{"id":"9c00b5c7-8d2f-4a2e-8ffc-0fd3f67307f9","arxiv_id":"2512.02076","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"FDRMFL combines MSE, mutual-information, KL-alignment, and InfoNCE losses to extract task-relevant features in federated multimodal regression, but the claimed ablations and baseline comparisons are not shown.","lead":"FDRMFL is a federated-learning loss that combines prediction error, a mutual-information surrogate, cross-modal alignment, and contrastive anchoring to extract features from multimodal data for regression. It reports large MSE gains over PCA and VAE on synthetic and near-infrared datasets, but the abstract's claims about beating FedAvg-family baselines and an ablation study are not present in the body.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract promises ablations and FedAvg-family baselines that never appear; the central claim is unsupported by the reported tables.","rationale":"The stress-test pass focused on what must be true for the central claim to hold: the four-term objective must be shown to outperform relevant federated baselines and each term must demonstrably contribute. The full text provides neither. Tables 1–4 compare FDRMFL only to PCA, TSVD, RP, and VAE; the five named federated algorithms never appear. The text also promises 'comprehensive ablation and sensitivity analyses' (abstract), but Section 3 contains no ablation of the three regularizers and no sensitivity analysis beyond reporting mean(std) across runs. This is not a matter of convention but of internal consistency: the abstract's two strongest assertions are untestable from the reported results. The reader's identified MI estimator flaw (Eq. 15) is also a real technical weakness—no derivation is given and the bound is not standard for continuous targets—but even a correct MI estimator would not rescue the paper without the missing comparisons/ablations, because the empirical contribution is the specific combination of losses. Thus the most load-bearing concern is the evidentiary gap. We do not doubt that FDRMFL may improve over PCA/VAE on these datasets; the problem is that the paper's central claims go far beyond the evidence. No code, hyperparameter details, or dataset splits are provided, so reproducing the favorable results is not currently possible. The verdict should remain REJECT (unchanged).","tokens_in":19902,"tokens_out":4992,"duration_ms":48559,"concrete_test":"Locate in Section 3 any table/figure containing columns or rows for FedAvg, FedProx, MOON, SCAFFOLD, or FedBN, and any ablation experiment setting λ1=0, λ2=0, or λ3=0. None exist in the provided text, so the abstract's claims are unsupported. To settle, run the identical federated protocol with those five baselines and with each of the three regularization coefficients zeroed on the same five datasets (or at least on Link-1 simulation and one real dataset). If FDRMFL does not achieve the lowest MSE in every scenario, or if removing a term does not increase MSE, the central claim fails. Alternatively, a line-by-line search of the PDF/LaTeX for 'FedAvg', 'FedProx', 'MOON', 'SCAFFOLD', 'FedBN', 'ablation', 'λ1', 'λ2', 'λ3' would confirm the missing support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript's central claim has two testable parts: (i) each of the four objective terms contributes to performance (abstract, §1, §4), and (ii) FDRMFL achieves the lowest MSE among six federated algorithms including FedAvg, FedProx, MOON, SCAFFOLD, FedBN (abstract). Neither is supported in the body. Section 3 reports experiments only against PCA, TSVD, RP, and VAE (Tables 1–4); no FedAvg-family method is ever run, and there is no ablation varying λ1, λ2, λ3 or removing terms. The only sentence resembling a baseline acknowledgement, §2.2 'Building upon FedAvg training…', explicitly concedes FedAvg is prior art but does not test against it. Therefore the reported gains cannot be attributed to the proposed MI/KL/contrastive terms, and the headline superiority over federated algorithms is entirely unverified. This is the load-bearing gap: the paper's contribution is the specific loss combination, and without ablation or federated baselines the empirical evidence does not distinguish FDRMFL from a plain supervised FedAvg with an MSE loss.