{"id":"49c6cf70-8baf-43c7-9ef0-78965561d754","arxiv_id":"2412.00452","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"FedGR exploits the slow memorization of noisy labels by the global federated model to sieve, relabel, distill, and regularize local training, outperforming existing F-LNL methods on three benchmarks.","lead":"This paper introduces FedGR, a method that uses the global federated model to find and fix mislabeled client data before each training round. It reports large accuracy gains over prior federated noise-robust methods on CIFAR-10, CIFAR-100, and Clothing1M, including under high label noise and non-IID data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Centralized sieving's GMM (Eq. 8) is load-bearing but validated only client-level on symmetric CIFAR-10 noise; per-sample misassignments at round α propagate wrong labels through Eq.","rationale":"Reading the paper in good faith, the central claim is that the slow-memorization property of the FL global model can be harvested by three modules (CS/LR, revised-EMA distillation, representation regularization) to reach SOTA accuracy across heterogeneous label-noise regimes; Tables 1-3 support this empirically, and Table 4 attributes the largest single gain to centralized sieving (83.91 vs 54.59 w/o CS, Sym phi=1.0 I.I.D.), with strong augmentation also critical (34.51 w/o it). Independent support: public code, 3-seed means with stds, consistent trends across three datasets, and ablations that honestly expose the strong-aug dependence. What must be true for the claim: the per-sample loss statistics on which CS and Eq. 15 depend must separate clean from noisy samples for every client and every claimed noise type, and the global model's slow memorization must persist in the reported regimes (100 clients, 0.1 sample ratio, 500 rounds). I examined alternatives: (i) the abstract/body baseline-count mismatch (eight vs seven) is cosmetic; (ii) the lack of sensitivity analysis for beta, epsilon, delta, kappa, mu weakens robustness claims but not the headline; (iii) the Eq. 23 piecewise definition leaves gamma_g undefined for r_k >= beta at t >= delta (since |D~_k|=|D^_k| by construction), which is a genuine but secondary reproducibility defect; (iv) the observation experiments use 10 clients with 0.2-1.0 sample ratios while the main runs use 100 clients at 0.1, but the method's own success in the main regime indirectly supports the observation there. The least-secure link is the sample-level fidelity of the GMM sieving. It is the foundation of the two largest ablation contributors, yet the paper's validation (Fig. 4, Pearson and client-level F-scores) covers only CIFAR-10 symmetric noise. The failure modes are concrete: early-memorized noisy samples have low mean loss at round alpha and can fall into the clean component, preserving wrong labels through Eq. 15 for the remaining rounds; for majority-noise clients, a 2-component GMM can misestimate r_k and bypass the beta=0.8 protection; asymmetric flips between confusable classes are the slow-memorization property's hardest case, and no Asym/Mixed per-sample metrics are reported. All of these are testable with the oracle-mask experiment and per-sample precision/recall described above. If the test shows a small gap, the concern does not land and the paper's mechanism is confirmed; if the gap is large, the claimed generality requires per-sample validation or redesign. Since the reader already conditioned acceptance on this assumption and my analysis sharpens but does not move it, the verdict stays CONDITIONAL.","tokens_in":23718,"tokens_out":26513,"duration_ms":254748,"concrete_test":"On CIFAR-10, rerun FedGR exactly as configured (alpha=100, beta=0.8, epsilon=0.9) under Sym phi=1.0 U(0.5,1.0), Asym phi=1.0, and Mixed phi=1.0, both I.I.D. and Non-I.I.D., replacing only the GMM selection at Eq. 8 with an oracle mask built from the ground-truth clean/noisy status of each sample while keeping all other modules, hyperparameters, and the 3-seed protocol identical. Also record, at round alpha, per-sample precision and recall of the GMM against ground-truth noise status, broken down by noise type and by client noise-rate bucket (rho<0.5, 0.5-0.8, >0.8). If the oracle-mask accuracy exceeds the reported FedGR by at least 3 points, or if noisy-sample recall falls below 90% in any bucket, GMM misassignments are material to the central claim and the paper must supply per-sample validation; if the gap is at most 1 point and recall is high everywhere, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on the sniffing-then-refining module: Table 4 shows that removing centralized sieving (CS) drops CIFAR-10 Sym phi=1.0 from 83.91 to 54.59, the largest single ablation in the paper. CS and label refining (Eq. 15) both presuppose that a two-component GMM over per-sample mean losses (Eq. 8), fit on the