{"id":"eacf56e7-2525-446f-b14c-3b27a803c614","arxiv_id":"2412.00410","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"FedPSD improves personalized federated learning by using previous-epoch model outputs as teachers and a locally prior-weighted softmax loss, reporting higher client and server accuracy on three benchmarks.","lead":"A new federated learning method, FedPSD, combines progressive self-distillation with a logits calibration loss to help each edge device keep both global and personalized knowledge during local training. Tests on MNIST, CIFAR-10, and CIFAR-100 show consistent accuracy gains over existing personalized federated learning baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (7) implements the opposite sign of the balanced objective derived in Eq (6), so the CLL component's theoretical motivation is internally inconsistent; the empirical claim may still hold, but the stated justification does not.","rationale":"The paper's central empirical claim is that FedPSD outperforms baseline PFL methods in Top-1 accuracy on client and server sides across MNIST, CIFAR-10, and CIFAR-100. The most load-bearing weakness in the argument is not the empirical comparison itself, which is broad and plausible, but the theoretical justification of the calibrated logits loss. Section IV-C contains a sign error: Eq (6) requires subtracting ln P(y) from logits to remove class-prior bias, while Eq (7) and the accompanying text use +ln P(y). This is not a matter of interpretation; the argmax of Eq (7) is provably argmax_y(f_y + ln P(y)). The stated equivalence is therefore false, and the entire motivation for the CLL component collapses unless the authors intended the inverse prior, which they did not say. The reader identified exactly this issue, and I agree. The concern is load-bearing because CLL is one of the three named contributions and the subject of a claimed theoretical demonstration; however, the error is localized and easily fixable, and it does not directly invalidate the empirical results. Missing error bars and some baseline omissions are additional weaknesses, but they are less decisive than the internal contradiction in the core derivation. The paper should be revised to correct the sign or reframe what Eq (7) actually accomplishes, with a corresponding update to the theoretical claims. Until then, conditional acceptance is appropriate: the method may work, but the paper's own justification for a key component is not correct as written.","tokens_in":13901,"tokens_out":3853,"duration_ms":39758,"concrete_test":"Algebraic check: substitute Eq (7) into the argmax. Since the denominator sum_{y'} P(y')e^{f_{y'}(x)} is independent of y, argmax_y P_calibrated(y|x) = argmax_y(f_y(x) + ln P(y)). The claimed balanced objective is argmax_y(f_y(x) - ln P(y)). These agree only if P(y) is constant across classes. To test practical impact, rerun the CIFAR-10 S=2 ablation from Table VI with CLL replaced by the correct logit adjustment (subtract ln P(y) from logits, or multiply softmax by 1/P(y)) while keeping all other settings identical. If the corrected variant does not match or exceed the reported 60.31 client accuracy, the CLL benefit is not attributable to the stated balanced objective.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-C derives the balanced objective argmax_y(f_y(x) - ln P(y)) in Eq (6), then states that the local cross-entropy is modified by adding ln P(y) to the logits. Equation (7) defines P_calibrated(y|x) proportional to P(y)e^{f_y(x)}. Because the normalization denominator is positive and independent of y, argmax_y P_calibrated(y|x) = argmax_y(f_y(x) + ln P(y)), which is the opposite of Eq (6). Unless P(y) is intended to denote the inverse prior (which is not stated), the claimed equivalence is false. The paper's own phrase 'guiding the model to prioritize optimization for locally frequent classes' describes the implemented +ln P(y) effect, which contradicts the stated goal of removing class-prior bias. This invalidates the theoretical motivation for one of the three core components (CLL). The Table VI ablation attributes large gains to CLL, but those gains cannot be explained by the balanced-objective argument as written. This is a localized mathematical error: it does not by itself falsify the empirical comparison, since the implemented CLL could still help through reweighting or regularization, but it undermines the paper's central theoretical claim for the calibrated loss.