{"id":"f063c3f2-5bed-4c59-84df-6b9b2355094b","arxiv_id":"2412.07062","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FLAYER combines performance-guided layer-wise initialization, adaptive layer-specific learning rates, and selective parameter masking to improve personalized federated learning accuracy and reduce training cost.","lead":"This paper introduces FLAYER, a personalized federated learning method that adjusts how much each neural network layer learns from local versus global data. It reports accuracy gains over six existing methods on image and text benchmarks, though the described training loop contains an unexplained aggregation step.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's server update averages zero-masked local parameters: any coordinate not selected by at least one sampled client becomes zero (or is shrunk) in the global model, so the reported results must rely on an unspecified extra aggregation rule.","rationale":"The reader's weakest assumption is exactly the point on which the paper's central claim rests. I checked the manuscript: Eq. (9) defines the uploaded value as the trained parameter vector multiplied by a binary mask whose zero entries are determined by Eq. (8); Algorithm 1 line 12 computes the next global model as a weighted sum of these masked vectors. Thus any coordinate absent from all sampled clients' selected sets is forced to zero, and any coordinate selected by only some clients is shrunk by the averaging weight of the non-selecting clients. This is a concrete mathematical consequence of the text, not a matter of tuning or consensus. The reported high accuracy cannot follow from this rule as written, so the paper needs either a corrected aggregation rule or a clear statement of an additional operation, such as carry-over of previous global values or masked-delta updates. I also verified that the 5.40% headline is a defensible but broad average across all six baselines and seven datasets; it is not itself the main problem. Because the algorithm-to-implementation gap is load-bearing, the reader's REJECT verdict should stand.","tokens_in":12573,"tokens_out":7737,"duration_ms":82467,"concrete_test":"Ask for the released code and inspect the server-side step. If code is unavailable, implement Algorithm 1 literally on CIFAR-100/ResNet-18 with Dir(0.1); after each round, record the sparsity and L2 norm of Θ_t_g. If the server averages masked θ_t_k exactly as line 12 states, unselected coordinates will be zero or heavily shrunk and accuracy should collapse well below the reported 60.68%. If the code instead preserves previous global values for unselected coordinates or aggregates masked deltas, the text of Eq. (9) and Algorithm 1 line 12 must be corrected and all reported numbers regenerated under the stated rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central accuracy claim depends on an aggregation rule that the paper never states. In Eq. (6)-(9), the client uploads θ_t_k = \\hatθ_t_k ⊙ M_t_k, where M_t_k is 0 for every parameter not in the top-UP_i set by weight fluctuation. Algorithm 1 line 12 then sets Θ_t_g = Σ_k (n_k/Σ_j n_j) Θ_t_k. For any coordinate that no sampled client selects, every uploaded Θ_t_k is zero, so the next global model has that coordinate exactly zero. Even for a coordinate selected by a strict subset of clients, the average includes zeros from the non-selecting clients, shrinking its magnitude roughly proportionally to the selection frequency. Neither behavior is described or analyzed in the sections 'Layer-wise Sparse Binary Masking' or 'Overall Performance'. The reported 60.68% on CIFAR-100 therefore cannot be produced by the algorithm as written; the implementation must be doing something else, such as averaging only selected coordinates while carrying over previous global values, or uploading masked deltas and adding them to Θ_g. Without that rule being specified and analyzed, the main empirical claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FLAYER, a personalized federated learning method with three layer-wise components: (i) performance-guided local aggregation that initializes head layers from a weighted mix of local and global parameters, (ii) an adaptive per-layer learning rate based on layer position and gradient norm, and (iii) a layer-wise sparse binary masking scheme that uploads only selected parameters to the server. The authors claim that FLAYER improves inference accuracy by 5.40% on average (up to 14.29%) over six pFL baselines while reducing computation cost, based on experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, and AG News.","tokens_in":12852,"tokens_out":8800,"duration_ms":87084,"significance":"If the method and results are reproducible, FLAYER would be a practically useful pFL technique: it combines a clear motivation (different layers play different roles) with three concrete mechanisms, and the evaluation is broad, covering four datasets, seven baselines, heterogeneity, scalability, applicability, and an ablation study. The reported computation-cost reductions are also practically valuable. However, the paper currently does not specify a load-bearing part of the server aggregation rule, so the experimental claims cannot be verified from the text as written. The paper also ships no code, which raises the bar for precision in the algorithmic description.","major_comments":[{"comment":"Algorithm 1 line 12 