server at round alpha=100, assigns clean/noisy status accurately in every claimed regime (Sym 50-100%, Asym 20-40%, Mixed, I.I.D. and Non-I.I.D.). That presupposition is the least-secure condition for three reasons. (1) Evidence: Fig. 4 reports only client-level Pearson between estimated and true noise ratios and client-level F-scores, and only for CIFAR-10 symmetric noise; per-sample precision/recall is never reported, and Asym/Mixed are unvalidated. (2) Timing: the GMM mask is computed once from Phase-I losses and then applied throughout Phase II; samples the global model memorizes early have low mean loss at round alpha and can fall into the clean component, so Eq. 15 keeps their wrong labels for the remaining ~400 rounds. (3) Majority-noise clients: when the true rho exceeds 0.8, the clean component is a minority in the pooled GMM and may absorb systematically easy mislabeled samples (e.g., asymmetric flips between confusable classes), so the estimated r_k drops below beta=0.8 and the all-pseudo-label protection is bypassed. A secondary internal inconsistency: in Eq. 23, since |D~_k|=|D^_k| by construction, gamma_g is undefined for r_k >= beta at t >= delta, so the EMA revision for the highest-noise clients is unspecified as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FedGR, a federated learning (FL) algorithm for label noise, based on the observation that the global FL model memorizes noisy labels more slowly than a centrally trained model. FedGR has three modules: (i) sniffing-then-refining, where the server fits a two-component GMM to per-sample mean losses computed with the global model, splits each client's data into clean/noisy sets, and refines labels via pseudo-labels from the global model; (ii) global revised EMA distillation, where the local EMA model is bootstrapped and periodically revised with the global model, and distilled into the client model; and (iii) global representation regularization, which distills the global model's representations to the local model. The method is evaluated on CIFAR-10, CIFAR-100, and Clothing1M under IID and Dirichlet-non-IID partitions, with symmetric, asymmetric, and mixed label noise, and is compared against seven baselines.","tokens_in":24104,"tokens_out":8569,"duration_ms":79612,"significance":"If the premises hold, FedGR is a meaningful advance for FL with noisy labels. The paper's observation about slow memorization in the global FL model is interesting and potentially useful. The experimental evaluation is broad, covering three benchmarks, two data partitions, multiple noise types and ratios, with ablations and three seeds. The main strengths are the global (server-side) noise sieving, which avoids per-client noise modeling under heterogeneity, and the strong empirical gains over existing F-LNL methods. However, the central noise-sniffing mechanism is validated only at client level on one setting, and the algorithm specification contains definitional gaps that affect the high-noise regime. The study also leaves the sensitivity to its many hyperparameters unexamined.","major_comments":[{"comment":"The centralized sieving step is the largest single contributor in the ablation: removing it drops CIFAR-10 Sym phi=1.0 from 83.91 to 54.59 in Table 4. Yet its validity is only demonstrated through client-level Pearson coefficients and client-level F-scores in Fig. 4, and only for CIFAR-10 symmetric noise. Please report per-sample precision/recall of the two-component GMM at round alpha for symmetric, asymmetric, and mixed noise under both IID and non-IID partitions, and show how per-sample errors evolve in Phase II (e.g., precision of the clean set over rounds). Without this, the central claim that the global model can reliably sniff noise is not directly supported.","section":"Section 3.2, Eqs. (8)-(15), Table 4, Fig. 4"},{"comment":"The definition of gamma_g is incomplete. For r_k >= beta and t >= delta, the first case requires r_k < beta, and the second case requires (r_k >= beta and |D~_k|/|D_k| < mu) or t < delta; since t >= delta and |D~_k|/|D_k| = 1 because Eq. (15) assigns a refined label to every sample, neither branch matches. Please specify the intended value of gamma_g for high-noise clients that successfully refine their data, and reconcile the ratio |D~_k|/|D_k| with the definition of D~_k.","section":"Section 3.3, Eq. (23)"},{"comment":"For clients with r_k >= beta, the refined label is y^pse_i, which is the zero vector whenever max(sigma(p^{g,w}_i)) <= epsilon. The paper does not say how the cross-entropy loss in Eq. (10) treats zero-vector targets; if low-confidence samples are masked out or assigned a uniform label, this should be stated explicitly. This branch is precisely the high-noise regime that differentiates the method from FedCorr, so it must be unambiguous.","section":"Section 3.2, Eqs. (10), (12), (15)"},{"comment":"The