{"id":"fe92d7da-17e6-46ec-8bdd-af74c5931419","arxiv_id":"2502.03340","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A loss-trajectory-based clustering method for federated learning, plus a Wasserstein-adjusted cluster metric, reports improved personalization on heterogeneous data.","lead":"FedGWC, a clustered federated learning method, groups clients by the similarity of their loss trajectories using Gaussian weighting, then trains a personalized model per cluster. The paper also introduces a Wasserstein-based cluster quality metric and reports accuracy gains on heterogeneous federated benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's evidence does not establish that FedGWC groups clients by class distribution: the WAS metric ranks class frequencies and is blind to label identity, and loss-to-mean rewards cannot distinguish label-permuted clients with symmetric losses.","rationale":"I read the paper in good faith as claiming that FedGWC clusters clients by similarity of their data distributions, using loss trajectories as a proxy, and that the Wasserstein Adjusted Score validates the resulting clusters. The reader's weakest-assumption analysis correctly identifies the proxy assumption as the fragile link. My stress-test sharpens this: the actual operational signal is only proximity of scalar losses to the federation mean, which is neither necessary nor sufficient for class-distribution similarity, and the paper's own evaluation metric ranks class frequencies, making it invariant to which classes a client holds. This means the main experimental support for the central claim (Tables 1 and 3) does not actually establish that FedGWC finds class-homogeneous clusters. The domain-detection experiments are more convincing but concern visual domain shift, not class-label distribution. The theoretical section adds a further concern because the stationarity and i.i.d. assumptions behind Theorems 5.1 and 5.2 are not met during optimization, but I do not treat that as the primary issue since the central claim is empirical. I nonetheless do not recommend rejection: the method may still work in practice for the tested heterogeneity levels, and a controlled synthetic test can settle whether the concern lands. The proposed test is straightforward, uses publicly available data and models, and directly targets the claimed mechanism rather than relying on the paper's own metric. For these reasons, the appropriate verdict remains CONDITIONAL, unchanged from the reader's assessment.","tokens_in":24564,"tokens_out":7988,"duration_ms":88606,"concrete_test":"Construct a synthetic Cifar10 federation with 50 clients holding only class 0 and 50 clients holding only class 1 (or any label-permuted pair with comparable per-class difficulty). Run FedGWC with FedAvg and compare the recovered partition to the ground-truth two-group labeling using Rand Index, and also compute the pairwise WAS distance between a class-0 client and a class-1 client. If the Rand Index is near 0.5 and/or the WAS distance between clients from different ground-truth groups is near 0, then loss-trajectory similarity and the proposed metric cannot certify class-distribution homogeneity, and the central empirical claim needs to be substantially qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FedGWC identifies homogeneous clusters, i.e., groups of clients with similar class distributions, and that this is validated by the new Wasserstein Adjusted Score. Both parts of this claim have a load-bearing weakness. First, the clustering signal is the Gaussian reward in Eq. 1, which measures how close a client's scalar loss trajectory is to the average loss trajectory of the sampled clients. This is a proximity-to-mean measure, not a measure of the shape of the loss landscape and not a measure of class identity. Two clients with disjoint but equally learnable class distributions can have nearly identical loss trajectories, so they receive high rewards and are kept together, even though their data distributions are maximally different. Second, the proposed WAS/WADB metrics cannot detect this failure: they rank each client's class-frequency vector and compute Wasserstein distance between the resulting multisets of frequencies. This makes the metric invariant under permutations of class labels. A client holding only class 0 and a client holding only class 1 have identical ranked frequency vectors, so their WAS distance is exactly 0. Thus the high WAS/low WADB values reported in Tables 1 and 3 do not demonstrate distributional homogeneity in the sense of shared class identity; they only demonstrate similar imbalance skewness. The domain experiments in Table 4 use Rand Index and are more direct, but domain shift (clean/noisy/blurred) is not the same as class-distribution shift. The theoretical convergence results in Section 5 also assume a stationary loss process and i.i.d. reward samples, which is not satisfied during training, but the more decisive issue is that the empirical validation of the central claim rests on a