{"id":"f68649d9-6b36-4551-95e2-3f047b0bff57","arxiv_id":"2506.21012","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"FedSC uses angular-neighbor relational prototypes and discrepancy-weighted consistent prototypes to reduce the accuracy loss of federated learning under non-IID client data.","lead":"FedSC is a federated learning method that improves training when client data is unevenly distributed by sharing class-level feature averages, called prototypes, between clients and the server. It builds two new prototype types, relational and consistent, and uses them in contrastive and regularization losses, reporting accuracy gains over existing methods on several image benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Convergence proof uses Assumption 3 as a lower bound though it is stated as an upper bound, and Eq. (10) mixes scalar similarities with vector norms; Theorems 1–3 are therefore unsupported.","rationale":"The reader's weakest_assumption pointed to Assumptions 1–4 being unverified for the prototype losses. I agree with that direction, but the more load-bearing problem is that even granting the assumptions, the proof of Theorems 1–3 contains invalid bounding steps. The strongest claim explicitly includes 'a guaranteed convergence rate'; that is a mathematical assertion, and it is falsifiable by checking the derivation. The empirical contribution is separate: the paper reports consistent gains in Tables 1–4, provides ablations for RPCL/CPDR and for M and τ, compares metrics in Table 5, and releases code. Those are real independent support for the empirical claim. However, the convergence theorem is not merely an accessory; Section 4.3 presents it as assurance. If the proof cannot be repaired, the authors should weaken the claim to 'empirically fast convergence' rather than 'guaranteed convergence rate'. The uniform-distribution assumption H in Eq. (10) is a modeling idealization, not a proof error; label-shift/long-tail experiments would test it, but it is not the central load-bearing flaw. The invalid use of Assumption 3 is internal inconsistency, not a disagreement with consensus, so it directly affects correctness. A corrected proof or a removed theoretical claim would resolve the issue; the empirical comparison could stand either way. I therefore keep the reader's CONDITIONAL verdict.","tokens_in":21905,"tokens_out":6809,"duration_ms":76474,"concrete_test":"Symbolically re-derive the inequality after Eq. (12) using Assumption 3 exactly as stated (E‖∇L‖² ≤ B). If the derivation of −(η − L1η²/2) E B² requires a lower bound, the proof fails; this can be checked by carrying the Σ∥∇L∥² term through without substituting B². Separately, choose a one-dimensional example for Eq. (10) step (ii): let ||R_{r+2}|| < ||R_{r+1}|| but reorient prototypes so that the RPCL log-sum-exp increases; if the claimed inequality fails, Theorem 1's round-to-round bound is invalid. Finally, numerically compute ∇_w L_RPCL including dU/dw for a small model; if the Lipschitz constant of L_RPCL is not bounded as U→0, Assumption 1 is not satisfied.","verdict_should_be":"UNCHANGED","load_bearing_attack":"FedSC's central claim includes a guaranteed convergence rate (Theorems 1–3). The proof in Appendix A.3.1 does not establish it. Two concrete defects: (i) In Eq. (10), step (ii) takes the similarity s(z_i,r_j) from Eq. (13), a scalar, and bounds it as though it were proportional to the prototype vector, then bounds a log-sum-exp of similarities by differences of prototype norms without any Lipschitz constant for exp. The normalizer U in Eq. (13) is data-dependent and not covered by Assumptions 1–4, so the round-to-round change of L_RPCL is not controlled. (ii) The decisive descent step after Eq. (12) replaces −(η − L1η²/2) Σ_e ∥∇L_{rE+e}∥² with −(η − L1η²/2) E B². Assumption 3 states Eζ∥∇L(w,ζ)∥² ≤ B; it upper-bounds the squared gradient norm. Using it as a lower bound to create the negative drift −E B² is backwards. Without a lower bound on gradient norms (or a Polyak–Łojasiewicz condition), the monotone decrease in Theorem 2 and the rate in Theorem 3 do not follow. Thus the 'guaranteed convergence' part of the central claim is not supported by the provided analysis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedSC, a federated learning method that augments standard cross-entropy training with two prototype-based losses: Relational Prototypes Contrastive Learning (RPCL), which pulls instance embeddings toward client-combined class prototypes selected by angular similarity, and Consistent Prototypes Discrepancy Regularization (CPDR), which penalizes distance to globally aggregated prototypes weighted by client distribution discrepancy. The authors claim that this semantic-aware collaboration improves accuracy under label shift, long-tailed imbalance, and few-shot settings, and they provide