slow-memorization observation is made on CIFAR-10 with 10 clients and client sample ratios 0.2, 0.5, and 1.0, but the main experiments use 100 clients (and 500 for Clothing1M), different backbones, and local epochs of 10. Because all three modules rely on the global model being less overfit to noisy labels, please report the memorization metric under the actual experimental configuration (e.g., 100 clients, CIFAR-10/100), or provide a principled argument for why the phenomenon should transfer. A sensitivity analysis on alpha (the sniffing length) would also help establish that the GMM mask is not overly sensitive to the warm-up point.","section":"Section 1 and App. A.1"}],"minor_comments":[{"comment":"The abstract and Section 4 disagree on the number of baselines (eight in the posted abstract vs. seven in the paper's abstract and Section 4); please harmonize.","section":"Abstract vs. Section 4"},{"comment":"The Non-IID symmetric phi=1.0 FedGR accuracy in Table 4 (64.24±5.10) differs from the corresponding entry in Table 1 (63.64±5.39) for what appears to be the same configuration; please verify which number is correct.","section":"Section 4.3, Table 4 vs. Table 1"},{"comment":"Figure 1's caption says the global model memorizes no more than 30% of noisy samples, while App. A.1 says no more than 50%; these statements should be reconciled.","section":"Figure 1 vs. App. A.1"},{"comment":"Equations (19) and (24) write KL(·, ·) on logits and on features, but KL divergence is defined between distributions; please clarify whether softmax (or a normalized form) is applied before the KL, and define the temperature tau's role in Eq. (24).","section":"Section 3.3, Eqs. (19), (24)"},{"comment":"Clothing1M results in Table 3 are reported without standard deviation and are based on a single seed (App. A.2); please report at least the average over a few seeds or state clearly why this is not feasible.","section":"Section 4.1, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution, but the acceptance decision should hinge on the authors' response to the GMM validation and the definitional gaps in Eqs. (15) and (23). The claim that FedGR beats a cleanly-trained model should be tempered, since the comparison includes strong augmentation in FedGR but not in the clean baseline. I also flagged the inconsistency between Tables 1 and 4. If these points are addressed convincingly, the paper would be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. First, the core observation—the federated global model memorizes noisy labels much more slowly than a centrally trained model—is real and well illustrated; it is the most interesting part of the work. Second, the method built on it, FedGR, reports large accuracy gains in several F-LN settings (e.g., CIFAR-10 symmetric 50–100%: 83.91 vs. 55.12 for FedCorr), with an extensive ablation and three-seed runs. I think the paper deserves a serious referee, but the mechanism that carries the weight is not yet properly supported.\n\nWhat is new: using the global model as a server-side sieve and label refiner is a clean idea that flips the usual client-side selection. The three modules are known ingredients, but the global-model framing is a genuine variation. The experiments are thorough by the standards of the area: three datasets, IID and Dirichlet-non-IID, symmetric/asymmetric/mixed noise, ablations of each module. The main tables are readable and the claims match the tables.\n\nWhere I'd push back. The ablation shows that removing centralized sieving (CS) drops CIFAR-10 Sym phi=1.0 from 83.91 to 54.59, so CS is clearly load-bearing. Yet the paper validates the GMM sieve only with client-level Pearson correlations and F-scores on CIFAR-10 symmetric noise (Fig. 4). No per-sample precision/recall is given, and asymmetric and mixed setups are not validated at sample level. Since Eq. 15 uses the GMM's per-sample clean probability to decide which labels to keep, the whole method depends on per-sample accuracy. The one-shot sieve at round α=100 is then kept fixed for the remaining ~400 rounds, and the paper doesn't analyze how early memorized noisy samples with low mean loss could slip into the clean component and keep their wrong labels.\n\nThere is also a concrete bug in Eq. 23. The condition |D~_k|/|D^_k| < µ can never be true because relabeling does not change the number of samples: |D~_k| = |D^_k| always. So for clients with r_k ≥ β and t ≥ δ, γ_g is undefined as written. This is probably a typo (they may mean the fraction of samples whose pseudo-label exceeded the confidence threshold), but it has to be fixed and the intended semantics stated.\n\nMinor issues: the abstract says 'eight state-of-the-art baselines' while the body says seven; hyperparameters in Tab. 7 are numerous and no sensitivity analysis is given.