metric that is structurally blind to the property being claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedGWC, a clustered federated learning algorithm that groups clients using a Gaussian reward mechanism applied to scalar loss trajectories. At each round, a client's reward is computed from proximity of its loss process to the mean loss process of sampled clients; the reward is averaged over local iterations and updated with an exponential running average. These weights populate an interaction matrix from which an RBF affinity matrix is built and spectral clustering is applied recursively, with the cluster count selected by a Davies-Bouldin criterion. The paper also introduces a Wasserstein Adjusted Score for evaluating cluster cohesion under class imbalance. Experiments cover Cifar100, Femnist, Google Landmarks, and iNaturalist, comparing FedGWC against IFCA, FeSEM, CFL, FedAvg, FedAvgM, FedProx, and FairAvg, and claiming superior accuracy and cluster quality.","tokens_in":24837,"tokens_out":4904,"duration_ms":53618,"significance":"The design is attractive for cross-device FL: clients communicate only scalar losses, the number of clusters is selected automatically, and the method is orthogonal to the aggregation rule. The paper ships a reasonably complete experimental campaign spanning small, large-scale, class-imbalanced, and domain-shift settings. If the central claim about distributional homogeneity were established, the contribution would be useful to the clustered-FL community. However, the evidence as written does not support the strong claims made in the abstract and contributions: Table 1 contradicts the accuracy claim on Femnist, the proposed WAS metric is invariant to class-label permutations and therefore cannot certify distributional homogeneity, and the theoretical results reduce to convergence of a running average to its mean. These are load-bearing gaps rather than presentation issues.","major_comments":[{"comment":"The central empirical claim that FedGWC outperforms existing FL algorithms is not supported on Femnist. In Table 1, FedGWC achieves 76.1±0.1 on Femnist, while FedAvgM reaches 83.3±0.3 and FedAvg reaches 76.6±0.1; on Cifar100 FedGWC ties FeSem at 53.4±0.4 vs 53.4±1.8 rather than clearly outperforming it. Table 9 further shows FedGWC+FedAvgM on Femnist gives 83.3±0.4, i.e. no improvement over no clustering. The text in §7.1 concedes that clustering-based methods perform worse on Femnist, which is in tension with the unqualified claim in the abstract and contributions. The claim should be scoped to the settings where the data support it, or additional evidence should be added.","section":"§7.1, Table 1 and Table 9"},{"comment":"The Wasserstein Adjusted Score is computed on rank-ordered class frequency vectors, making it invariant under permutations of class labels. Consequently, a client holding only class A and a client holding only class B have identical ranked frequency vectors, so their WAS distance is exactly zero. The high WAS and low WADB values reported in Tables 1 and 3 therefore do not establish that FedGWC groups clients by shared class identity; they at most measure similarity of imbalance skewness. This is a load-bearing limitation for the paper's cluster-quality validation and should be addressed directly, for example by also reporting label-aware metrics or by testing on label-permuted synthetic partitions.","section":"§6, Eq. (7), and Appendix B"},{"comment":"The clustering signal is a proximity-to-mean measure on scalar loss trajectories, not a measure of class identity. Two clients with disjoint but equally learnable class distributions can produce nearly identical loss trajectories, receive similar Gaussian rewards, and be placed in the same cluster even when their data distributions are maximally different. The paper's premise in §1 that similar data distributions produce similar loss landscapes is an assumption, not an established fact, and the experiments do not test it separately from the reward construction. A concrete test would be a synthetic experiment in which class labels are permuted between two equal-size client groups; this would show whether FedGWC separates clients by data identity or only by loss-difficulty profile.","section":"§4.1, Eq. (1)"},{"comment":"The theoretical results are considerably weaker than the text suggests. Since Eq. (2) defines the Gaussian weight as an exponential running average of rewards whose expectation is µk, Theorems 5.1 and 5.2 essentially state that the average converges to its mean, which is the standard Robbins-Monro guarantee and does not depend on the specific Gaussian construction. The proofs do not analyze cluster recovery, identifiability of the interaction matrix, or consistency of the split decision. Calling this a rigorous framework that provides 'theoretical guarantees for our clustering approach' (§1, Contributions) overstates the scope; the theorems can at most support the claim that the weight estimates are consistent estimators of the expected