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedPSD, a personalized federated learning method for IIoT edge clients, built from three client-side components: a logits-calibrated cross-entropy loss that rescales logits by local class priors, a progressive self-distillation loss in which the previous local epoch serves as the teacher for the next epoch, and calibrated fusion labels constructed from historical personalized model outputs that guide the first local epoch. The authors motivate the method with an analysis of global and personalized knowledge forgetting, and evaluate it on MNIST, CIFAR-10, and CIFAR-100 under pathological sharding and LDA partitions, reporting consistent Top-1 accuracy improvements on both client and server sides as well as reduced communication rounds to reach FedAvg-level accuracy.","tokens_in":14114,"tokens_out":5341,"duration_ms":57375,"significance":"The empirical study is extensive: eight baselines, three datasets, two non-IID partition families, ablations for each component, and robustness checks across network architectures, local epoch counts, and client participation rates. The experimental protocol is described in enough detail to be reproduced, and the reported gains are large and consistent. If the results hold, FedPSD is a practically useful contribution to personalized federated learning for resource-constrained edge devices. However, the central theoretical justification for the calibrated logits loss is internally inconsistent: the derivation in Eq. (6) and the implementation in Eq. (7) use opposite signs. This means the stated balanced-objective motivation for one of the three core components is invalid as written, and the empirical gains attributed to that component in Table VI currently lack the claimed theoretical explanation.","major_comments":[{"comment":"The derivation in Eq. (6) correctly obtains argmax_y(f_y(x) - ln P(y)) as the balanced objective that removes the class-prior bias. However, Eq. (7) defines P_calibrated(y|x) proportional to P(y)e^{f_y(x)}, whose argmax is argmax_y(f_y(x) + ln P(y)), because the normalization denominator is positive and independent of y. Thus the implemented loss in Eq. (8) has the opposite sign of the derived objective, and the text's statement that adding ln P(y) to the logits achieves the objective of Eq. (6) is false. This invalidates the theoretical motivation for the CLL component; the large CLL gains in Table VI cannot be explained by the balanced-objective argument as written. The authors should either change Eq. (7) to use P(y)^{-1}e^{f_y(x)} (equivalently, subtract ln P(y) from the logits) and rerun the affected experiments, or explicitly reframe CLL as a positive reweighting that prioritizes locally frequent classes and provide a separate justification for that design choice.","section":"IV-C, Eqs. (5)-(8)"},{"comment":"The claimed 'theoretical demonstration' that alternating aggregation and local training causes continual forgetting of global and personalized knowledge is not a proof. The decomposition w_t^k = w_global + w_local,k and the observation that the weighted average of personalized components may be small restate the effect of averaging but do not establish a mechanism by which a specific client's historical personalized knowledge is forgotten, nor do they quantify forgetting. The contribution bullet 'We theoretically demonstrate...' is therefore overstated. Please either provide a formal statement with assumptions and a quantitative bound or temper the claim to an analytical motivation.","section":"III-A.1 and contribution list (Section I)"}],"minor_comments":[{"comment":"The ablation text refers to 'GSD' and 'RELogits' components, but the table columns are labeled RHPK, PSD, and CLL; additionally, the reported Δ values mix incremental gains over the previous configuration with gains compared to the baseline. Please align the terminology and state explicitly what each Δ measures.","section":"VI-A, Table VI"},{"comment":"The caption contains the phrase 'the sharing pathological sharding strategy'; this appears to be a typo for 'the sharding pathological sharding strategy' and should be corrected.","section":"Figure 2 caption"},{"comment":"The results are reported as averages over three runs without standard deviations or per-seed ranges. Given the small number of runs, reporting variance information would substantially strengthen the reliability of the comparisons.","section":"V-D and Tables I-III"},{"comment":"The within-round progressive self-distillation label H^{t,e-1}_k uses the same alpha from Eq. (2), which depends only on the communication round t, not on the local epoch e. Please clarify whether this is intentional, since the teacher here is the previous epoch's output within the same round rather than a historical personalized model.","section":"IV-B-2, Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eqs. (6)-(8) is the main