{"id":"f5e284df-f279-4e17-b638-18e2ce1bc82a","arxiv_id":"2508.12042","paper_version":3,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FairGrad and FairGrad* add gradient variance penalties to federated learning, and the authors report these reduce accuracy variance across clients while keeping or improving mean accuracy, though the exact version's gains are inconsistent.","lead":"The paper proposes new federated learning methods that reduce performance differences between clients by penalizing how much each client's gradient deviates from the average. It reports improvements in fairness and accuracy on image benchmarks, but the exact variant performs inconsistently and the theory rests on questionable assumptions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own tables contradict the central claim that FairGrad and FairGrad* consistently improve fairness; for CIFAR-10 α=0.1, FairGrad* is worse than FedAvg in both accuracy and variance.","rationale":"The central claim of the paper is empirical: FairGrad and FairGrad* consistently improve fairness while maintaining competitive or improved mean performance. A load-bearing condition for that claim is that the reported tables actually show such improvements. They do not. Table 2 and Table 4 contain direct counterexamples where the proposed methods have equal or higher variance than FedAvg, or lower accuracy, or both. This is an internal inconsistency, not a disagreement with an external consensus, so it directly undercuts the abstract and the conclusion's statement that regularized methods improve both fairness and overall model performance. The reader's weakest_assumption identified the Theorem 4.1 proof gap, which is real: E[∇fi(x*) | fi(x*) - F(x*)] = 0 does not follow from homogeneity because fi and ∇fi are computed from the same samples and are generally correlated. I treat that as a separate, serious correctness problem for the theoretical section, but it is less load-bearing than the empirical contradiction, because the paper's headline contribution is the claimed consistent empirical improvement. The best-epoch selection protocol in Section 6.1 is another important concern: selecting the best-performing epoch per run can inflate apparent method quality and, if selection uses test data, constitutes leakage. A fixed-protocol rerun would settle whether any fairness improvement is real and consistent. For these reasons, the reader's REJECT verdict stands unchanged.","tokens_in":26617,"tokens_out":5899,"duration_ms":62024,"concrete_test":"Re-run Tables 1-4 under a fixed protocol: for each seed and method, evaluate at the final global epoch only, or at a single validation-based early-stopping epoch chosen without using test labels. For each dataset and each α, count whether FairGrad and FairGrad* have strictly lower mean test-variance than FedAvg, AAggFF, and q-FFL, and accuracy not lower by more than one standard error. If the same counterexamples appear (e.g., CIFAR-10 α=0.1 for FairGrad*, Tiny ImageNet α=0.05 and α=0.1 for FairGrad), the claimed 'consistent' improvement is unsupported. If fixed-protocol results show strict variance improvements in every cell, the central empirical claim would survive despite the current protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's headline claim is that 'FairGrad and FairGrad* consistently improve fairness by reducing variance in client accuracies, while maintaining competitive or improved mean performance compared to existing fairness-aware baselines.' The reported results fail this assertion in multiple cells. On CIFAR-10 (Table 2, α=0.1), FairGrad* has accuracy 84.82±0.16 vs FedAvg's 84.89±0.36 and variance 1.55±0.09 vs FedAvg's 1.50±0.14; at α=0.5 its variance is identical to FedAvg (0.75) while accuracy is essentially tied. On Tiny ImageNet (Table 4), FairGrad has variance 0.52 vs FedAvg's 0.47 at α=0.05 and 0.56 vs 0.53 at α=0.1, so it does not reduce variance even though accuracy improves. Thus the 'consistently improve fairness' assertion is not merely statistically underpowered; it is directly contradicted by the paper's own point estimates. The evaluation protocol compounds this: Section 6.1 states that 'the best-performing epoch was identified' for each training run, so the reported numbers are post-hoc selections over training time. Without a fixed, pre-registered epoch-selection rule, apparent advantages of one method over another are not reliable estimators of the method's typical behavior. The proof of Theorem 4.1 in Appendix A.1 also contains a genuine gap, since E[∇fi(x*) | fi(x*) - F(x*)] = 0 is not implied by identical client distributions, but the empirical inconsistency alone is sufficient to remove support