{"id":"a7ba464c-dfb5-4650-89ae-e99613842590","arxiv_id":"2505.21010","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"UAP, an alternating two-stage training protocol, improves unseen-domain accuracy in semi-supervised federated learning by aligning client and server features to a Gaussian distribution defined by the classifier weights.","lead":"This paper presents a method for training federated models when client data is unlabeled and the final test data looks different from the training data. The method, UAP, aligns server and client feature representations to a common distribution and reports large accuracy gains over prior federated and semi-supervised baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The zero-overhead transfer of the server feature distribution depends on an unenforced equality between classifier weights and per-class feature means.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue: the equality w_k^G = mu_k asserted in Section 4.1 is not enforced by the stated objective. This is the most serious correctness risk because the entire communication-free alignment protocol depends on it. If the equality fails, client-side alignment targets a distribution that the server does not generate, so the method loses its theoretical grounding. The concern is concrete and testable without needing the full pipeline; it can be checked by measuring empirical feature means after Stage-I. The paper's empirical gains could still be real if the mismatch is small or if L_CDD and L_COV provide enough alignment on their own, but the current manuscript provides no evidence for the asserted equality, no error bars, and no code. There are also reproducibility inconsistencies (e.g., the Sketch server-domain UAP accuracy is 61.67 in Table 1 but 67.92 in Table 5), which weaken the SOTA claim further. Because the reader's verdict was already CONDITIONAL and this stress-test reinforces the same concern, no verdict change is needed.","tokens_in":17246,"tokens_out":2903,"duration_ms":35691,"concrete_test":"Re-run Stage-I alone on PACS with the paper's hyperparameters. After training, compute the empirical per-class feature means mu_hat_k = (1/N_k) * sum_{x in D_s, y=k} F(x) on held-out server data and compare them with the classifier weight vectors w_k^G. Report Delta_k = ||mu_hat_k - w_k^G||_2 / ||w_k^G||_2 and the cosine similarity for every class and server domain. If any Delta_k remains large (e.g., > 0.1) at convergence, the communicated distribution N(w_k^G, lambda I) is not the actual server feature distribution, and the paper must state which loss term enforces the asserted equality or weaken the zero-overhead claim accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 asserts that after Stage-I the classifier weights satisfy w_k^G = mu_k for all k, so clients can reconstruct the server distribution as N(w_k^G, lambda I) at no communication cost. However, no term in L_server enforces that equality. L_CDD pushes features toward N(w_k^G, lambda I), and L_CE updates the classifier weights, but nothing penalizes the difference w_k^G - mu_k. For a softmax classifier, the optimal weight vector is not generally the class-mean vector; that identity holds only under restrictive conditions (for example, spherical class-conditional Gaussians with equal priors and a particular bias choice), which the paper neither states nor derives. If the equality is only approximate, then in Stage-II clients align their features to a distribution that the server features do not actually follow, so the central alignment mechanism and the 'no additional communication overhead' claim rest on an unverified premise. A loose version of the argument may still work if the mismatch is small, but the paper reports no measurement of this mismatch, and the loss design does not obviously drive it to zero.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a new problem setting, Semi-Supervised Federated Domain Generalization (S-FDG), in which a central server has limited labeled data, clients have unlabeled data, and the deployed model must generalize to an unseen domain. The proposed Unified Alignment Protocol (UAP) alternates between two stages: Stage-I trains the server so that per-class feature distributions become Gaussian with mean equal to the classifier weight vector and a known diagonal covariance; Stage-II trains clients, using pseudo-labels, to align their feature distributions to the same parametric distribution. The authors claim that the server feature distribution can be transferred to clients at zero additional communication cost because the classifier weights themselves are the distribution means. Experiments on PACS, VLCS, OfficeHome, RotatedMNIST, and TerraIncognita report large gains over SSFL and FDG baselines, including a roughly 37% absolute accuracy improvement on an unseen PACS domain.","tokens_in":17521,"tokens_out":4125,"duration_ms":44165,"significance":"If the central mechanism is sound, UAP addresses a real and under-studied gap between SSFL and FDG, with a communication-efficient way to share distributional information. The paper is the first to formulate S-FDG, and it evaluates across multiple datasets and architectures, which is a useful contribution. However, the central technical premise—that classifier weights equal per-class feature means after server training—is asserted rather than enforced or verified, and the reported experimental numbers are internally inconsistent. These issues must be resolved before the claimed SOTA results can be relied upon. The paper currently does not ship code, multiple seeds, or error bars, so the magnitude of the reported improvements cannot be independently assessed.","major_comments":[{"comment":"The