reward.","section":"§5, Theorems 5.1 and 5.2, and Appendix A"},{"comment":"The domain-detection success is only partial. In the (40, 30, 30) clean/noise/blur split of Cifar100, the ground truth has three domains, yet FedGWC detects four clusters and obtains a Rand Index of only 0.60, whereas the two-domain rows reach 1.0. The sentence 'Consistently FedGWC accurately distinguishes all visual domains' is contradicted by this row. Additionally, in Table 3, FedGWC reports C=5 on two-group Cifar100 splits (10,0,90) and (10,90,0), indicating that the automatic cluster-count selection can substantially over-split. These discrepancies should be discussed and either explained or used to temper the clustering-quality claims.","section":"§7.3, Table 4"}],"minor_comments":[{"comment":"The interaction-matrix update uses the same ωt_k for every co-sampled client j in Pt×Pt, so each row is constant across the sampled clients at a given round. This makes the raw update non-pairwise; the paper should clarify how the UPV and RBF stages convert this row-constant signal into a valid pairwise interaction, or justify why the simplification is harmless.","section":"§4.2, Eq. (3)"},{"comment":"The paragraph beginning 'In Table 1, we present a comparative analysis' is duplicated almost verbatim later in the same section; one copy should be removed.","section":"§7.1"},{"comment":"The word 'anaylsis' should be corrected to 'analysis', and the phrase 'Wasserstein's Adjusted Score' should be made consistent with the term 'Wasserstein Adjusted Score' used elsewhere.","section":"§7.3"},{"comment":"The text refers to a baseline as 'e.g. FedA'; this appears to be a typo for 'FedAvg' and should be corrected.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a useful algorithmic core and a broad evaluation, but the abstract and contribution statements currently promise more than the evidence delivers. The label-permutation blindness of WAS and the Femnist results are the two most serious issues; both are addressable with additional experiments and more careful claims. A revision should also include an explicit limitations paragraph, as the current text never acknowledges the possibility that loss-process similarity may fail to track class-distribution similarity."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: FedGWC is a plausible new entry in clustered FL, but the headline claim — that it groups clients by class distribution — is not established by the evidence. The evaluation metric is blind to label identity, and the loss signal doesn't track class identity either. Still worth a serious referee: the algorithm is clean, the domain-separation results are real, and the flaws are fixable.\n\nThe genuinely new piece is the combination: Gaussian rewards on empirical loss trajectories, an interaction matrix, RBF affinities, spectral clustering with automatic cluster-count selection. That combination isn't in the literature, and the algorithm is simple and cheap: clients send scalar losses, the server does the clustering, and any aggregator can be dropped in. The domain experiments are the most convincing part — on clean/noisy/blurred Cifar100 and Cifar10, FedGWC separates domains with Rand Index 1.0 in most settings. The theory (Robbins-Monro convergence of the weights) is correct but modest: it says the running average converges to its mean, which is true by construction.\n\nThe problems start with the abstract's \"outperforms existing FL algorithms.\" Table 1 shows a tie with FeSEM on Cifar100 (53.4 vs 53.4) and a clear loss to FedAvgM on Femnist (76.1 vs 83.3). The balanced accuracy gains appear mainly on Cifar100, where clustering is most needed. That's a qualified claim, not the one in the abstract.\n\nThe bigger issue is the central claim about class-distribution homogeneity. The stress-test note holds up: the Wasserstein Adjusted Score is computed on ranked class-frequency vectors, which makes it invariant under permutations of class labels. A client with only class 0 and a client with only class 1 have identical ranked vectors and WAS distance 0. So the high WAS/low WADB values in Tables 1 and 3 don't demonstrate that clusters share class identity — they only show that clients within a cluster have similar imbalance skewness. Similarly, the Gaussian reward measures proximity to the average loss trajectory, which can be high for clients with disjoint but equally learnable class distributions. The domain results are evidence that the method can separate visual domains, but domain shift is not the same as class-label shift, and the paper never tests the label-permuted case directly.\n\nThe convergence proofs assume a stationary loss process and i.i.d. reward samples, which is a real assumption, but that is a minor issue relative to the metric problem.