{"id":"ffaeb513-a9b0-418a-ac98-fd9a160369fa","arxiv_id":"2506.10102","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SFMTL-Graph builds a dynamic client similarity graph, partitions it with Louvain community detection, and restricts federated model aggregation to within communities to personalize learning while cutting communication.","lead":"This paper proposes a federated learning method that groups similar clients into communities and lets them share only compact feature summaries and lightweight classification heads, lowering communication costs. A reader interested in personalized machine learning over distributed data might care because the method claims to improve accuracy, fairness, and efficiency over prior federated baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section V-C's claim of less than 10^8 cumulative transmitted bits is contradicted by the paper's own stated model sizes and round count.","rationale":"The reader's conditional verdict already captures the need for corrected and reproducible numbers, and my concern reinforces that verdict rather than changing it. I identify a different weakest point than the reader did: the reader focused on whether the similarity metric correctly ranks task relatedness, which is a plausible but hard-to-calibrate empirical assumption. The communication-cost claim is more decisively testable because it can be checked against the paper's own stated dimensions. Under those dimensions, the headline '<10^8 bits' appears to be off by an order of magnitude, which undermines one of the two central promises of the paper. The accuracy comparison is also troubled by the Section V-B versus Table II discrepancy (text says FedU is about 9% less accurate; Table II reports 0.92 vs 0.91), but that could be a metric or typo issue and is less cleanly falsifiable from the text alone. The paper does provide a public code link and a reproducible experimental protocol, which is genuine supporting evidence; a straightforward logging run would settle the communication question. Therefore the appropriate outcome remains conditional acceptance pending a corrected, reproducible communication-cost analysis and clarification of the accuracy numbers, not outright rejection.","tokens_in":16521,"tokens_out":6865,"duration_ms":84153,"concrete_test":"Run the released repository's Heterogeneous CIFAR-10 configuration for 200 rounds, instrumenting the code to log the exact number of bits transmitted per client per round (classification head plus feature anchor, 32-bit floats) and the size of the sampled client subset S(t). Compute the cumulative transmitted bits across all clients and rounds. If the logged value is approximately 1.18 × 10^9 or larger while Section V-C reports less than 10^8, the communication-efficiency claim is false as stated; if the value is below 10^8, the code must be using an unreported configuration such as quantization, sparse participation, or not transmitting the head every round.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is in the communication-efficiency result, which is a headline contribution of the method. Section V-C states that SFMTL-Graph transmits less than 10^8 bits cumulatively over 200 rounds, and that this is even less than one baseline communication round. But using the paper's own stated quantities (dh = 512, Ck = 2 for CIFAR-10, head size 512 × 10, 32-bit floats), each client transmits a head of 5120 floats plus an anchor of 1024 floats per round, i.e., 6144 floats = 196,608 bits per client per round. With 30 clients and 200 rounds, this gives 30 × 200 × 196,608 = 1.18 × 10^9 bits, more than an order of magnitude above the claimed 10^8. To stay below 10^8 with this per-client payload, fewer than 3 clients would have to transmit per round on average, which is not stated and is implausible for a 30-client experiment. The related claim that cumulative SFMTL-Graph traffic is less than a single baseline round also depends on unreported model-size and client-participation assumptions. This is not a matter of benchmark disagreement; it is an internal arithmetic inconsistency in a core stated advantage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SFMTL-Graph, a federated multi-task learning method in which clients share only a compact feature anchor and a lightweight classification head with the server. The server builds a dynamic similarity graph from cosine similarities between anchors and between head responses on anchors, partitions clients into communities using the Louvain modularity heuristic, and performs per-community aggregation and Laplacian-style head regularization. The authors report experiments on two heterogeneous image classification benchmarks (CIFAR-10 with 30 clients and Rotated & Masked MNIST with 40 clients), claiming