obstacle. It is localized and fixable in principle, but the authors must clarify which expression was actually implemented: if Eq. (7) was used, the method is a frequent-class-prior reweighting rather than the balanced-objective calibration claimed in the paper, and the entire Section IV-C motivation needs to be rewritten; if the intended implementation was the corrected sign, the experiments must be rerun. The empirical benchmark is otherwise competent and the framework is plausible, so I would not reject on this basis alone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plain speaking: this is a useful empirical paper that overreaches in its theory. The proposed FedPSD combines three ingredients — progressive self-distillation where the previous local epoch's outputs teach the next, fusion labels from historical personalized models to seed the first epoch, and a prior-weighted local cross-entropy — and the combination is new. The storage overhead is genuinely low, which matters for the IIoT framing.\n\nThe experiments are the best part. The authors benchmark against eight baselines across MNIST, CIFAR-10, CIFAR-100, under both sharding and LDA, and report both client-side and server-side accuracy. The gains over FedLMD/FedNTD are large and consistent (e.g., CIFAR-10 S=2 client accuracy 60.31 vs 42.42). The ablations attribute gains to each component, though the names shift between text and table (GSD/RELogits vs RHPK/CLL) — a sloppiness that needs cleaning.\n\nThe soft spot is not soft: the theoretical motivation for the calibrated loss has a sign error. Eq. (6) correctly derives argmax(f_y(x) - ln P(y)) as the balanced objective. Eq. (7) then defines P_calibrated ∝ P(y) e^{f_y(x)}, which is argmax(f_y(x) + ln P(y)) — the opposite sign. The text says adding ln P(y) 'guides the model to prioritize locally frequent classes,' which directly contradicts the preceding derivation's goal of removing class-prior bias. The paper's central claim that CLL realizes the balanced objective is false as written. The ablation shows CLL accounts for a large chunk of the gain, so the method may work through some other reweighting effect, but the stated justification collapses.\n\nThe 'theoretical demonstration' in Section III is also not a proof; it is informal reasoning that the aggregation averages out personalized parameters, which is fine as intuition but not a theorem.\n\nThe empirical comparison is missing error bars, and several related methods cited in the related work (FedGKD, pFedSD, DKD-pFed) are not included as baselines. That is worth fixing.\n\nOverall, the empirical core is plausible and the error is localized. I would send this to review with a request to fix the sign error or reframe CLL as a prior-weighted loss rather than a balanced-loss implementation, and to add error bars. The paper is likely to be a useful contribution to PFL if the authors make that repair.","headline":"A storage-light personalized FL recipe with big empirical gains, but the calibration component's theoretical motivation has a sign error that the authors need to fix.","tokens_in":14641,"tokens_out":1962,"would_cite":false,"duration_ms":18363,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A federated learning method counteracts knowledge forgetting by calibrating logits and distilling each client's own past outputs.","keywords":["personalized federated learning","knowledge forgetting","self-distillation","logits calibration","Non-IID data","IIoT edge intelligence","knowledge distillation"],"falsifier":"Evaluate the decision rule implied by Eq (7) on two classes with logits $f_1=0$, $f_2=1$ and priors $P(1)=0.9$, $P(2)=0.1$: the calibrated probability in Eq (7) favors class 1, while the derived balanced rule $\\arg\\max_y(f_y(x) - \\ln P(y))$ favors class 2, so the claimed equivalence is directly checkable. A decisive experiment would train on an imbalanced client with Eq (8) and with the explicit logit-subtraction loss, then compare per-class balanced accuracy; if the calibrated loss does not improve balanced accuracy over standard cross-entropy, the calibration claim is not doing the attributed work.","tokens_in":13671,"feed_emoji":"🧠","tokens_out":7252,"duration_ms":63060,"temperature":0.7,"pith_summary":"Federated learning on non-IID edge data suffers from two kinds of forgetting: local training drifts away from the global model's general knowledge, and global aggregation erases each client's historical personalized knowledge. The paper proposes FedPSD, a client-side method that combines logits calibration with progressive self-distillation to prevent both kinds of forgetting. Each client's previous-round personalized model output becomes a soft