for the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies performance-equitable fairness in federated learning, defining fairness as low variance of client accuracy. It evaluates q-FFL, FedAvg, AAggFF, and two newly proposed gradient-variance-regularized variants, FairGrad and FairGrad*, together with loss-variance variants FairLoss and FairLoss*. The theoretical section analyzes stationary points of these regularized objectives under homogeneous and heterogeneous data, claiming that homogeneous settings preserve FedAvg stationary points while heterogeneous settings can shift them. The experiments on MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet with Dirichlet heterogeneity report test accuracy and across-client variance, and the paper claims that the proposed gradient-alignment methods consistently improve fairness while maintaining competitive accuracy.","tokens_in":27043,"tokens_out":6865,"duration_ms":64461,"significance":"The paper addresses a relevant problem and makes a useful organizational contribution by explicitly comparing several regularized fairness objectives and by noting a formal connection between q-FFL (q=1) and loss-variance regularization. The idea of regularizing gradient variance for performance-equitable fairness is reasonable, and the approximate/exact algorithmic distinction is sensible. However, the central empirical claim is contradicted by the paper's own tables, and the main theoretical lemma has a conditional-independence gap. In its current form the paper does not support its headline conclusions. No code repository is mentioned, which limits reproducibility of the experiments.","major_comments":[{"comment":"The abstract and Section 7 claim that FairGrad and FairGrad* 'consistently improve fairness' by reducing variance in client accuracies. The reported point estimates contradict this claim. On CIFAR-10 with α=0.1 (Table 2), FairGrad* has test variance 1.55±0.09 versus FedAvg's 1.50±0.14, and lower accuracy (84.82 vs 84.89). On Tiny ImageNet (Table 4), FairGrad has variance 0.52 vs FedAvg's 0.47 at α=0.05 and 0.56 vs 0.53 at α=0.1, so it fails to reduce variance in both settings. Since 'consistently improve' is the paper's headline claim, this empirical inconsistency is load-bearing and cannot be resolved by restating the results as trends.","section":"Abstract; Section 7; Tables 2 and 4"},{"comment":"The proof of Theorem 4.1 invokes the law of total expectation and asserts that E[∇fi(x*) | fi(x*) − F(x*)] = 0 under data homogeneity. Identical client distributions do not imply this conditional independence, because fi(x*) and ∇fi(x*) are computed from the same samples and are generally correlated. Without this assumption the conclusion E[∇Lλ(x*)] = 0, and the analogous statement for Jγ, does not follow. This gap also undermines the later claim in Theorem 4.2 that E[∇H1(x*)] = 0.","section":"Appendix A.1, Theorem 4.1"},{"comment":"The evaluation protocol is confounded with the target metric. Section 6.1 states that 'the best-performing epoch was identified' for each training run without specifying the selection rule, and Appendix C selects hyperparameters by maximizing acc − 1.96·sqrt(var/n), where var is the validation variance of client accuracies. Because the reported fairness metric is exactly the test variance, selecting configurations and epochs on variance and then reporting test variance does not provide an unbiased estimate of a method's typical fairness. A fixed epoch rule or a pre-registered protocol is needed before the empirical superiority claim can be assessed.","section":"Section 6.1; Appendix C"},{"comment":"The proof of Theorem 4.6 only shows that ∇L̂λ(x*) ≠ 0 for a chosen perturbation o, i.e., that x* is not a stationary point of L̂λ. The theorem's conclusion that L̂λ 'has a different stationary point than F' additionally requires existence of a stationary point ŷ ≠ x* of L̂λ, which is not established. As stated, the result is unsupported.","section":"Appendix A.2, Theorem 4.6"}],"minor_comments":[{"comment":"In Algorithm 1 the local gradient update is written as g_t^i = ∇f_i(x_t) + λ(f_i(x_t) − a_t)(∇f_i(x_t) − g_t^i), with g_t^i appearing on the right-hand side. From the description in Appendix B.1, the intent is to use the server-aggregated gradient g_t from the previous iteration; the pseudocode should be corrected.","section":"Algorithm 1"},{"comment":"The text says FairGrad* 'attained the lowest variance at α=0.05' on Tiny ImageNet, but Table 4 shows a tie with FedAvg at 0.47. Please clarify that the comparison includes ties.","section":"Section 6.1, Table 4 discussion"},{"comment":"In the proof of Theorem 4.2, the expression '∇F(x⋆)F(x⋆) = 0' appears to be a typo; as