superior personalized accuracy, lower communication cost, and better fairness than FedAvg, pFedMe, and FedU. They also provide a qualitative analysis of the dynamic graph and community structure.","tokens_in":16736,"tokens_out":2934,"duration_ms":38036,"significance":"If the claims were fully supported, the paper would make a useful contribution to personalized federated learning by combining communication-efficient anchors with community-restricted collaboration. The idea of using functional similarity of classification heads on shared feature anchors is sensible, and the paper ships a public code repository, which aids reproducibility. However, the central communication-efficiency claim is internally inconsistent with the paper's own stated model sizes, and the client-side anchor update appears not to be computed as described. These issues affect the two headline contributions: communication efficiency and the dynamic similarity graph. The work is therefore not yet ready for publication in its current form, but the core approach is defensible and could be acceptably revised.","major_comments":[{"comment":"The claim that SFMTL-Graph transmits 'less than 10^8 bits cumulatively' over 200 rounds is contradicted by the paper's own stated quantities. With dh = 512, Ck = 2, head size 512×10, and 32-bit floats, each client transmits 5120 + 1024 = 6144 floats = 196,608 bits per round. For 30 clients and 200 rounds this is 30 × 200 × 196,608 ≈ 1.18×10^9 bits, more than an order of magnitude above the claimed 10^8. The related statement that cumulative traffic is less than one baseline round also lacks the baseline model sizes needed to verify it. This is an internal arithmetic inconsistency in a headline contribution; please correct the accounting or report the exact participation counts and payload sizes used to produce Figures 2c and 3c.","section":"Section V-C"},{"comment":"The feature anchor sent to the server is not actually 'newly computed' from local features. In Algorithm 1, line 11 initializes the local anchor h_k^(t) to the received community anchor h_cj, and lines 14–17 update only the model parameters w via the local loss; the anchor h is never updated during local training. Therefore the quantity sent in line 18, h_{k,R}, is just the community anchor received at the start of the round, not a data-dependent summary of the client's learned features. This breaks the definition in Eq. (5), where g_k(θ_k) uses class-wise anchors as targets, and it makes the representation similarity in Eq. (10) largely a measure of prior community identity rather than current feature alignment. Please specify how h is recomputed locally, or revise the algorithm and the similarity construction accordingly.","section":"Algorithm 1 and Eq. (5)"},{"comment":"The text states that FedU, the closest baseline, is 'around 9% less in accuracy' than SFMTL-Graph on Heterogeneous CIFAR-10. Table II reports mean accuracies of 0.92 ± 0.06 for SFMTL-Graph and 0.91 ± 0.08 for FedU, a difference of 0.01, i.e., about 1% relative, and well within one standard deviation. The claim as written is not supported by the reported numbers. Please either report the accuracy gap as a percentage in a way that is consistent with Table II, or provide per-client accuracy distributions and a significance test that justifies the 'significantly superior' wording.","section":"Section V-B and Table II"},{"comment":"The similarity weights in Eq. (7) combine cosine similarity of head logits on the two clients' anchors and cosine similarity of the anchors themselves, but the paper does not calibrate this proxy against ground-truth task relatedness. The qualitative t-SNE analysis in Section V-D is suggestive, but it does not quantify whether the resulting communities consistently group clients with the same or overlapping label sets, nor whether misgrouped clients suffer negative transfer. Please add a quantitative evaluation of the similarity metric (e.g., correlation with label-overlap ground truth, or an ablation with a known-correct similarity oracle) to substantiate the load-bearing premise that the graph correctly orders task relatedness.","section":"Eqs. (7)–(10) and Section V-D"}],"minor_comments":[{"comment":"The indicator function I(C_k^(t), C_ℓ^(t)) is used before the community partition has been formally defined; please define the community assignment function before Eq. (11) or rewrite the indicator in terms of the partition C^(t).","section":"Eq. (11)"},{"comment":"In the Parameters line, the learning rate η appears twice (once as learning rate and once inside the server-side update τ = η × R). Please use distinct symbols for the local learning rate and the server learning rate.","section":"Algorithm 1"},{"comment":"The description of pFedMe's local rounds is