label for the first local epoch, each previous local epoch serves as a virtual teacher for the next, and the local cross-entropy loss is reweighted by the class prior. On MNIST, CIFAR-10, and CIFAR-100 under sharding and LDA partitions, the paper reports higher Top-1 accuracy than eight federated learning baselines on both client and server models, plus faster convergence in communication rounds. The motivation is that personalized federated learning can then run on resource-constrained IIoT edge devices without storing an extra teacher model.","feed_headline":"Logits calibration plus self-distillation beats eight FL baselines","feed_subtitle":"Keeps global and historical knowledge intact, with higher Top-1 accuracy in fewer rounds.","key_machinery":"The central mechanism is a pair of calibrated fusion labels and two client-side losses. The calibrated fusion label $H = \\alpha P + (1-\\alpha)Y$ with $\\alpha = t/t_{\\mathrm{total}}$ provides a smoothed teacher for the first local epoch, reviewing historical personalized knowledge. The progressive self-distillation loss $L_{KD} = \\mathrm{KL}(H^{t,e-1}_k \\| P^{t,e}_k)$ makes each epoch's output the teacher for the next, so global knowledge inherited from the server is passed along gradually. The calibrated cross-entropy loss uses $P_{\\mathrm{calibrated}}(y|x) = P(y)e^{f_y(x)} / \\sum_{y'} P(y')e^{f_{y'}(x)}$, and the total local loss is $L_{\\mathrm{local}} = L_{CE} + L_{KD}$. The design keeps all modifications on the client side and requires no retained global teacher model.","core_discovery":"FedPSD claims that the alternating pattern of local training and global aggregation in federated learning produces continual forgetting of both global generalization knowledge and historical personalized knowledge, and that both can be countered with changes confined to the client. For global knowledge, the local cross-entropy loss is replaced by a calibrated loss that reweights logits by the local class prior, and the model's output from the previous local epoch distills into the next epoch, so knowledge from the global model is inherited gradually. For personalized knowledge, the output of the client's personalized model from the previous communication round is fused with the ground-truth label into a calibrated fusion label that guides the initial epoch of local training. On MNIST, CIFAR-10, and CIFAR-100 under pathological sharding and LDA partitions, the paper reports consistent Top-1 accuracy gains over FedAvg, FedProx, FedCurv, FedNova, SCAFFOLD, MOON, FedNTD, and FedLMD, including CIFAR-10 S=2 client accuracy of 60.31 versus 42.42 for FedLMD and 42.66 for FedNTD, and faster convergence to FedAvg's final accuracy in fewer communication rounds.","pith_inferences":["The derivation in Eqs (5)–(6) implies the balanced objective requires subtracting $\\ln P(y)$ from logits, while Eq (7) adds it; if the intended calibration is the derived one, the method's theoretical justification would need a sign reversal, and the empirical gains might come from the progressive distillation rather than the calibrated loss.","A straightforward testable extension is to replace Eq (7) with the explicit logit adjustment $f_y(x) - \\ln P(y)$ and compare balanced accuracy on the same partitions; this would isolate whether the calibration mechanism or the distillation schedule drives the reported gains.","The same calibrated-fusion idea could be applied to centralized class-imbalanced training, where past-epoch predictions could serve as smoothed teachers without any federated aggregation.","Storing one soft-label vector per client per round is cheap, but across many rounds memory could grow; averaging or compressing historical outputs is a natural follow-up that the paper does not explore."],"forward_implications":["Edge clients can retain both global and personalized knowledge using only their own historical outputs, avoiding the storage cost of a separate teacher model.","Fewer communication rounds are needed to reach FedAvg's final accuracy; on CIFAR-10 sharding S=2, the client-side model reaches it in 10 rounds instead of 200.","The method's benefit grows as data heterogeneity increases, making it suited to IIoT deployments with skewed and missing classes.","Consistent gains across SimpleCNN, ResNet10, VGG11, and across 1–20 local epochs suggest the approach transfers across architectures and training budgets."],"supporting_citations":[{"why":"FedNTD, the not-true distillation baseline that preserves global knowledge; the paper compares against it and reports higher accuracy.","marker":"[9]"},{"why":"FedLMD, the label-masking distillation baseline for preserving minority label knowledge; the paper compares