written it is dimensionally inconsistent.","section":"Appendix A.1, Theorem 4.2 proof"}],"recommendation":"reject","confidential_remarks":"The manuscript's headline claim is directly contradicted by its own reported numbers, and the proof gap in Theorem 4.1 cannot be fixed by local editing alone. The empirical evaluation would need to be redone under a pre-registered protocol, so I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is reasonable: penalize gradient variance across clients as a way to push performance-equitable fairness in federated learning, with an approximate and an exact variant plus a theoretical comparison to loss-variance regularization and q-FFL. The algorithms are clearly specified, and the empirical study covers four datasets and multiple heterogeneity levels. Credit where due: FairGrad on CIFAR-10 does show consistent gains in both accuracy and fairness, and the attempt to connect q-FFL at q=1 to loss-variance regularization is a genuine contribution to the subfield.\n\nThe soft spots are real and load-bearing. The abstract and conclusion claim that FairGrad and FairGrad* \"consistently improve fairness,\" but the paper's own tables contradict that. On CIFAR-10 with alpha=0.1, FairGrad* has lower accuracy and higher variance than FedAvg. On Tiny ImageNet with alpha=0.05 and 0.1, FairGrad has higher variance than FedAvg. These are not statistically marginal differences; they are the reported point estimates. So the central claim as written is false, even before worrying about significance.\n\nThe theory also has a genuine gap. The proof of Theorem 4.1 asserts E[\\nabla f_i(x*) | f_i(x*) - F(x*)] = 0 under data homogeneity. Identical client distributions do not imply this conditional independence, because f_i and \\nabla f_i are computed from the same local samples and are generally correlated. The same issue appears in the proof of Theorem 4.2, which also contains a variance-identity error: it treats the variance of a single client's loss as if it were the sample variance across clients, with incorrect scaling. These need fixing before the theoretical claims can be taken seriously.\n\nThe evaluation protocol compounds the problem. Section 6.1 states that the best-performing epoch was identified for each run, and the hyperparameter selection criterion explicitly trades accuracy against variance. Selecting the best epoch post hoc makes reported results optimistic, and since the fairness metric is the variance the methods are trained to reduce, the evaluation is partially circular.\n\nWho is this for? Researchers working on performance-equitable fairness in FL might find the FairGrad idea worth building on, and the paper is a useful cautionary example of overclaiming from mixed results. But as it stands, the central claim is unsupported, and the theory is not dependable. A serious revision that reports all cells honestly, fixes the conditional-independence step, and adopts a pre-registered evaluation protocol could make this a solid paper. Right now I would not cite it, but I would send it to reviewers if the authors asked for feedback, because the idea and the effort are there and the flaws are correctable.","headline":"The proposed FairGrad methods are a sensible idea, but the paper's own tables contradict the central claim of consistent fairness improvement, and the theory rests on an unjustified conditional-independence step.","tokens_in":27517,"tokens_out":1976,"would_cite":false,"duration_ms":22918,"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":"Adding a penalty on the spread of client gradients in federated training improves fairness across clients without sacrificing average accuracy.","keywords":["federated learning","performance equitable fairness","variance regularization","gradient variance","client accuracy disparity","data heterogeneity","q-FFL","FairGrad"],"falsifier":"On a dataset with provably identical client distributions, train FedAvg to a stationary point $x^\\star$, then estimate $\\mathbb{E}[\\nabla f_i(x^\\star)\\mid f_i(x^\\star)-F(x^\\star)]$ from client samples; if the estimate is significantly nonzero, the homogeneous-case proof of Theorem 4.1 collapses, and the claim that the regularizers are unnecessary when data are identical loses its theoretical footing.","tokens_in":26425,"feed_emoji":"⚖️","tokens_out":12795,"duration_ms":112676,"temperature":0.7,"pith_summary":"In federated learning, skewed client data can leave some participants with a noticeably worse global model, and this paper asks whether a small change to the training objective can close that gap. The authors propose adding a variance penalty—on client losses for FairLoss and FairLoss*, and on client gradients