481,"tokens_out":288,"duration_ms":3316,"temperature":1.0,"reasoning_tokens":210,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:44:51.404457+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same SMoE architecture with the load-balancing loss removed (set the regularization coefficient to 0) and compare AUC on the three datasets. If removing the loss leaves AUC essentially unchanged in all three, the paper's stated contribution is not supported. Alternatively, if the loss is necessary, the AUC should drop measurably when it is disabled.","supporting_citations":[],"review_version":1}