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FDRMFL, a federated multimodal feature-extraction method for regression. It combines an MSE prediction loss with three regularizers: a mutual-information surrogate (Eq. 15), a symmetric KL divergence between modality feature distributions (Eq. 17), and an InfoNCE-style contrastive loss against historical global representations (Eq. 19). The authors claim each component contributes and that FDRMFL achieves the lowest MSE among six federated algorithms including FedAvg, FedProx, MOON, SCAFFOLD, and FedBN. Experiments are reported on three synthetic link-function settings and two near-infrared spectroscopy datasets, with PCA, TSVD, RP, and VAE as baselines.","tokens_in":20252,"tokens_out":6060,"duration_ms":57818,"significance":"The problem of task-driven multimodal feature extraction under federated non-IID constraints is timely, and the proposed loss combination is plausible. If fully validated with proper ablations and federated baselines, the method could be a useful contribution. The paper ships a concrete algorithm and reports run-level statistics. However, as written, the central empirical claims are not supported: no ablation or sensitivity analysis appears, no federated algorithm is compared against, and the theoretical justifications for the MI surrogate and KL simplification are not established.","major_comments":[{"comment":"The abstract claims 'comprehensive ablation and sensitivity analyses' and that FDRMFL 'attains the lowest overall mean MSE among six federated algorithms including FedAvg, FedProx, MOON, SCAFFOLD, and FedBN.' The body contains no ablation study, no sensitivity analysis, and no comparison with any federated algorithm. Tables 1–4 list only PCA, TSVD, RP, and VAE; §3.1 explicitly names 'PCA, TSVD, RP, MAE and FDRMFL'. The only acknowledgment of FedAvg is §2.2 ('Building upon FedAvg training...'), which does not test it. Thus the load-bearing claims of component contribution and superiority over federated baselines are unsupported.","section":"Abstract; §3.1, §3.2"},{"comment":"Eq. (15) is presented as a 'variational lower bound' for I(Z_ij; y_ij), but no derivation is given. The expression uses binary logistic losses on f(Z)·y and f(Z)·ỹ; for a continuous target y this is not any standard InfoNCE or variational MI estimator. Moreover f is the same prediction network trained by the MSE term in Eq. (14), so the 'information' being maximized is not independent of the predictor. The claim that this term 'preserves dependence' is therefore unsupported and arguably circular.","section":"§2.2, Eq. (15)"},{"comment":"The KL alignment term is reduced to a per-sample Euclidean penalty under the per-sample homoscedastic Gaussian assumption (Eq. 10). But p(z_im) is a distribution over features, not defined per sample; Eq. (16) computes symKL on point estimates z_ijm, z_ijn. The shared variance σ² is a free parameter, and the reduction to ||z_ijm−z_ijn||²/(2σ²) turns a distributional alignment objective into a pointwise distance. The mechanistic interpretation as 'cross-modal alignment at the distribution level' (§4) is not supported.","section":"§2.1, §2.2, Eqs. (10), (16), (17)"},{"comment":"The experiments do not actually create non-IID partitions. §3.1 says 'Samples are then randomly partitioned across three clients'; this is an IID split. The real-data descriptions (§3.2) mention qualitative differences ('client1 contains mostly lean samples') but no quantitative distribution-shift measure or partition protocol is given. Since non-IID is one of the three core challenges motivating the method, the reported results cannot be interpreted as evidence for the non-IID setting.","section":"§3.1, §3.2"},{"comment":"No ablation or sensitivity study is performed for the weights λ1, λ2, λ3, temperature τ, or σ². The abstract's promise of 'comprehensive ablation and sensitivity analyses' is absent. Without these, one cannot tell whether the observed MSE reductions come from the three added regularizers or simply from the supervised federated training itself.","section":"§3.1, Tables 1–4"}],"minor_comments":[{"comment":"The baseline list includes 'MAE' but no MAE results appear; presumably 'VAE' is meant. This typo obscures the baseline set and should be fixed.","section":"§3.1"},{"comment":"This section is a long textbook exposition of Transformer, CNN, LSTM, and MLP architectures. It is unrelated to the experiments and could be condensed or moved to an appendix.","section":"§2.3"},{"comment":"The target is labeled 'Water' in the table but 'moisture' in the text; standardize the terminology.","section":"Table 3"},{"comment":"The phrase 'six federated algorithms' is ambiguous: if it means FDRMFL plus five baselines, the names should be explicit; if it means six baselines, the list is incomplete. In any case, no such comparisons appear in the body.","section":"Abstract"},{"comment":"The symbol μ_im is used in the Gaussian assumption but never defined. Eq. (17) then switches to sample-level notation z_ijm, which adds confusion about what is being aligned.","section":"Eq. (10)"}],"recommendation":"reject","confidential_remarks":"The abstract and conclusions overclaim ablation studies and federated baselines that are entirely absent from the experiments. The theoretical surrogates are not justified. Even with additional experiments, the manuscript would require major restructuring; in its current form the central claims are not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper proposes a four-term federated objective (MSE, an MI surrogate, symmetric KL alignment, InfoNCE contrastive) for multimodal regression. What's new is the specific combination, not the parts; the author acknowledges that. The experiments show consistent wins over PCA/RP/TSVD/VAE on synthetic and two NIR datasets, 9/9 sub-scenarios in simulation. That's a real result, and the writing is clear enough that someone could reimplement the method.