\n\nBottom line: the paper is serious, the central observation is valuable, and the method is worth engaging with. But the load-bearing sieve needs per-sample validation, and Eq. 23 must be clarified. I'd send it to review with a request for those revisions.","headline":"FedGR's headline gains on federated label noise are plausible and the slow-memorization observation is worth taking seriously, but the load-bearing GMM sieve is under-validated and Eq. 23 has an undefined branch.","tokens_in":24644,"tokens_out":3915,"would_cite":true,"duration_ms":36664,"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":"The federated global model memorizes noisy labels slowly, and FedGR converts that into robust label-noise handling.","keywords":["federated learning","label noise","noisy labels","memorization effect","sample selection","label refinement","knowledge distillation","representation regularization"],"falsifier":"Construct a federated benchmark where noisy labels are deliberately high-confidence (e.g., instance-dependent noise generated from a near-perfect teacher), so noisy samples have systematically lower mean loss than clean ones; if FedGR's centralized sieving mis-splits the sets, its accuracy should collapse toward the FedAvg baseline, confirming the GMM assumption is load-bearing.","tokens_in":23535,"feed_emoji":"🧠","tokens_out":4853,"duration_ms":42846,"temperature":0.7,"pith_summary":"Federated learning struggles when client labels are corrupted, and the noise differs from client to client. This paper claims there is a saving grace built into federated training itself: the aggregated global model fits noisy labels much more slowly than a centrally trained model, so it stays a reliable teacher throughout training. On that observation the authors build FedGR, which uses the global model to sniff out noisy samples on the server, refine their labels, revise each client's exponential moving average model, and regularize local representations. The result, they report, is state-of-the-art accuracy across CIFAR-10, CIFAR-100, and Clothing1M under symmetric, asymmetric, and mixed label noise, including settings where every client is noisy.","feed_headline":"FedGR exploits the global model's slow memorization of noisy labels","feed_subtitle":"Using the server's model to sieve and relabel corrupted samples lifts CIFAR-10 from 55% to 84% at extreme noise.","key_machinery":"Three modules, all driven by the global model, carry the argument. Centralized sieving computes each sample's mean loss across the rounds its client participated, fits a two-component Gaussian Mixture Model on the server, and labels each sample clean or noisy; this is what the whole method relies on. Label refining replaces noisy labels with confident pseudo labels $\\text{onehot}(\\arg\\max \\sigma(p_g))$ when the estimated noise ratio $r_k \\geq \\beta$, and otherwise mixes the original label with the pseudo label according to the GMM clean probability (Eq. 15). Global revised EMA distillation bootstraps each client's local EMA model from the global model at round $\\delta$ and periodically mixes global weights into it, then distills its logits into the client model; global representation regularization distills the global backbone's representation of weakly augmented images into the local model's strong-augmentation branch. The load-bearing observation behind all three is that the global model's slow memorization keeps these pseudo-labels and representations trustworthy.","core_discovery":"The paper's central claim is that the global model of federated learning is a naturally reliable reference for handling label noise: while a centrally trained network ends up memorizing over 80% of corrupted labels, the federated global model memorizes no more than 30–50% and never shows the characteristic test-drop of overfitting. FedGR operationalizes this. During a warm-up phase, clients compute per-sample cross-entropy losses with the global model and upload the mean losses; the server fits a two-component Gaussian Mixture Model to these statistics and returns clean/noisy splits plus estimated noise ratios. In the refinement phase, each client's labels are rebuilt by mixing the original label with a confident pseudo-label from the global model, with the mixing determined by the estimated noise ratio. Parallel modules revise the client's local EMA model with global parameters and distill its logits, and distill the global representation into the local model, so that noisy samples still contribute knowledge without their wrong labels. On CIFAR-10 with 50–100% symmetric noise and no clean clients, the method reports 83.91% accuracy versus 55.12% for FedCorr.","pith_inferences":["The slow-memorization effect likely has a mechanical explanation the paper leaves implicit: averaging many client models at the server smooths the loss landscape and delays the phase where individual networks lock onto wrong