\n\nBottom line: this deserves peer review, not desk rejection. The core idea is worth replicating, the domain detection results are genuinely good, and the metric issue is fixable — either change the metric to something label-aware or temper the homogeneity claim. I'd ask the authors for code and a label-permutation experiment before acceptance.","headline":"A clean new clustering algorithm for FL with real domain-separation results, but the class-distribution homogeneity claim is under-supported because the evaluation metric is blind to label identity.","tokens_in":25457,"tokens_out":3017,"would_cite":false,"duration_ms":28033,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H30","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"FedGWC claims that grouping federated clients by the similarity of their loss trajectories, via Gaussian rewards on the empirical loss process, yields homogeneous clusters and better classification accuracy than existing…","keywords":["federated learning","clustered federated learning","data heterogeneity","client clustering","Gaussian weighting","spectral clustering","Wasserstein distance","class imbalance"],"falsifier":"Run FedGWC on synthetic federations with known per-client class distributions and compare the within-cluster Wasserstein distance of the detected clusters to that of random same-size partitions: if the clustered distances are not clearly lower, loss trajectories are not tracking distribution similarity, and the central claim collapses.","tokens_in":24340,"feed_emoji":"📊","tokens_out":6827,"duration_ms":61818,"temperature":0.7,"pith_summary":"FedGWC claims that in federated learning the evolution of each client's training loss is enough to recover which clients share a data distribution, and it turns that intuition into a clustering algorithm: each client gets a Gaussian reward for how well its loss trajectory tracks the federation average, rewards accumulate into an interaction matrix, and spectral clustering on that matrix partitions the clients. The paper argues this loss-only route is cheaper and more automatic than existing clustered-FL methods, because clients send only scalar loss values (plus their usual model updates), the number of clusters is decided by a split rule rather than a user-supplied constant, and any aggregation rule can be plugged in underneath. On Cifar100, Femnist, Google Landmarks, and iNaturalist, FedGWC is reported to beat both standard federated averaging and clustered baselines (IFCA, FeSEM, CFL) in balanced accuracy and cluster-quality metrics. A companion contribution is the Wasserstein Adjusted Score, which adapts silhouette or Davies-Bouldin scores to class imbalance by sorting clients' class frequencies before measuring distances, a step the paper proves equivalent to using the Wasserstein distance between empirical class distributions.","feed_headline":"Loss-curve clustering lifts federated accuracy 12 points","feed_subtitle":"FedGWC groups clients by Gaussian-weighted loss similarity, detects the cluster count itself, and works with any aggregator.","key_machinery":"The engine is the Gaussian reward combined with a running-average interaction matrix. At each round, the server forms the mean loss $m_{t,s}$ and sample variance $(\\sigma_{t,s})^2$ across sampled clients, rewards each client with a Gaussian kernel $r^{t,s}_k$, averages over the $S$ local steps to get $\\omega^t_k$, and updates the client weight $\\gamma^{t+1}_k = (1-\\alpha_t)\\gamma^t_k + \\alpha_t \\omega^t_k$ and the interaction matrix $P^{t+1}_{kj} = (1-\\alpha_t)P^t_{kj} + \\alpha_t\\omega^t_k$ for co-sampled clients. The weight update is a Robbins-Monro stochastic approximation, which the paper uses to prove convergence of $\\Gamma^k_t$ to $\\mu_k$ (the expected reward) almost surely under an $\\ell^2/\\ell^1$ step-size condition and in expectation for constant $\\alpha$, plus a variance-reduction bound. The interaction matrix is then transformed into an affinity matrix via unbiased perception vectors and an RBF kernel, and spectral clustering over this affinity matrix produces the client partition.","core_discovery":"The central discovery is that pairwise client similarity can be estimated from the empirical loss process alone, without gradients or data: rewards defined as $r^{t,s}_k = \\exp\\left(-(l^{t,s}_k - m_{t,s})^2 / (2(\\sigma_{t,s})^2)\\right)$ measure how close client $k$'s loss at round $t$, local step $s$, stays to the mean loss of the sampled clients, and running averages of these rewards converge (as $t\\to\\infty$) to the client's theoretical expected reward, with strictly smaller variance. The paper then organizes these estimates in an interaction matrix $P$, symmetrizes it through unbiased perception vectors and an RBF kernel into an affinity matrix $W$, and applies spectral clustering; a Davies-Bouldin-based rule decides whether to split a cluster, so the number of clusters is detected rather than fixed. Experiments reported on Cifar100, Femnist, Google Landmarks, and iNaturalist show accuracy gains (for example, 53.4 vs. 41.6 balanced accuracy versus FedAvg on Cifar100) and better separation of clients by heterogeneity level and visual domain.","pith_inferences":["Because the server only needs scalar loss traces, FedGWC's privacy story is