sets the next global model to a weighted average of the uploaded masked parameters Theta_t_k, where Theta_t_k = hat_Theta_t_k ⊙ M_t_k by Eq. (9). For any coordinate that no sampled client selects, every uploaded Theta_t_k is zero, so that coordinate of the global model becomes exactly zero. For a coordinate selected by only a subset of clients, the average is shrunk by the zero contributions of the non-selecting clients. The paper never states a mechanism that prevents this behavior, such as carrying over the previous global values for unselected coordinates, averaging only selected coordinates, or uploading masked deltas and adding them to the global model. This is the central aggregation rule of the method, and the reported accuracies in Table 2 cannot be produced by the algorithm as written unless an additional rule is assumed. The authors must specify the exact aggregation rule, analyze its effect on the global model, and confirm that the reported experiments use that rule.","section":"Layer-wise Sparse Binary Masking / Algorithm 1"},{"comment":"The computation-cost comparison is based on the number of iterations and total time \"until convergence,\" but the paper does not define the convergence criterion. Without a stopping rule (for example, a threshold on accuracy change or validation loss, or a fixed number of rounds), the iteration counts in Table 3 are not reproducible and the claimed average reduction of 58.9% in training time is not well grounded. Please state the stopping criterion used for each method and dataset.","section":"Computation cost / Table 3"},{"comment":"Eq. (2) uses aggregation weights A_{k,l} and A_{k,g} with A_{k,l}+A_{k,g}=1, but the paper only says that \"the local model inference accuracy ... sets the local weight A_{k,l}.\" This is not a precise definition: it is unclear whether A_{k,l} equals raw accuracy, a normalized or clipped accuracy, or some other mapping, and how accuracy is evaluated (on the local training set or a validation split). Since this performance-guided aggregation is the first contribution of the paper, the mapping must be specified explicitly so that the method is reproducible.","section":"Performance-guided Layer-wise Local Aggregation / Eq. (2)"}],"minor_comments":[{"comment":"The abstract's \"5.40% average improvement\" is supported by Table 2 only if it is interpreted as an average absolute percentage-point difference across all 42 baseline-task pairs; it is not a relative improvement. The text should state this explicitly, since the current wording is ambiguous and could mislead readers.","section":"Abstract / Overall Performance"},{"comment":"Table 5 includes GPFL and FedCP as compared methods, but these are not described in Section 2 or in the related-work discussion, and FedCP has no citation. Please add the missing references and briefly describe their experimental settings.","section":"Table 5"},{"comment":"The adaptive learning rate formula grows without bound as the gradient norm approaches zero. The paper does not discuss clipping or a lower bound on the gradient norm; a brief statement on numerical stability would be helpful.","section":"Eq. (4)"},{"comment":"The head size s is tuned per architecture (s=1 for CNN, s=2 for ResNet-18), but the sensitivity analysis is only reported for CIFAR-100. Please state how s was selected for the other datasets and whether the results are stable with respect to this choice.","section":"Evaluation on Personalization Layers / Table 4"},{"comment":"In the CIFAR-10 CNN row for FedAMP, the table reports \"#517\" with a stray marker; this appears to be a typesetting error that should be corrected.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The missing aggregation rule in Algorithm 1 is the critical issue. The authors appear to have a working implementation, so the problem is likely fixable by stating and analyzing the actual update rule rather than an inherent flaw in the idea. I would not reject solely because the method is empirical; the broad evaluation is a strength. However, if the aggregation rule cannot be clarified or if the reported experiments used a different rule than the one eventually specified, the central accuracy claim would be unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere's the short version: FLAYER is a sensible engineering recipe for personalized FL, but the paper as written has a load-bearing inconsistency in the server-side aggregation that makes the reported gains impossible to reproduce from the equations alone. The abstract's \"5.40% average improvement\" is actually defensible if read as percentage points against Table 2—the average absolute gain across the seven settings is about 5.4 points—so I wouldn't lean on that as the main critique.\n\nWhat's genuinely new: the combination of three mechanisms—performance-weighted head aggregation, per-layer adaptive learning rates, and selective parameter masking for upload. Each piece echoes existing work (FedPer/FedRep, Singh et al., FedMask), but the integration is coherent and the evaluation is broad: seven settings, two architectures, CV and NLP, plus scaling to 100 clients and an ablation showing the learning-rate component is doing most of the work. The CKA analysis is a nice attempt to explain why the method lands between full personalization and full sharing.