assertion that after Stage-I training the classifier weights satisfy w_k^G = mu_k for all k is not enforced by any term in the server loss L_server = L_CE + alpha*L_CDD + beta*L_COV. L_CDD pulls features toward the dynamic Gaussian N(w_k^G, lambda*I), but nothing pulls the weights toward the empirical feature means. For a softmax classifier trained with cross-entropy, the optimal weight vector is not generally the class-mean vector; the identity holds only under restrictive conditions (e.g., spherical class-conditional Gaussians with equal priors and a specific bias, or terminal-phase neural collapse, which is not demonstrated here). Since this equality is the basis for the zero-communication transfer of the distribution parameters, the authors must either add an explicit loss term penalizing ||w_k^G - mu_k||, provide empirical measurements showing that the mismatch is negligible, and/or analyze how a residual mismatch affects the client alignment objective. Without this, the central premise of the method is unverified.","section":"4.1, Eq. (2)"},{"comment":"The reported UAP results on PACS are internally inconsistent. Table 1 reports UAP accuracy on the unseen Art Painting domain as 64.84 for Photo and 61.67 for Sketch, while Table 5 reports 64.40 for Photo and 67.92 for Sketch for the same experimental setting (server trained on Photo/Sketch, test on Art Painting). Table 15 lists Sketch=67.92 for the gamma=1.0 row, which conflicts with Table 1. The abstract and Section 1 also cite a ~37% improvement based on these numbers, but the gain depends on which table entry is used. The authors should reconcile these numbers and state explicitly whether different seeds, training configurations, or table construction errors are responsible.","section":"Tables 1, 2, 5, 15"},{"comment":"The experimental evaluation reports no error bars, no multiple seeds, and no code release. Given that the claimed gains are large (e.g., average improvements of 22-33% over SSFL on several domains) and that the method has several hyperparameters (alpha, beta, lambda, gamma), single-run results are insufficient to support the SOTA claim. The paper should report mean and standard deviation over at least three independent runs with different seeds, and ideally release code to enable verification.","section":"5 and 6"}],"minor_comments":[{"comment":"The hyperparameter values are inconsistent: Section 5 states gamma=100, but the ablation in Table 15 uses gamma values of 0.5, 1.0, and 2.0, with the main UAP results corresponding to gamma=1.0. Please clarify the actual value used in the main experiments.","section":"5"},{"comment":"The caption contains a typo: 'Sever Feature Alignment' should be 'Server Feature Alignment'.","section":"Figure 2 caption"},{"comment":"Reference [7] is incomplete ('Enmao Diao and et al.'); the full author list should be provided.","section":"References"},{"comment":"The pseudo-label generation via weighted k-means clustering is mentioned but not described; details on how the clustering is performed in the federated setting, how initial centroids are selected, and how the 'weighted' aspect is implemented would improve reproducibility.","section":"4.2"},{"comment":"The reference covariance matrix Sigma is defined vaguely: the text says Sigma = gamma*Sigma_k, but Sigma_k is not defined. Clarify whether it is the per-class covariance, a fixed diagonal matrix, or the empirical covariance of the current mini-batch.","section":"4.1, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is interesting, but the unverified equality between classifier weights and feature means is load-bearing, and the inconsistent numbers across tables make the empirical claims hard to trust. I would encourage the editor to ask for a revision that adds an explicit enforcement or verification of the weight-mean equality, reconciles the reported numbers, and provides multi-seed results with code. The scope fit is fine for a vision or federated learning venue, though the paper would benefit from a more rigorous treatment of the theoretical assumptions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know first: this paper introduces a genuinely new problem setting, Semi-Supervised Federated Domain Generalization (S-FDG), where clients have unlabeled data, the server has a small labeled set, and the model is evaluated on an unseen domain. That framing is timely and practical, and the method—two-stage alignment where the server first shapes its feature distribution into class-wise Gaussians and clients then align to that distribution via CDD loss plus covariance regularization—is a reasonable recipe. The reported gains are large: on PACS, up to ~22–37% absolute over the SemiFL baseline.\n\nThe paper deserves credit for the problem statement and for an ablation that shows both new loss terms contribute. The experimental setup (one domain as server, one held out, rest to clients) is sensible and matches the FDG literature.\n\nThe soft spots are mostly about evidence, not about the idea. The numbers disagree across tables: UAP on PACS with Photo server and Art Painting test is 64.84 in Table 1, 64.40 in Table 2; with Sketch server it is 61.67 vs 67.92. That is not a computational artifact; it points to inconsistent experimental runs. There are no error bars, no seeds, no code. For a claim of SOTA accuracy on a new task, that is insufficient. The ablation is helpful, but it is only on PACS.