Theorems 1–3 claiming a convergence guarantee and rate under non-convex objectives. Experiments on CIFAR10, CIFAR100, TinyImageNet, and FC100 compare against several FL baselines and report consistent accuracy gains, with ablations for the two modules and hyperparameters.","tokens_in":22199,"tokens_out":3091,"duration_ms":37414,"significance":"If the empirical claims hold, FedSC is a simple and practically relevant contribution: it combines prototype-based semantic information with contrastive and regularization objectives, and the authors release code and evaluate across several heterogeneity regimes. The conceptual distinction between relational prototypes and consistent prototypes is reasonable, and the ablation study suggests both modules contribute. However, the paper's stated theoretical contribution—a guaranteed convergence rate—is not established by the provided proofs, and the experimental reporting lacks statistical uncertainty measures. The empirical headline is directionally supported, but the convergence guarantee is a load-bearing part of the paper's claims and is currently unsupported.","major_comments":[{"comment":"The step labeled (ii) is mathematically invalid: it treats the scalar similarity s(z_i, r_j) from Eq. (13) as if it were comparable to the vector r_j, writing s(x,r_j) <= r_j and then bounding log-sum-exp differences by differences of prototype norms. A scalar similarity cannot be bounded by a vector norm in this way, and the log-sum-exp function is not controlled by the proposed inequality without Lipschitz constants for the exponential. The normalizer U in Eq. (13) is data-dependent and is not covered by Assumptions 1–4, so the round-to-round change of L_RPCL is not established.","section":"Appendix A.3.1, Eq. (10)"},{"comment":"Assumption 3 states E[||∇L(w,zeta)||^2] <= B, which is an upper bound on the squared gradient norm. In Eq. (13) the coefficient (L1/2 eta^2 - eta) is negative for the small learning rates considered, so replacing the sum of squared gradient norms by E B^2 reverses the inequality: one would need a lower bound on gradient norms, or a Polyak–Lojasiewicz condition, to obtain the claimed negative drift. Consequently, the monotone decrease in Theorem 2 and the rate in Theorem 3 do not follow from the provided analysis.","section":"Appendix A.3.1, final step of Eq. (13); Appendix A.3.2, Eq. (14)"},{"comment":"The convergence analysis assumes L = L_CE + L_RPCL + L_CPDR satisfies smoothness, bounded variance, bounded stochastic gradients, and Lipschitz continuity of the feature extractor, but the paper never verifies these properties for the prototype-based losses. The similarity in Eq. (13) involves the data-dependent normalizer U and a softmax over relational prototypes, and the CPDR term depends on consistent prototypes formed by discrepancy-weighted aggregation; none of these are shown to satisfy the assumptions. Theorems 1–3 therefore rely on unverified assumptions.","section":"Section 4.3 and Appendix A.2"},{"comment":"The central accuracy claim is supported only by point estimates. No standard deviations or seed counts are reported, and the hyperparameters M and tau are tuned on the same CIFAR10 NID1 0.2 configuration used for the headline results, which makes the statement that FedSC 'universally outperforms' other methods stronger than the evidence supports.","section":"Tables 1–3 and Fig. 5"},{"comment":"The consistent prototypes assume a hypothetical global category distribution H that is uniform over classes. In long-tailed scenarios this is an idealization that may bias the discrepancy weights and the resulting prototypes; the paper does not discuss how sensitive the results are to this choice or whether a non-uniform H would change the conclusions.","section":"Section 3.4, Eq. (10)"}],"minor_comments":[{"comment":"The displayed formula for the similarity s(z_i, r_j) contains malformed spacing and an unclear decomposition; please rewrite it so that the normalization factor U and the cosine similarity are explicit.","section":"Eq. (13)"},{"comment":"The CPDR loss is written as a sum of squared differences; the notation is fine, but the paper should state explicitly whether this is a squared L2 distance (which is what Eq. (15) computes) rather than an unsquared distance.","section":"Eq. (15)"},{"comment":"The comparison of neighbor-selection metrics reports only final accuracies without variance; adding repeated-run statistics would strengthen the claim that angle-wise selection is superior.","section":"Table 5"},{"comment":"The global model aggregation uses n_k/N weights over the selected client subset, but N is defined as the total number of samples among all clients; if the subset is not representative, the aggregation weights should be clarified.","section":"Algorithm 1, line 8"},{"comment":"The sentence 'FedSC universally outperforms other methods' is a summary of the tables, but the evidence is limited to the particular hyperparameter configuration and datasets studied; please temper the wording or provide statistical support.