\n\nThe soft spot is real and it's central. Eq. (9) defines the uploaded parameters as the locally trained parameters multiplied by a binary mask, and Algorithm 1 line 12 averages those masked parameters on the server. Any coordinate that no sampled client selects becomes exactly zero in the next global model, and coordinates selected by only a subset get shrunk by the zero contributions from the rest. The text never explains how the global model avoids collapsing to zero. The reported results therefore cannot be produced by the algorithm as written; the implementation must be using a different rule, such as averaging only selected coordinates while carrying over previous global values, or uploading masked deltas. Until that rule is specified and analyzed, the main empirical claim is unsupported.\n\nOther, smaller issues: the head size s is tuned per dataset (Table 4) without a sensitivity discussion, the adaptive LR formula has no convergence guarantee, and FedMask is cited but not included as a baseline. These are minor next to the aggregation problem.\n\nWho should read this: practitioners working on communication-efficient pFL who care about accuracy and wall-clock time. The idea is worth engaging with, and the experimental setup is serious.\n\nMy recommendation: send it to review, but with a clear request to fix the aggregation description and re-run the experiments. If the authors can reconcile the text with the implementation, this is a credible engineering contribution. If not, reject.","headline":"Good engineering recipe for personalized FL, but the server aggregation rule as written would zero out the global model; the paper needs a major revision before its central accuracy claim can be trusted.","tokens_in":13326,"tokens_out":6064,"would_cite":false,"duration_ms":58057,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FLAYER treats each neural network layer differently to lift personalized federated learning accuracy by 5.4%.","keywords":["personalized federated learning","layer-wise optimization","adaptive learning rate","parameter masking","non-IID data","federated learning efficiency","model personalization","deep neural networks"],"falsifier":"Run the published Algorithm 1 verbatim on CIFAR-100 with ResNet-18 and 20 clients; if masked-out coordinates enter the server average as zeros, the global model loses those parameters entirely, so the accuracy claim collapses. The alternative—keeping previous global values for masked coordinates—is a rule the paper does not state, and the difference between the two variants is measurable in one training run.","tokens_in":12401,"feed_emoji":"📈","tokens_out":10134,"duration_ms":95449,"temperature":0.7,"pith_summary":"FLAYER is a personalized federated learning method built on one idea: a local model's layers should not be treated uniformly, because shallow layers carry shared features and deep layers carry client-specific ones. The paper argues that adapting each stage of local training to this split—blending global and local weights when initializing head layers, scaling each layer's learning rate by depth and gradient, and uploading only the most changed parameters per layer—prevents over-personalization and preserves global knowledge. The reported payoff is concrete: across image and text benchmarks with non-IID data, FLAYER improves inference accuracy by 5.40% on average over six prior personalized federated learning methods, with gains up to 14.29%, while reaching convergence in fewer rounds. If correct, this is a low-overhead recipe for personalizing federated models under data heterogeneity.","feed_headline":"Layer-wise learning lifts personalized federated accuracy 5.4%","feed_subtitle":"Adaptive aggregation, layer-wise learning rates, and selective uploads cut training time even as accuracy improves.","key_machinery":"The load-bearing machinery is a three-stage layer-wise controller applied inside each client. First, head layers (the last $s$ layers) are initialized as $\\tilde{\\theta}_k^{t} = [\\theta_g^{(1:L-s,t-1)},\\ A_{k,l}^{t-1}\\odot\\theta_k^{(L-s+1:L,t-1)} + A_{k,g}^{t-1}\\odot\\theta_g^{(L-s+1:L,t-1)}]$, with $A_{k,l}+A_{k,g}=1$ and $A_{k,l}$ set by the local model's recent accuracy (Eq. 2). Second, each layer $i$ trains with learning rate $\\eta^{(i,t)} = \\eta\\left(1+\\log\\left(1+1/\\|g^{(i,t)}\\|_2\\right)\\cdot i/L\\right)$ (Eq. 4), so deeper layers and smaller gradients take larger steps. Third, each client uploads $\\hat{\\theta}_k^t \\odot M_k^t$, where the binary mask keeps the top $UP_i = \\min(\\max(i/L,0.1),1)$ fraction of weights by absolute change per layer (Eqs. 5-9). This three-step design is what carries the argument: it decides, at every stage, how much global versus local information enters each layer.","core_discovery":"The central claim is that the conflict between personalization and shared global knowledge in federated learning can be managed layer by layer, cheaply, at all three points where local and global information meet. FLAYER's performance-guided aggregation sets the head-layer initialization from the client's recent local accuracy; its adaptive layer-wise learning rate lets deeper, more client-specific layers take larger steps while shallow layers stay close to the global representation; and its layer-wise masking uploads a depth-dependent fraction of each layer's most-changed parameters, so server averaging retains essential base features. On CIFAR-10, CIFAR-100, Tiny-ImageNet, and AG News under