stronger than gradient-based clustering; a natural next test is whether the loss trajectories themselves leak class information under membership inference, which the paper does not analyze.","The same loss-reward mechanism could be reused outside federated learning, for example to partition tasks among agents with different difficulty regimes in decentralized multi-task learning.","The convergence theorems assume the reward process is stationary, but during real training losses drift downward; a testable extension is to analyze whether the stationary assumption holds under the paper's 'single cluster initially, then split' schedule, or whether the weights need a forgetting factor.","A direct empirical check of the core premise would be to measure the correlation between the learned interaction matrix $P$ and the true Wasserstein distances between clients' class distributions, a number the paper does not report directly."],"forward_implications":["Automatic cluster detection: FedGWC decides the number of clusters from the data via the Davies-Bouldin score, removing the preset cluster count that IFCA, FeSEM, and similar methods require.","Plug-in personalization: FedGWC sits on top of any existing FL aggregator; the reported tables show FedAvg, FedAvgM, and FedProx all gain accuracy in heterogeneous settings when run per cluster.","Out-of-distribution detection: experiments that separate clean, noisy, and blurred domains with near-perfect Rand Index suggest the same machinery can flag anomalous or malicious clients by their loss behavior.","A fairer clustering benchmark: the Wasserstein Adjusted Score gives class-imbalance-aware cluster-quality numbers, so future FL clustering methods can be compared on distributional cohesion rather than accuracy alone."],"supporting_citations":[{"why":"Defines the federated averaging setup that FedGWC builds on and provides the main standard aggregation baseline.","marker":"(McMahan et al., 2017)"},{"why":"Supplies the premise that clients with similar data distributions exhibit similar loss landscapes, backing the loss-based clustering idea.","marker":"(Cho et al., 2022)"},{"why":"CFL is the principal clustered-FL baseline FedGWC compares against in automatic cluster detection and accuracy.","marker":"(Sattler et al., 2020)"},{"why":"IFCA is the fixed-cluster-count baseline that FedGWC must match or beat without requiring clients to evaluate every cluster model.","marker":"(Ghosh et al., 2020)"},{"why":"FeSEM is the multi-center clustering baseline used in the comparison tables.","marker":"(Long et al., 2023)"},{"why":"Provides the spectral clustering algorithm used to partition the affinity matrix into client clusters.","marker":"(Ng et al., 2001)"},{"why":"Supplies the Davies-Bouldin score that decides whether to split a cluster and is also adapted into the Wasserstein Adjusted Score.","marker":"(Davies & Bouldin, 1979)"},{"why":"The Wasserstein distance on class-frequency vectors is the theoretical basis for the paper's new clustering metric.","marker":"(Kantorovich, 1942)"},{"why":"Provides the stochastic approximation framework used to prove convergence of the Gaussian weight updates.","marker":"(Robbins & Monro, 1951)"},{"why":"Defines the large-scale real-world federated partitions on Google Landmarks and iNaturalist used in the evaluation.","marker":"(Hsu et al., 2020)"}],"fun_headline_variants":["Gauss-weighted loss curves cluster federated clients","Loss similarity clustering lifts federated accuracy 12 points","Auto-detect client clusters via loss rewards in federated learning","Cluster federated clients by loss interaction, no data needed","Gaussian reward interaction matrix improves federated clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that clients with similar data distributions produce similar loss trajectories, so loss similarity can stand in for the unseen data-distribution similarity that clustering actually wants.","fun_headline_variants_meta":{"raw":{"variants":["Gauss-weighted loss curves cluster federated clients","Loss similarity clustering lifts federated accuracy 12 points","Auto-detect client clusters via loss rewards in federated learning","Cluster federated clients by loss interaction, no data needed","Gaussian reward interaction matrix improves federated clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1701,"prompt_tokens":961,"completion_tokens":740,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":663}},"tokens_in":577,"tokens_out":740,"duration_ms":7956,"temperature":1.0,"reasoning_tokens":663,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:04:41.278511+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FedGWC on synthetic federations with known per-client class distributions and compare the within-cluster Wasserstein distance of the detected clusters to that of random same-size partitions: if the clustered distances are not clearly lower, loss trajectories are not tracking distribution similarity, and the central claim collapses.","supporting_citations":[],"review_version":1}