for the new FairGrad and FairGrad*—to the standard federated average of losses. They show theoretically that under identical client distributions these penalties leave the FedAvg solution unchanged, but under heterogeneity they move the stationary point so that client losses spread less. Empirically, across MNIST, CIFAR-10, CIFAR-100, and Tiny ImageNet with Dirichlet-based heterogeneity, the gradient-variance variants reduce the variance of client test accuracies while keeping or improving mean accuracy relative to FedAvg, q-FFL, and AAggFF. If this holds, fairer federated learning can be obtained by reweighting local gradients rather than redesigning aggregation.","feed_headline":"Gradient-variance penalty makes federated learning fairer","feed_subtitle":"New FairGrad methods shrink accuracy gaps across clients without sacrificing overall performance.","key_machinery":"The load-bearing object is the gradient-variance regularizer $J_\\gamma(x)=F(x)+\\frac{\\gamma}{2n}\\sum_{i=1}^n\\|\\nabla f_i(x)-\\nabla F(x)\\|^2$, where $F$ is the average client loss and $\\nabla F$ the average gradient; the FairGrad variant approximates it with the previous round's average gradient (one communication round), while FairGrad* recomputes the average in the current round (two rounds). Together with the loss-variance objective $L_\\lambda(x)=F(x)+\\frac{\\lambda}{2n}\\sum_{i=1}^n(f_i(x)-F(x))^2$, it forms the pair of regularizers the analysis compares. The key identity connecting the two families shows that q-FFL with $q=1$ satisfies $\\mathbb{E}[H_1(x)]=\\frac{1}{2n}\\sum_i \\mathbb{E}[f_i(x)^2]+\\frac{n}{n-1}\\mathbb{E}[\\frac{1}{2n}\\sum_i(f_i(x)-F(x))^2]$, and when $\\mathbb{E}[f_i(x)]\\le 1$ this is bounded above by $\\mathbb{E}[L_\\lambda(x)]$ with $\\lambda=n/(n-1)$, so the loss-variance regularizer and the q-FFL reweighting coincide at this parameter point. The theorems on stationary points (4.1–4.7) then delineate when the regularizers preserve the FedAvg optimum (homogeneous data), shift it (heterogeneous data with misaligned losses or gradients), or can be shown to reduce the variance of client losses at the new optimum under convexity.","core_discovery":"The central claim is that performance equitable fairness in federated learning—minimizing the spread of client accuracies—can be achieved by regularizing the variance of client gradients during training, not just the variance of client losses. The paper introduces the objective $J_\\gamma(x)=F(x)+\\frac{\\gamma}{2n}\\sum_{i=1}^n\\|\\nabla f_i(x)-\\nabla F(x)\\|^2$, with a cheaper approximate variant (FairGrad, using the previous round's average gradient) and an exact variant (FairGrad*, using the current round's average). It proves that in a homogeneous setting the regularized objectives share FedAvg's stationary points in expectation, so the penalties only matter when client distributions differ; in heterogeneous settings they generically shift the optimum away from the FedAvg solution (Theorems 4.1–4.7). A companion identity relates q-FFL with $q=1$ to the loss-variance objective with $\\lambda=n/(n-1)$. The empirical section reports that on four standard image benchmarks with 10–100 clients and Dirichlet heterogeneity, FairGrad and FairGrad* achieve the lowest client-accuracy variance in most configurations and the highest or competitive mean accuracy, corroborating the fairness mechanism.","pith_inferences":["The exact variants send per-client losses and gradients to the server before updating, which weakens the privacy guarantee that motivates federated learning; a natural extension is to combine the gradient-variance penalty with secure aggregation or gradient compression and test whether the fairness gains survive.","Because the regularizer acts on per-client gradients and losses, it could be dropped into other FL objectives (personalized heads, multi-task splits) where the variance penalty would then apply only to the shared parameters; the paper does not test this.","The empirical gains are selected by tuning $\\lambda,\\gamma,q$ with a validation criterion that explicitly rewards low variance, so a stricter test of the fairness claim would fix the same hyperparameters across methods or constrain the tuning budget."],"forward_implications":["With identical client distributions, the variance regularizers change nothing in expectation: their stationary points coincide with FedAvg's, so any fairness benefit must come from data heterogeneity.","Under heterogeneity, the regularizers generically move the optimum away from the FedAvg solution, which is the mechanism by which client performance gaps narrow.","q-FFL with $q=1$ is approximately the loss-variance regularizer with $\\lambda=n/(n-1)$, so