\n\nBut the abstract goes further: it says 'comprehensive ablation and sensitivity analyses' demonstrate each component contributes, and claims the lowest MSE among six federated algorithms including FedAvg, FedProx, MOON, SCAFFOLD, FedBN. None of that is in the body. No ablation is reported, no federated baseline is run, and the only federated training is vanilla FedAvg itself. Given that the paper's contribution is the loss design, this gap is load-bearing: you cannot attribute the gains over unsupervised baselines to the MI/KL/contrastive terms, because the baseline methods don't have the MSE head in the loop at all.\n\nTwo more soft spots. Eq (15) presents a mutual-information lower bound for continuous targets without derivation; it uses the same prediction network f as the MSE term, so the 'information' being preserved isn't independent of the predictor. Eq (17) reduces distributional alignment to a per-sample Euclidean penalty under an unstated Gaussian assumption, so the KL story is weaker than claimed. Minor: no code, no hyperparameter details, and the conclusion calls the work 'breakthrough progress' — overstated.\n\nOn balance, I think the direction is reasonable and the empirical wins, while limited to non-federated baselines, are consistent. But the paper is not ready. It needs an ablation of the four terms, at least one FedAvg-family comparison, and a corrected abstract. That's major revision territory, not a desk reject.","headline":"Abstract overclaims—no ablations or FedAvg baselines exist in the body—but the core loss combination is sensible and the reported wins over PCA/VAE are consistent.","tokens_in":20645,"tokens_out":2672,"would_cite":false,"duration_ms":27468,"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":"FDRMFL claims that a four-term objective—prediction error, mutual-information retention, cross-modal KL alignment, and contrastive anchoring—makes federated multimodal regression more accurate and stable under non-IID data.","keywords":["multi-modal data analysis","feature extraction","federated learning","contrastive learning","non-IID data","mutual information","KL divergence","near-infrared spectroscopy regression"],"falsifier":"Run FDRMFL head-to-head against FedAvg, FedProx, MOON, SCAFFOLD, and FedBN on the paper's own simulation protocol; the abstract claims the lowest overall mean MSE among these six, but Tables 1-4 report only PCA/TSVD/RP/VAE comparisons. Also, remove each of the three regularizers one at a time; the paper claims all contribute, yet the promised ablation tables do not appear in the full text.","tokens_in":19777,"feed_emoji":"📉","tokens_out":7001,"duration_ms":63886,"temperature":0.7,"pith_summary":"FDRMFL is a task-driven feature extraction method for federated multimodal regression. The paper argues that when data cannot leave clients and local samples are scarce and non-IID, dimensionality reduction must be supervised: instead of maximizing variance or reconstruction quality, the extractor should be trained jointly with the downstream regressor. The proposed objective combines mean-squared-error prediction loss with three regularizers—a mutual-information surrogate that keeps target-relevant information, a symmetric KL penalty that aligns modality distributions before fusion, and an InfoNCE contrastive term that anchors local representations to previous-round global representations. In simulations on three synthetic link functions and two real near-infrared spectroscopy datasets, FDRMFL reports the lowest test MSE in every client-target combination, with mean reductions of 33.8% over PCA and 43.0% over a VAE baseline, and smaller cross-client variance. If true, this provides a privacy-preserving recipe for extracting predictive features from distributed multimodal data, with direct application to food and agricultural quality control.","feed_headline":"Four-term loss beats PCA and VAE in federated non-IID regression","feed_subtitle":"Adds mutual-information, KL-alignment, and contrastive anchors to MSE to keep task-relevant features