against it and reports higher accuracy.","marker":"[10]"},{"why":"FedSSD, a selective self-distillation baseline that motivates the paper's self-distillation design.","marker":"[11]"},{"why":"FedGKD, a global knowledge distillation baseline that uses cached historical global models as teachers.","marker":"[12]"},{"why":"A prior method that preserves both global and personalized knowledge via an adaptive knowledge matrix, the closest antecedent to FedPSD's dual goal.","marker":"[8]"},{"why":"pFedSD, which distills previous personalized models into current local models, used as a baseline for personalized knowledge forgetting.","marker":"[15]"},{"why":"DKD-pFed, a decoupled knowledge distillation approach for personalized federated learning, used as a baseline for personalized knowledge forgetting.","marker":"[16]"},{"why":"The knowledge distillation method that supplies the teacher-student loss formulation used by FedPSD.","marker":"[13]"},{"why":"FedAvg, the foundational federated learning algorithm and the primary baseline for the communication-efficiency comparisons.","marker":"[25]"}],"fun_headline_variants":["Calibrated logits plus progressive self-distillation top eight FL baselines","FedPSD: calibration and self-distillation curb forgetting in FL","Stop forgetting in federated learning: a calibrated self-distillation trick","FedPSD: logits calibration plus progressive self-distillation beats baselines","Progressive self-distillation with logits calibration preserves FL knowledge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the calibrated loss of Eq (7), which multiplies the softmax by the class prior $P(y)$, realizes the balanced objective derived in Eq (6), $\\arg\\max_y(f_y(x) - \\ln P(y))$; the two expressions have opposite signs, so if the premise gives way the method's theoretical motivation collapses.","fun_headline_variants_meta":{"raw":{"variants":["Calibrated logits plus progressive self-distillation top eight FL baselines","FedPSD: calibration and self-distillation curb forgetting in FL","Stop forgetting in federated learning: a calibrated self-distillation trick","FedPSD: logits calibration plus progressive self-distillation beats baselines","Progressive self-distillation with logits calibration preserves FL knowledge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000707,"raw_usage":{"total_tokens":3214,"prompt_tokens":1005,"completion_tokens":2209,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":2115}},"tokens_in":621,"tokens_out":2209,"duration_ms":14844,"temperature":1.0,"reasoning_tokens":2115,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:24:55.370416+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the decision rule implied by Eq (7) on two classes with logits $f_1=0$, $f_2=1$ and priors $P(1)=0.9$, $P(2)=0.1$: the calibrated probability in Eq (7) favors class 1, while the derived balanced rule $\\arg\\max_y(f_y(x) - \\ln P(y))$ favors class 2, so the claimed equivalence is directly checkable. A decisive experiment would train on an imbalanced client with Eq (8) and with the explicit logit-subtraction loss, then compare per-class balanced accuracy; if the calibrated loss does not improve balanced accuracy over standard cross-entropy, the calibration claim is not doing the attributed work.","supporting_citations":[{"cited_title":"Federated learning with label-masking distillation,","cited_arxiv_id":null,"evidence_quote":"FedLMD, the label-masking distillation baseline for preserving minority label knowledge; the paper compares against it and reports higher accuracy."},{"cited_title":"Learning critically: Selective self-distillation in federated learning on non-iid data,","cited_arxiv_id":null,"evidence_quote":"FedSSD, a selective self-distillation baseline that motivates the paper's self-distillation design."},{"cited_title":"F ed gkd: Towards heterogeneous federated learning via global knowledge distillation,","cited_arxiv_id":null,"evidence_quote":"FedGKD, a global knowledge distillation baseline that uses cached historical global models as teachers."},{"cited_title":"Rethinking personalized federated learning from knowledge perspective,","cited_arxiv_id":null,"evidence_quote":"A prior method that preserves both global and personalized knowledge via an adaptive knowledge matrix, the closest antecedent to FedPSD's dual goal."},{"cited_title":"Dkd-pfed: A novel framework for personalized federated learning via decoupling knowledge distillation and feature decorrelation,","cited_arxiv_id":null,"evidence_quote":"DKD-pFed, a decoupled knowledge distillation approach for personalized federated learning, used as a baseline for personalized knowledge forgetting."}],"review_version":1}