confusing: the text says 'uses 20 local rounds' and then 'we make 30 inner rounds for pFedMe'. Please clarify the number of local updates and inner-loop iterations used for pFedMe.","section":"Section V-A"},{"comment":"Several typos and grammatical errors appear in Sections V-D and the figure captions (e.g., 'reprentation', 'similairty', 'simmilarity'), and the t-SNE plots in Figure 4 would benefit from higher resolution and a legend that distinguishes client classes.","section":"Overall"}],"recommendation":"major_revision","confidential_remarks":"The communication-cost arithmetic inconsistency is the most serious issue because it is a headline claim and is easy to verify from the paper's own numbers. The missing client-side anchor update is equally important methodologically: without a local recomputation of h, the similarity graph does not reflect the learned representations as claimed. Both issues are fixable within the scope of a revision, so I recommend major revision rather than rejection. I would also ask the authors to clarify whether the reported communication figures include only uplink client-to-server traffic or both directions, since this affects the comparison with baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a reasonable incremental contribution to personalized FL, but the headline communication-efficiency claim doesn't survive contact with the paper's own model sizes. The stress-test arithmetic is right: with dh=512, Ck=2, head 512×10, 32-bit floats, each client sends 6,144 floats per round (196,608 bits). Over 30 clients and 200 rounds that's 1.18×10^9 bits, an order of magnitude above the claimed 10^8. Even if only a fraction of clients participate each round, the curve in Fig. 2c looks inconsistent with the stated setup. This is not a minor typo; the paper's second headline advantage is built on it.\n\nWhat is genuinely new is the composition: feature anchors plus classification heads as a similarity signal, Louvain-based community partitioning to restrict Laplacian head regularization to the discovered communities. That's a sensible combination, and the authors do give credit to the component ideas (FedFA, FedU, clustered FL). The dynamic graph analysis (t-SNE and community tables) lends some support to the claim that the similarity metric tracks label overlap, which is more than many graph-based FL papers do. Code is available.\n\nThe soft spots beyond the communication issue: Section V-B says FedU is 'around 9% less accurate' while Table II reports 0.91 vs 0.92 — that's about 1%. The closest baseline to their own method, FedFA, is cited but not compared. The client-side anchor update is underspecified: Algorithm 1 uses h_{k,R} in Eq. (5) but never shows how h is updated locally. Lambda is not reported. And the similarity graph is built from representations that are themselves regularized inside the same communities, so the 'ground truth' in Fig. 4 is partly a product of the algorithm. None of these are fatal to the central idea, but they need to be fixed.\n\nI'd send this to peer review — the idea is worth scrutinizing — but the authors should not be allowed to keep the current communication numbers.","headline":"A reasonable incremental composition of FedFA-style anchors, FedU-style Laplacian regularization, and Louvain clustering, but the headline communication-efficiency claim is internally inconsistent with the paper's own model sizes.","tokens_in":17272,"tokens_out":3110,"would_cite":false,"duration_ms":36021,"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 multi-task method restricts collaboration to similar client communities and beats FedAvg, pFedMe, and FedU on heterogeneous benchmarks.","keywords":["federated multi-task learning","personalized federated learning","task similarity graph","community detection","Louvain modularity","feature anchors","communication efficiency","non-IID data"],"falsifier":"A direct falsifier is an ablation on Heterogeneous CIFAR-10 in which the similarity weights entering Eq. (15) are replaced by random permutations of the true weights: if test accuracy does not drop materially, the learned similarity and community structure are not what produce the reported gains. A complementary check is a synthetic benchmark with known task clusters, asking whether Louvain communities recover the ground-truth clusters each round; persistent mismatch would disprove the claim that the metric orders task relatedness correctly.","tokens_in":16291,"feed_emoji":"🤝","tokens_out":8015,"duration_ms":76968,"temperature":0.7,"pith_summary":"This paper tries to establish that personalized federated learning can be made communication-efficient and more accurate by having clients share only compact feature anchors and linear classification heads, not full models, and by letting