across clients.","key_machinery":"The central object is the four-term local objective L = L_pred + λ1 L_mi + λ2 L_kl + λ3 L_fcl. L_pred is mean squared error between prediction and true label. L_mi is an empirical surrogate for mutual information between the fused representation and the regression target, implemented as a sigmoid/binary cross-entropy bound with a negative-sample label. L_kl is a symmetric KL divergence between pairs of modality feature distributions; with a shared-variance Gaussian assumption it reduces to a scaled Euclidean distance between per-sample modality means. L_fcl is an InfoNCE-style contrastive loss using cosine similarity, with the previous round's global representation as the positive sample and","core_discovery":"The central claim is that the four-term objective works as a unit: the MSE term pins down regression accuracy, the mutual-information surrogate preserves dependence between the fused representation and the continuous target, the symmetric KL term aligns per-modality feature distributions before fusion, and the InfoNCE term stabilizes federated training by treating the previous global representation as the positive anchor and earlier global representations as negatives. Under a Gaussian homoscedastic approximation, the KL alignment collapses to a per-sample Euclidean penalty on modality-feature means. The paper reports that this combined objective yields the lowest MSE in all nine simulated n","pith_inferences":["Because the KL term collapses to per-sample Euclidean distances under the Gaussian assumption, we suspect the claimed 'distribution alignment' may mostly be centroid matching; a Wasserstein or MMD-based alignment could reveal whether higher-order distributional information matters.","The same objective could be adapted to classification or survival targets by replacing the MSE term and the MI surrogate, an extension the paper does not explore.","The paper's rationale implies an untested ordering of components: if MI retention is the main driver, simpler supervised dimensionality reduction might capture part of the gain; a direct comparison would isolate the contribution of the contrastive and KL terms."],"forward_implications":["If the central claim holds, federated multimodal regression can be built around a supervised objective instead of unsupervised dimension reduction, preserving task-relevant information that PCA, TSVD, and RP discard.","The same four-term recipe should transfer to different encoder families (Transformer, CNN, LSTM, MLP) and fusion functions, so the method is not tied to one architecture.","The reported reduction in cross-client MSE spread implies that contrastive anchoring and KL alignment jointly act as a stabilizer under non-IID partitions.","For near-infrared spectroscopy, the method offers a privacy-preserving pipeline for food and agricultural quality prediction without centralizing sensitive spectral data.","The performance on two real datasets suggests the method can handle small-sample, high-dimensional functional covariates."],"fun_headline_variants":["Four-term loss beats PCA and VAE on non-IID federated data","Info-max loss cuts federated regression MSE vs PCA by 34%","MI + contrastive anchors improve federated non-IID regression","Federated regression: 4-term objective outperforms FedAvg and SCAFFOLD"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The method's load-bearing premise is that the sigmoid-based surrogate in Eq. (15) truly measures how much the fused representation depends on the continuous target; if it does not, the regularization can distort features rather than preserve predictive information.","fun_headline_variants_meta":{"raw":{"variants":["Four-term loss beats PCA and VAE on non-IID federated data","Info-max loss cuts federated regression MSE vs PCA by 34%","MI + contrastive anchors improve federated non-IID regression","Federated regression: 4-term objective outperforms FedAvg and SCAFFOLD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2277,"prompt_tokens":746,"completion_tokens":1531,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":1449}},"tokens_in":490,"tokens_out":1531,"duration_ms":15283,"temperature":1.0,"reasoning_tokens":1449,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:17:20.067427+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FDRMFL head-to-head against FedAvg, FedProx, MOON, SCAFFOLD, and FedBN on the paper's own simulation protocol; the abstract claims the lowest overall mean MSE among these six, but Tables 1-4 report only PCA/TSVD/RP/VAE comparisons. Also, remove each of the three regularizers one at a time; the paper claims all contribute, yet the promised ablation tables do not appear in the full text.","supporting_citations":[],"review_version":1}