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The proof of Theorems 1–3 is not just missing details; the inequality direction in the final descent step is reversed and the scalar/vector confusion in Eq. (10) is substantive. Since the convergence guarantee is advertised as a contribution, this needs a genuine rework, not a cosmetic fix. The empirical study is broad and useful, but the lack of error bars and the tuning of M and tau on the headline benchmark should be addressed in revision. I would not reject outright because the proposed method is well motivated and the empirical results are directionally consistent; however, the theoretical claims must be substantially corrected or explicitly weakened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The method is sensible and the empirical results are directionally supportive, but the convergence proof in the appendix doesn't hold up, so the \"guaranteed convergence\" claim should be treated with skepticism.\n\nWhat's new: FedSC combines angular-neighbor prototype selection (Eq. 8-9), discrepancy-weighted consistent prototypes (Eq. 10-12), and an InfoNCE-style prototype contrastive loss plus a prototype regularization loss. The specific combination isn't in FedProto, FedNH, FedRCL, or FedCDA. That's a legitimate incremental contribution. The method is specified clearly enough to reimplement, and the authors release code, which is good.\n\nWhat it does well: The empirical evaluation is broad: label shift, long-tailed, few-shot, across four datasets, with comparisons against recent baselines. The accuracy gains are consistent, around 1-3 points, and the ablations show both RPCL and CPDR contribute. The convergence rates table also shows faster reaching of accuracy thresholds.\n\nSoft spots: The main problem is the convergence proof. The stress-test note is correct: in Eq. (10) of the appendix, the proof treats the scalar similarity s(z_i, r^j) as if it were a vector norm, then bounds a log-sum-exp of similarities by prototype norms without a Lipschitz constant for exp. The normalizer U in Eq. (13) is data-dependent and not covered by the assumptions. More seriously, the step after Eq. (12) replaces the sum of squared gradient norms with E B^2, but Assumption 3 gives an upper bound on each squared norm, not a lower bound. With the negative coefficient in front, you cannot tighten the upper bound that way. Without a lower bound on gradients (or a PL condition), the monotone decrease in Theorem 2 and the rate in Theorem 3 do not follow. So the theory as written doesn't establish what it claims.\n\nTwo more minor concerns: no standard deviations or seed counts, which matters when gains are a couple of points; and key hyperparameters (tau, M) appear to be tuned on the same benchmark used for the main comparison. The sensitivity analysis in Fig. 5 helps, but it doesn't fully rule out overfitting to that setting.\n\nWho this is for: groups working on prototype-based FL will find the method useful as a baseline or an idea to build on. The flawed theory should not prevent using the method, but it does mean the paper's theoretical contribution is not credible in its current form.\n\nRecommendation: This deserves a serious referee. The empirical work is substantial, but the paper needs major revisions: fix or remove the convergence theorems, add seed variability, and clarify hyperparameter selection. If the theory can't be repaired, it should be replaced with a more modest statement about empirical behavior.","headline":"Sensible FL method with a broken convergence proof; the empirical work is substantial and the method is worth trying, but the theory should be fixed or dropped.","tokens_in":22706,"tokens_out":4667,"would_cite":true,"duration_ms":43111,"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":"FedSC shares class prototypes across clients and beats nine baselines under heterogeneous data.","keywords":["federated learning","data heterogeneity","prototype learning","contrastive learning","label shift","long-tailed distribution","few-shot classification","convergence analysis"],"falsifier":"One concrete test: run FedSC under a deliberately constructed extreme label shift, for example each client holds exactly two disjoint classes on CIFAR10, and measure per-class accuracy of the final global model. If tail or absent-class accuracy falls below FedAvg, or if the consistent prototypes $o^j$ drift toward head-class directions, the discrepancy-weighting premise fails. On the theory side, empirically estimate $\\mathbb{E}_\\zeta[\\|\\nabla L(w,\\zeta)-\\nabla L(w)\\|^2]$ for the RPCL term with the normalization factor $U$; if