the server restrict collaboration to detected communities of similar clients. The authors claim that a dynamically updated similarity graph, built from cosine similarities between anchors and between heads' responses, correctly identifies which clients should cooperate, and that clustering this graph with the Louvain method prevents negative knowledge transfer. If the claim holds, heterogeneous federated networks gain personalized accuracy with orders of magnitude less communication and fairer outcomes across clients. The experiments aim to show this on two non-IID image benchmarks, Heterogeneous CIFAR-10 and Rotated & Masked MNIST.","feed_headline":"Client communities cut federated communication 100x and lift accuracy","feed_subtitle":"Sharing only anchors and heads, clients learn from similar peers, beating three baselines on accuracy, cost, and fairness.","key_machinery":"The carrying object is the pair (feature anchor $h_k$, classification head $\\phi_k$). A feature anchor is a per-class prototype vector of the feature extractor's output, summarizing the local data distribution, and the head is the lightweight linear layer mapping features to logits; both are transmitted to the server instead of the full model. The server builds a weighted graph with edge weight $a_{k\\ell} = \\alpha \\cdot \\mathrm{Sim}_{\\text{head}}(k,\\ell) + (1-\\alpha) \\cdot \\mathrm{Sim}_{\\text{repr}}(k,\\ell)$, where $\\mathrm{Sim}_{\\text{head}}$ averages the cosine similarity of the two heads' logits on both clients' anchors for each shared class and $\\mathrm{Sim}_{\\text{repr}}$ is the cosine similarity of the anchors themselves, and then solves a modularity maximization with the Louvain algorithm to partition clients into communities. Collaboration is executed by aggregating heads within a community via the regularized update $\\phi_k^{t+1} = \\phi_{k,R}^{t} - \\lambda \\tau \\sum_{\\ell \\in C_j} a_{k,\\ell}^{t}(\\phi_{k,R}^{t} - \\phi_{\\ell,R}^{t})$, which pulls each head toward community peers proportionally to $a_{k\\ell}$, and by replacing each client's local anchor with the sample-weighted community anchor. This mechanism is what the authors claim enforces positive transfer and preserves personalization.","core_discovery":"The central claim is that selective collaboration through community detection outperforms unselective collaboration and global averaging in federated multi-task learning. On Heterogeneous CIFAR-10, the proposed SFMTL-Graph reaches about 0.92 mean accuracy, roughly 9 percentage points above the closest baseline FedU, while transmitting less than $10^{8}$ bits over 200 rounds; FedU, FedAvg, and pFedMe each transmit over $10^{10}$ bits, more than SFMTL-Graph's cumulative total in a single round. The authors attribute the gain to a similarity weight that combines functional similarity of classification heads (cosine similarity of their logits evaluated on both clients' feature anchors) with representational similarity of anchors, followed by Louvain community detection that restricts the regularized head update to peers inside the same community. On Rotated & Masked MNIST the method again ranks first, with a smaller margin, and the fairness analysis reports the lowest variance in client accuracies and the best mean accuracy for the worst 10% of clients.","pith_inferences":["An implication the authors leave implicit is that the communication budget scales with the number of local classes and the head width, not the backbone size, so the scheme becomes increasingly attractive for large backbones; whether the accuracy gap survives on harder tasks such as ImageNet-scale heterogeneity is untested.","The similarity proxy is never calibrated against ground-truth task distances; a direct test would construct a synthetic federated benchmark with known label distributions and check whether Louvain communities recover the true clusters, and whether randomizing the edge weights collapses the accuracy gain.","Because clients need the same number of local classes for the class-wise anchor comparison, the method as stated cannot handle clients with different label sets; a natural extension is to allow partial class overlap by comparing only shared classes or by aligning anchor sets with optimal transport.","The hard community assignment could be relaxed to soft membership, letting a client interpolate between its own head and a global mixture; the paper does not explore this, but it would test whether the sharp community boundary is what drives the gain or whether a soft version is equally effective."],"forward_implications":["On Heterogeneous CIFAR-10, SFMTL-Graph's mean accuracy of about 0.92 exceeds the closest baseline FedU by roughly 9 points, and it is the top method on Rotated & Masked MNIST as well.","Cumulative