Dirichlet(0.1) heterogeneity, the paper reports FLAYER as the most accurate of seven compared methods, averaging a 5.40% inference-accuracy improvement over the six pFL baselines and up to 14.29% on individual tasks, while reducing total training time to convergence by 58.9% on average relative to FedAvg.","pith_inferences":["If the layer-position monotonicity in Eq. (5) is what drives the gain, then models with skip connections or attention heads may need non-monotone upload fractions, because residual paths make some shallow parameters load-bearing in later layers; this is a testable variant.","The accuracy-based aggregation weight in Eq. (2) creates a feedback loop: a client that reports high local accuracy receives more of its own weights back, so in adversarial deployment the rule would need a trustworthy accuracy estimate.","A natural cheap extension is replacing the gradient-norm term in Eq. (4) with a depth-only schedule, since gradient norms and layer depth are correlated; if accuracy holds, the per-layer norm computation can be skipped on resource-limited devices."],"forward_implications":["If the central claim holds, personalized federated learning need not choose between global generalization and local specialization: the same layer-wise rules can be applied to any client model.","Communication per round falls for early layers because each layer uploads only the top $UP_i$ fraction of parameters by absolute change, with $UP_i$ as low as 0.1 for shallow layers.","Because FLAYER is a wrapper rather than a new network architecture, the paper reports that applying it to other pFL methods improves their accuracy by 4.90 to 11.97 percentage points, meaning the layer-wise controls transfer across methods.","The reported convergence speedup (e.g., 78 versus 152 iterations on CNN/CIFAR-10 against FedALA) implies that adaptive layer-wise learning rates are the main driver of faster training, with masking mattering more in deeper residual networks."],"supporting_citations":[{"why":"Element-wise FedALA baseline that FLAYER extends; supplies the strongest accuracy comparison and the base learning-rate setup.","marker":"Zhang et al. 2023b"},{"why":"FedPer baseline that introduces the base/head layer split FLAYER builds on.","marker":"Arivazhagan et al. 2019"},{"why":"FedRep baseline that trains base and head layers separately, the layer-wise precedent FLAYER compares against.","marker":"Collins et al. 2021"},{"why":"APPLE model-wise baseline whose high per-iteration cost motivates FLAYER's cheaper layer-wise aggregation.","marker":"Luo and Wu 2022"},{"why":"Ditto baseline in the six-method comparison that FLAYER claims to outperform.","marker":"Li et al. 2021b"},{"why":"FedAMP baseline in the six-method comparison that FLAYER claims to outperform.","marker":"Huang et al. 2021"},{"why":"Source of the layer-specific learning-rate idea that FLAYER adapts to personalized federated learning.","marker":"Singh et al. 2015"},{"why":"Evidence that shallow layers learn general features and deep layers learn task-specific features, the premise for treating base and head layers differently.","marker":"Yosinski et al. 2014"},{"why":"Supplies the CKA similarity metric used to measure layer-wise generalization versus personalization across clients.","marker":"Kornblith et al. 2019"}],"fun_headline_variants":["Layer-wise federated learning lifts accuracy 5.4%","Adaptive layers cut federated training time 58.9%","FLAYER: layer-wise personalization boosts accuracy 5.4%","Personalized federated learning improved via adaptive layers +5.4%","Layer-wise approach raises federated accuracy 5.4% and speeds training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy depends on an unstated server-side rule for parameters that no sampled client uploads; Algorithm 1 as written would set those coordinates to zero, and the paper never says how the global model keeps them intact.","fun_headline_variants_meta":{"raw":{"variants":["Layer-wise federated learning lifts accuracy 5.4%","Adaptive layers cut federated training time 58.9%","FLAYER: layer-wise personalization boosts accuracy 5.4%","Personalized federated learning improved via adaptive layers +5.4%","Layer-wise approach raises federated accuracy 5.4% and speeds training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1367,"prompt_tokens":989,"completion_tokens":378,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":283}},"tokens_in":605,"tokens_out":378,"duration_ms":3821,"temperature":1.0,"reasoning_tokens":283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:09:57.698088+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the published Algorithm 1 verbatim on CIFAR-100 with ResNet-18 and 20 clients; if masked-out coordinates enter the server average as zeros, the global model loses those parameters entirely, so the accuracy claim collapses. The alternative—keeping previous global values for masked coordinates—is a rule the paper does not state, and the difference between the two variants is measurable in one training run.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the layer-specific learning-rate idea that FLAYER adapts to personalized federated learning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Evidence that shallow layers learn general features and deep layers learn task-specific features, the premise for treating base and head layers differently."}],"review_version":1}