labels; a theory of aggregation as implicit regularization could make the observation predictive rather than empirical.","The GMM-on-mean-loss assumption is only validated at client level (Pearson correlation on noise ratios), not sample level; a direct sample-level precision-recall evaluation would show whether the sieving holds when noise is instance-dependent.","One testable extension is to make the confidence threshold $\\epsilon$ and noise-ratio threshold $\\beta$ adaptive per client, since the paper fixes them globally; the mixed-noise results suggest the optimal thresholds depend on the noise type.","Because FedGR only needs the global model plus loss statistics, it could be combined with differential-privacy mechanisms on the uploaded mean losses if a formal privacy bound were required."],"forward_implications":["When some clients are clean ($\\phi=0.6$), FedGR matches or exceeds training on clean labels, on both CIFAR-10 and CIFAR-100.","Under the hardest tested setting — every client noisy, symmetric noise between 50% and 100% — FedGR keeps 83.91% accuracy on CIFAR-10, where FedCorr drops to 55.12%.","The method transfers to a large real-world noisy dataset, Clothing1M, where it leads across both IID and Non-IID partitions.","Only per-sample loss statistics and model parameters are exchanged with the server, so the approach avoids transmitting class centers or other high-risk local proxies used by prior methods.","Ablations attribute most of the gain under high noise to centralized sieving and label refining, with the two distillation and regularization modules contributing further gains."],"supporting_citations":[{"why":"Supplies the memorization-effect premise that deep networks fit noisy labels late, which the paper verifies holds for the federated global model.","marker":"[2]"},{"why":"Provides the two-phase training and small-loss-trick template that FedGR's sniffing-then-refining adapts to the federated setting.","marker":"[7]"},{"why":"Contributes the GMM-based sample selection and label-refurbishment strategy that FedGR ports to the server with the global model.","marker":"[16]"},{"why":"Defines the FedAvg aggregation rule that produces the global model whose slow memorization is the paper's key observation.","marker":"[23]"},{"why":"Supplies the confidence-thresholded pseudo-labeling mechanism used in Eq. 12 for label refinement.","marker":"[25]"},{"why":"The main F-LNL baseline and the source of the multi-stage training setup and data-augmentation schedule that FedGR inherits.","marker":"[33]"},{"why":"A strong baseline for heterogeneous label noise in FL that FedGR compares against and improves upon.","marker":"[10]"}],"fun_headline_variants":["Global model's slow noise memorization powers FedGR's label fixing","FedGR turns slow global memorization into label-noise armor","How FedGR sieves noisy labels via the global model's pace","FedGR's global revisor lifts CIFAR-10 from 55% to 84% noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a two-component Gaussian Mixture Model fit to per-sample mean losses cleanly separates clean from noisy samples on every client, no matter the noise type or data distribution.","fun_headline_variants_meta":{"raw":{"variants":["Global model's slow noise memorization powers FedGR's label fixing","FedGR turns slow global memorization into label-noise armor","How FedGR sieves noisy labels via the global model's pace","FedGR's global revisor lifts CIFAR-10 from 55% to 84% noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001091,"raw_usage":{"total_tokens":4576,"prompt_tokens":986,"completion_tokens":3590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":3509}},"tokens_in":602,"tokens_out":3590,"duration_ms":21878,"temperature":1.0,"reasoning_tokens":3509,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:22:07.637047+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a federated benchmark where noisy labels are deliberately high-confidence (e.g., instance-dependent noise generated from a near-perfect teacher), so noisy samples have systematically lower mean loss than clean ones; if FedGR's centralized sieving mis-splits the sets, its accuracy should collapse toward the FedAvg baseline, confirming the GMM assumption is load-bearing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contributes the GMM-based sample selection and label-refurbishment strategy that FedGR ports to the server with the global model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The main F-LNL baseline and the source of the multi-stage training setup and data-augmentation schedule that FedGR inherits."},{"cited_title":"FedFixer: Mitigating Hetero- geneous Label Noise in Federated Learning","cited_arxiv_id":null,"evidence_quote":"A strong baseline for heterogeneous label noise in FL that FedGR compares against and improves upon."}],"review_version":2}