\n\nThe stress-test worry about the classifier weights being the feature means is less damaging than it looks. L_CDD pulls each class's features to N(w_k^G, lambda I), so by construction the feature means should track the weights; the equality is enforced from the feature side, not the weight side. However, the paper simply asserts the equality without measuring the mismatch. A quick diagnostic (mean distance between w_k^G and empirical class means after Stage-I) would settle it. The \"zero communication overhead\" claim rests on this, so the authors should show the numbers.\n\nThere are also citation problems: PACS is cited as the 2022 audio-visual commonsense dataset, and reference [37] contains two different papers. These look like careless errors, but they need fixing.\n\nBottom line: the problem is worth working on, the method is plausible, and the results could be important for federated learning in healthcare and on-device settings. But the paper in its current form overclaims. Send it to peer review—a serious referee should engage—and ask for code, seeds, corrected tables, and a direct verification of the weight-mean equality. I would cite it once it clears those hurdles.\n\nLet me know if you want to discuss.","headline":"A genuinely new federated problem setting with a plausible method and large claimed gains, but the evidence as presented is too uneven to accept the SOTA claim at face value.","tokens_in":18026,"tokens_out":3841,"would_cite":false,"duration_ms":41470,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A communication-free feature-alignment protocol lets unlabeled clients generalize to unseen domains in semi-supervised federated learning.","keywords":["semi-supervised federated learning","domain generalization","feature alignment","contrastive domain discrepancy","covariance matching","zero-overhead communication","pseudo-labeling","domain shift"],"falsifier":"Measure the distance between each trained classifier weight vector and the true mean of its class's server features, on any standard DG dataset, at the end of Stage-I. If that distance is comparable to the within-class feature spread, clients reconstruct a distribution the server features do not actually follow, and UAP's gains should shrink toward the SSFL baseline; conversely, a near-zero distance would confirm the mechanism.","tokens_in":17096,"feed_emoji":"🧩","tokens_out":4197,"duration_ms":41942,"temperature":0.7,"pith_summary":"The paper introduces Semi-Supervised Federated Domain Generalization (S-FDG), a setting in which only the server has labeled data, clients hold unlabeled data from their own domains, and the trained model must generalize to a domain none of them saw. It claims existing semi-supervised federated learning methods fail under domain shift, and existing federated domain generalization methods fail without client labels. To close this gap, it proposes the Unified Alignment Protocol (UAP), an alternating two-stage training scheme in which the server first learns per-class Gaussian feature distributions whose means are stored in the classifier weights, then clients align their pseudo-labeled features to that communicated distribution. The paper reports substantial gains over state-of-the-art baselines, including roughly 37% higher accuracy on an unseen PACS test domain, with no additional communication overhead.","feed_headline":"Zero-overhead alignment lifts unseen-domain accuracy by ~37%","feed_subtitle":"Unlabeled clients align to the server's Gaussian features, with no extra communication cost.","key_machinery":"The load-bearing object is the per-class dynamic Gaussian distribution $q_k = N(w_k^G, \\lambda I)$ built from the server classifier's weight vectors. Contrastive Domain Discrepancy (CDD) loss, a class-conditioned maximum mean discrepancy, pulls server features toward $q_k$ while pushing different classes apart, and a covariance matching loss $L_{\\text{COV}}$ pushes both server and client feature covariances toward a diagonal reference $\\gamma\\Sigma_k$. The trick that makes the protocol communication-free is that the client never needs the mean vector $\\mu_k$ sent separately: the protocol assumes the classifier weights end up equal to the feature means, so the weights themselves are the distribution parameters.","core_discovery":"On the paper's own terms, the central discovery is that a semi-supervised federated system can generalize to unseen domains if the server's per-class features are shaped into a known parametric distribution whose parameters ride for free inside the already-sent model weights. The server trains with $L_{\\text{CE}} + \\alpha L_{\\text{CDD}} + \\beta L_{\\text{COV}}$ so that classifier weight $w_k^G$ equals the class-$k$ feature mean $\\mu_k$ and the class covariance collapses to $\\lambda I$; clients then reconstruct $N(w_k^G, \\lambda I)$ as the server feature distribution, pseudo-label their unlabeled data, and minimize the same CDD and covariance losses against it. The result is a training protocol that claims state-of-the-art S-FDG performance, with the headline PACS gain from 52.20% to 75.73% on the Art Painting test domain compared to the SOTA SSFL method.","pith_inferences":["If the weight-mean equality is only approximate, one could add an explicit loss term that pulls $w_k^G$ toward the feature means, which would likely strengthen the client-side alignment; the paper's own sensitivity to $\\lambda$ suggests the margin in the current protocol is testable.","The same embedding trick could generalize to other parametric families, such as mixture models or von Mises-Fisher distributions for directional features, letting clients reconstruct richer server distributions at zero communication cost.","In cross-institution healthcare deployment, the protocol suggests that unlabeled hospital clients could improve a shared model's performance on new patient populations without sharing raw data; a natural next test is on medical imaging benchmarks with