the variance grows without bound as $U$ varies or fails Assumption 2, the convergence rate in Theorem 3 does not apply to the actual objective.","tokens_in":21716,"feed_emoji":"🤝","tokens_out":5814,"duration_ms":55506,"temperature":0.7,"pith_summary":"Federated learning across clients with different label distributions usually produces a global model that drifts because each local model overfits its own skewed data. This paper claims that the fix is to let clients exchange class-level semantic knowledge, not only model weights: each client computes class prototypes, and the server fuses them into relational prototypes by angular similarity and consistent prototypes by discrepancy-weighted aggregation. These prototypes feed two complementary losses, a contrastive loss that pulls samples toward their class's relational prototypes and pushes them away from other classes, and a regularization term that pulls features toward stable consistent prototypes. If correct, FedSC gives a more generalizable global model under label shift, long-tailed imbalance, and few-shot tasks, while converging in fewer rounds than existing federated learning methods.","feed_headline":"Sharing semantic prototypes tames non-IID federated learning","feed_subtitle":"FedSC fuses client class prototypes by angle and discrepancy, beating FedAvg on label shift, long-tail, and few-shot tasks.","key_machinery":"The argument runs on two constructed objects. Relational prototypes $r^j_k$ are client-prototype means merged over the top-$M$ clients whose class-$j$ prototypes are closest in angle to the global prototype $g^j$, captured by an adjacency matrix $A^j$; they supply diverse, class-relevant semantic knowledge. Consistent prototypes $o^j$ are weighted sums of relational prototypes with weights $e_k = \\mathrm{Sigmoid}(a n_k - b d_k)$, where $d_k$ is the distance between the client's local class distribution and a uniform global distribution; they supply a flat, stable convergence target. The two prototype sets carry the learning through loss terms $L_{RPCL}$, a normalized InfoNCE-style contrastive loss with similarity $s(z_i, r^j_k)$ divided by the average feature distance $U$, and $L_{CPDR}$, a squared distance between the sample feature and $o^j$. The convergence theorems then bound the deviation of the combined objective using smoothness, bounded variance, bounded dissimilarity, and Lipschitz continuity assumptions.","core_discovery":"On its own terms, the paper establishes that exploiting intra-client semantic information via prototypes is enough to mitigate the negative effects of data heterogeneity in federated learning. The central mechanism is a local objective $L = L_{CE} + L_{RPCL} + L_{CPDR}$: relational-prototype contrastive learning ($L_{RPCL}$) treats the relational prototypes $R^j$ of class $j$ as virtual teachers, attracting each query embedding and repelling prototypes of other classes; consistent-prototype discrepancy regularization ($L_{CPDR}$) constrains each sample feature to stay close to the corresponding consistent prototype $o^j$. The paper reports that FedSC \"universally outperforms\" baselines across CIFAR10, CIFAR100, TinyImageNet, and FC100 under three heterogeneity settings, requires fewer rounds to reach target accuracy, and is robust in long-tailed and few-shot scenarios. It also proves, under four stated assumptions, a monotone-decreasing objective and a guaranteed convergence rate with an explicit learning-rate window.","pith_inferences":["The prototype-communication design suggests a privacy and usefulness trade-off: sharing only class means and aggregated angular relations may leak less than sharing model weights, but the paper does not quantify this, so an adversarial membership-inference test on the exchanged prototypes would be a natural next step.","Because the consistent-prototype weights favor clients with large sample size and low distributional discrepancy, the aggregation rule could double as a client-selection or fairness signal; the paper explores neither, but its own Eq. 11 makes the connection immediate.","The angular neighbor selection treats class semantics as living on a hypersphere, so a direct extension would be to test whether replacing the top-$M$ hard selection with a soft, temperature-weighted angular kernel preserves the reported gains at larger client counts.","The empirical gains under long-tail and few-shot settings suggest the method may also act as a regularizer in a single-client long-tail problem, but that transfer is not tested in the paper."],"forward_implications":["Under label shift (Dirichlet and NID2 partitions), FedSC improves final accuracy over nine federated learning baselines on all four tested datasets, with the largest reported gains in the hardest heterogeneity regime.","In long-tailed settings with imbalance ratio $\\rho$ up to 200, FedSC