the two families are not separate approaches but the same regularization at a particular parameter setting.","When the one-round approximate variants (FairLoss, FairGrad) converge, their limits are stationary points of the exact regularized objectives $L_\\lambda$ and $J_\\gamma$, so the cheaper updates do not change the optimization target.","On CIFAR-10, CIFAR-100, and Tiny ImageNet with many clients, gradient-variance regularization (FairGrad or FairGrad*) yields the lowest variance of client accuracies and the highest or tied mean accuracy in most configurations, improving on FedAvg, AAggFF, and q-FFL."],"supporting_citations":[{"why":"Defines performance equitable fairness and the q-FFL baseline that the new regularizers are compared against; also the source of the fairness definition adapted in Definition 3.1.","marker":"Li et al. [2019]"},{"why":"Introduces FedAvg, the baseline objective $F(x)$ that all regularized methods reduce to when their penalty parameters are zero.","marker":"McMahan et al. [2017]"},{"why":"Proposes the loss-variance penalty that FairLoss and FairLoss* implement, giving the paper its direct predecessor method.","marker":"Ling et al. [2024]"},{"why":"AAggFF, the fairness-aware aggregation baseline used in the experiments; the paper compares its server-side aggregation against the local regularizers.","marker":"[Hahn et al., 2024]"},{"why":"Supplies the Dirichlet-based data partitioning procedure used to create the heterogeneous client data splits in all experiments.","marker":"[Tam et al., 2023]"},{"why":"Earlier use of gradient variance regularization in FL (for out-of-distribution generalization), which the paper distinguishes from its new fairness-oriented application.","marker":"Guo et al. [2023]"},{"why":"Documents the performance disparities caused by non-IID client data, the problem the paper's fairness methods target.","marker":"[Zhao et al., 2018]"},{"why":"Provides the fairness definition that Definition 3.1 modifies, anchoring the variance-based fairness criterion.","marker":"Nguyen et al. [2024]"}],"fun_headline_variants":["Gradient variance penalty boosts fairness in federated learning","FairGrad reduces client accuracy gaps in FL","Variance-regularized gradients make FL fairer","Exact gradient variance method shrinks client disparities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof that the regularizers are inert when data are identical assumes a client's gradient is unrelated to how far that client's loss sits from the group average; identical distributions alone do not guarantee this, and if it fails the homogeneous-case theory no longer follows.","fun_headline_variants_meta":{"raw":{"variants":["Gradient variance penalty boosts fairness in federated learning","FairGrad reduces client accuracy gaps in FL","Variance-regularized gradients make FL fairer","Exact gradient variance method shrinks client disparities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000453,"raw_usage":{"total_tokens":2278,"prompt_tokens":946,"completion_tokens":1332,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1271}},"tokens_in":562,"tokens_out":1332,"duration_ms":10956,"temperature":1.0,"reasoning_tokens":1271,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:26:26.794800+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a dataset with provably identical client distributions, train FedAvg to a stationary point $x^\\star$, then estimate $\\mathbb{E}[\\nabla f_i(x^\\star)\\mid f_i(x^\\star)-F(x^\\star)]$ from client samples; if the estimate is significantly nonzero, the homogeneous-case proof of Theorem 4.1 collapses, and the claim that the regularizers are unnecessary when data are identical loses its theoretical footing.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"Introduces FedAvg, the baseline objective $F(x)$ that all regularized methods reduce to when their penalty parameters are zero."},{"cited_title":"FL-bench : A federated learning benchmark repository, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the Dirichlet-based data partitioning procedure used to create the heterogeneous client data splits in all experiments."},{"cited_title":"Out-of-distribution generalization of federated learning via implicit invariant relationships","cited_arxiv_id":null,"evidence_quote":"Earlier use of gradient variance regularization in FL (for out-of-distribution generalization), which the paper distinguishes from its new fairness-oriented application."},{"cited_title":"Federated Learning: Theory and Practice","cited_arxiv_id":null,"evidence_quote":"Provides the fairness definition that Definition 3.1 modifies, anchoring the variance-based fairness criterion."}],"review_version":2}