stricter privacy guarantees."],"forward_implications":["This is the first formulation of S-FDG; the paper's evaluation setup, with one labeled server domain, unlabeled client domains, and an unseen test domain, gives future work a concrete benchmark to compare against.","UAP consistently improves unseen-domain accuracy over SSFL baselines across PACS, VLCS, OfficeHome, RotatedMNIST, and TerraIncognita, and across VGG11, ResNet18, DenseNet121, and DeiT-B architectures.","The alignment costs no additional communication overhead, so the generalization gains come within the same bandwidth budget as a plain SSFL training round.","The combination of CDD and covariance matching is robust to noisy pseudo-labels; ablation results show each loss component contributes to the final gain.","The method is insensitive to the number of clients in the tested range, with only a small drop as decentralization increases, matching the usual federated learning trend."],"supporting_citations":[{"why":"Defines the SOTA SSFL method with alternate training that UAP extends and compares against as its primary baseline.","marker":"[7]"},{"why":"Supplies the Contrastive Domain Discrepancy loss that aligns features to the dynamic Gaussian distributions.","marker":"[13]"},{"why":"Provides the source hypothesis transfer approach with weighted k-means pseudo-labeling used for client unlabeled data.","marker":"[20]"},{"why":"Introduces the first semi-supervised federated learning framework that the paper builds upon.","marker":"[12]"},{"why":"A federated domain generalization baseline and the source of the RotatedMNIST network architecture used in experiments.","marker":"[32]"},{"why":"A federated domain generalization baseline that relies on client labels and underperforms in the S-FDG setting.","marker":"[23]"},{"why":"A federated domain generalization baseline used in the comparative tables.","marker":"[47]"},{"why":"Motivates the diagonal covariance reference matrix by showing decorrelated representations generalize better.","marker":"[4]"},{"why":"A semi-supervised federated learning baseline that is compared in the SOTA tables.","marker":"[17]"}],"fun_headline_variants":["Zero-overhead alignment lifts unseen-domain accuracy","Federated clients align to server's Gaussian for unseen domains","Server's feature shape guides unlabeled clients to new domains","No communication spares: align client features to server's distribution","SSFL gains on new domains via server's parametric alignment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole protocol assumes that after server training the classifier weight $w_k^G$ actually equals the per-class feature mean $\\mu_k$, but no loss term explicitly enforces that equality; it is only encouraged indirectly by pulling features toward the current weights.","fun_headline_variants_meta":{"raw":{"variants":["Zero-overhead alignment lifts unseen-domain accuracy","Federated clients align to server's Gaussian for unseen domains","Server's feature shape guides unlabeled clients to new domains","No communication spares: align client features to server's distribution","SSFL gains on new domains via server's parametric alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000316,"raw_usage":{"total_tokens":1812,"prompt_tokens":987,"completion_tokens":825,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":745}},"tokens_in":603,"tokens_out":825,"duration_ms":8518,"temperature":1.0,"reasoning_tokens":745,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:42:59.798562+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the distance between each trained classifier weight vector and the true mean of its class's server features, on any standard DG dataset, at the end of Stage-I. If that distance is comparable to the within-class feature spread, clients reconstruct a distribution the server features do not actually follow, and UAP's gains should shrink toward the SSFL baseline; conversely, a near-zero distance would confirm the mechanism.","supporting_citations":[{"cited_title":"Class balanced adap- tive pseudo labeling for federated semi-supervised learning","cited_arxiv_id":null,"evidence_quote":"A semi-supervised federated learning baseline that is compared in the SOTA tables."},{"cited_title":"Semifl: Semi-supervised feder- ated learning for unlabeled clients with alternate training","cited_arxiv_id":null,"evidence_quote":"Defines the SOTA SSFL method with alternate training that UAP extends and compares against as its primary baseline."},{"cited_title":"Contrastive adaptation network for unsupervised do- main adaptation","cited_arxiv_id":null,"evidence_quote":"Supplies the Contrastive Domain Discrepancy loss that aligns features to the dynamic Gaussian distributions."},{"cited_title":"Do we really need to access the source data? source hypothesis transfer for un- supervised domain adaptation","cited_arxiv_id":null,"evidence_quote":"Provides the source hypothesis transfer approach with weighted k-means pseudo-labeling used for client unlabeled data."},{"cited_title":"Tuan Nguyen, Philip Torr, and Ser-Nam Lim","cited_arxiv_id":null,"evidence_quote":"A federated domain generalization baseline and the source of the RotatedMNIST network architecture used in experiments."},{"cited_title":"Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous fre- quency space","cited_arxiv_id":null,"evidence_quote":"A federated domain generalization baseline that relies on client labels and underperforms in the S-FDG setting."},{"cited_title":"Federated domain general- ization with generalization adjustment","cited_arxiv_id":null,"evidence_quote":"A federated domain generalization baseline used in the comparative tables."}],"review_version":1}