communication over 200 rounds stays below 10^8 bits, less than a single round of any baseline, because clients transmit only a small anchor (1024 values) and a 512 by 10 head.","Local compute stays at the level of FedAvg and FedU (about 1.25 times 10^8 FLOPS on CIFAR-10), while pFedMe's inner-loop optimization costs roughly 1.5 times 10^10 FLOPS for the same task.","Fairness improves: the worst-10% mean accuracy is 0.75 with standard deviation 0.11, the standard deviation across clients is the lowest at 0.08, and clients sharing the same label set move closer in representation space over rounds."],"supporting_citations":[{"why":"Defines FedAvg, the standard global-model baseline the paper must beat in non-IID settings.","marker":"[1]"},{"why":"pFedMe is the personalized-federated-learning baseline compared in the experiments.","marker":"[14]"},{"why":"FedU is the closest baseline, the graph-Laplacian federated multi-task method whose regularized-update idea the paper extends.","marker":"[19]"},{"why":"Introduces class-wise feature centroids sent to the server, the direct precursor of the paper's feature anchor.","marker":"[23]"},{"why":"FedFA supplies the learned feature anchor and feature-alignment mechanism that SFMTL-Graph adapts into its compact client representation.","marker":"[24]"},{"why":"The Louvain algorithm is the community-detection heuristic that partitions the similarity graph, the load-bearing selection mechanism.","marker":"[33]"}],"fun_headline_variants":["Community detection makes federated multi-task learning selective and efficient","Selective federated learning: share anchors, group similar clients via communities","Federated learning with community-based selective collaboration cuts communication","Community-based collaboration boosts federated learning accuracy and efficiency","Selective peer grouping via community detection improves federated learning performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the premise that the cosine-similarity weights in Eq. (7) correctly rank how related two clients' tasks are, together with the assumption that every client has the same number of local classes so class-wise anchor comparisons are valid; if either fails, communities will group incompatible clients and collaboration will push models the wrong way.","fun_headline_variants_meta":{"raw":{"variants":["Community detection makes federated multi-task learning selective and efficient","Selective federated learning: share anchors, group similar clients via communities","Federated learning with community-based selective collaboration cuts communication","Community-based collaboration boosts federated learning accuracy and efficiency","Selective peer grouping via community detection improves federated learning performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001737,"raw_usage":{"total_tokens":6877,"prompt_tokens":969,"completion_tokens":5908,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":5824}},"tokens_in":585,"tokens_out":5908,"duration_ms":42252,"temperature":1.0,"reasoning_tokens":5824,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:35:10.536922+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct falsifier is an ablation on Heterogeneous CIFAR-10 in which the similarity weights entering Eq. (15) are replaced by random permutations of the true weights: if test accuracy does not drop materially, the learned similarity and community structure are not what produce the reported gains. A complementary check is a synthetic benchmark with known task clusters, asking whether Louvain communities recover the ground-truth clusters each round; persistent mismatch would disprove the claim that the metric orders task relatedness correctly.","supporting_citations":[{"cited_title":"Personalized federated learning with moreau envelopes,","cited_arxiv_id":null,"evidence_quote":"pFedMe is the personalized-federated-learning baseline compared in the experiments."},{"cited_title":"A new look and convergence rate of federated multitask learning with laplacian regularization,","cited_arxiv_id":null,"evidence_quote":"FedU is the closest baseline, the graph-Laplacian federated multi-task method whose regularized-update idea the paper extends."},{"cited_title":"Fedfa: Federated learning with feature anchors to align features and classifiers for heterogeneous data,","cited_arxiv_id":null,"evidence_quote":"FedFA supplies the learned feature anchor and feature-alignment mechanism that SFMTL-Graph adapts into its compact client representation."},{"cited_title":"Fast unfolding of communities in large networks,","cited_arxiv_id":null,"evidence_quote":"The Louvain algorithm is the community-detection heuristic that partitions the similarity graph, the load-bearing selection mechanism."}],"review_version":1}