holds its advantage, which the paper attributes to discrepancy-weighted consistent prototypes counteracting head-class bias.","In few-shot FC100 benchmarks, FedSC outperforms prior prototype- and contrastive-based methods, indicating that the semantic-level knowledge transfers to low-data regimes.","FedSC reaches target accuracies in substantially fewer rounds than FedAvg, FedProx, FedProto, MOON, FedNTD, FedNH, FedRCL, and FedCDA.","Ablation shows RPCL contributes more than CPDR in most settings, and combining both gives the best average accuracy, supporting the paper's complementarity claim."],"supporting_citations":[{"why":"Defines the FedAvg aggregation loop that FedSC inherits and compares against.","marker":"[33]"},{"why":"Supplies the heterogeneous-FL formulation, local-drift motivation, and the smoothness and bounded-variance assumptions used in the convergence proof.","marker":"[28]"},{"why":"Establishes the class-prototype sharing paradigm in FL that FedSC extends with relational and consistent prototypes.","marker":"[44]"},{"why":"Provides the model-contrastive FL baseline (MOON) and the contrastive-loss-in-FL comparison.","marker":"[26]"},{"why":"Provides the recent relaxed-contrastive FL baseline (FedRCL) that table comparisons must beat.","marker":"[41]"},{"why":"Gives the SimCLR contrastive learning formulation that RPCL adapts to prototype-vs-instance pairs.","marker":"[4]"},{"why":"Defines the InfoNCE objective whose log-sum-exp form is used by the RPCL loss.","marker":"[38]"},{"why":"Defines prototypical networks, the few-shot benchmark backbone used for the FC100 evaluations.","marker":"[42]"}],"fun_headline_variants":["Prototype collaboration defeats label shift in federated learning","Semantic prototypes cure non-IID federated learning","Prototype-wise collaboration tames heterogeneous federated learning","Relational prototypes plus consistency fix non-IID FL","Semantic prototypes tame label shift and long-tail in FL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee depends on assuming that the combined prototype-based loss is smooth, has unbiased stochastic gradients with bounded variance, has bounded gradient dissimilarity, and that feature extraction is Lipschitz continuous; the paper assumes these properties rather than verifying them for the prototype and contrastive terms, and it further assumes a uniform global class distribution when computing the discrepancy weights.","fun_headline_variants_meta":{"raw":{"variants":["Prototype collaboration defeats label shift in federated learning","Semantic prototypes cure non-IID federated learning","Prototype-wise collaboration tames heterogeneous federated learning","Relational prototypes plus consistency fix non-IID FL","Semantic prototypes tame label shift and long-tail in FL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001375,"raw_usage":{"total_tokens":5597,"prompt_tokens":998,"completion_tokens":4599,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":4522}},"tokens_in":614,"tokens_out":4599,"duration_ms":31487,"temperature":1.0,"reasoning_tokens":4522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:36:54.935576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One concrete test: run FedSC under a deliberately constructed extreme label shift, for example each client holds exactly two disjoint classes on CIFAR10, and measure per-class accuracy of the final global model. If tail or absent-class accuracy falls below FedAvg, or if the consistent prototypes $o^j$ drift toward head-class directions, the discrepancy-weighting premise fails. On the theory side, empirically estimate $\\mathbb{E}_\\zeta[\\|\\nabla L(w,\\zeta)-\\nabla L(w)\\|^2]$ for the RPCL term with the normalization factor $U$; if the variance grows without bound as $U$ varies or fails Assumption 2, the convergence rate in Theorem 3 does not apply to the actual objective.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the FedAvg aggregation loop that FedSC inherits and compares against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the heterogeneous-FL formulation, local-drift motivation, and the smoothness and bounded-variance assumptions used in the convergence proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the class-prototype sharing paradigm in FL that FedSC extends with relational and consistent prototypes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the model-contrastive FL baseline (MOON) and the contrastive-loss-in-FL comparison."},{"cited_title":"Nguyen, Chien Duong, Tu Phan, Erman Tjiputra, and Quang D","cited_arxiv_id":null,"evidence_quote":"Provides the recent relaxed-contrastive FL baseline (FedRCL) that table comparisons must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines prototypical networks, the few-shot benchmark backbone used for the FC100 evaluations."}],"review_version":1}