{"id":"511bcd70-91d5-4518-8008-9db401f981f4","arxiv_id":"2501.02662","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"FLamma claims to balance fairness and accuracy in federated learning via a Stackelberg game with an adaptive decay factor, but the theory has derivation errors and the experiments use fixed local epochs.","lead":"This paper proposes FLamma, a federated learning framework where a server adjusts a decay factor and clients choose local training epochs through a Stackelberg game, aiming to improve fairness in non-IID settings. The simulations show lower accuracy variance and higher accuracy than baselines, but the game-theoretic mechanism is not actually tested in the experiments.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Server utility derivation drops the sum over clients, so the claimed Stackelberg equilibrium gamma* is algebraically invalid.","rationale":"The reader's weakest_assumption focuses on omega_i possibly becoming negative, which is a real concern for the client best response, but the server utility derivation is more fundamental: even with all omega_i > 0, the paper's gamma^* is not the Stackelberg leader's best response, so the equilibrium claim fails across all regimes. The reader's rationale does mention the dropped sum, so my read partially agrees with the reader. The paper has no machine-checked proof or released code; the only independent support would be the empirical tables, but those do not implement the optimal epoch selection or the derived gamma^*, so they cannot rescue the mechanism claim. I would keep the rejection: the central theoretical contribution is invalid as written, and a major rewrite plus mechanism-faithful experiments would be needed.","tokens_in":15717,"tokens_out":4002,"duration_ms":37934,"concrete_test":"Re-derive the server's first-order condition by substituting tau_i^* = gamma * omega_i / (2 c_i) into the full sum in U_server(gamma, tau_i) and solving dU_server/dgamma = 0. Compare with the displayed gamma^* in Section IV.C. If the resulting expression contains sum_i omega_i and sum_i omega_i / c_i (it does) and the displayed formula does not, the proof of Lemma 3 fails. To make this numerical, instantiate N=2 with omega_1=0.8, omega_2=0.2, c_1=c_2=1, t=10; evaluate the displayed gamma^* and the correct gamma^*, and verify only the latter satisfies the stationary condition of the full utility.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing flaw is in Section IV.C. The server utility in Section IV.B is U_server(gamma, tau_i) = sum_{i=1}^N [ gamma * (omega_i + tau_i) ] - t * gamma^2, with omega_i = 1 - ||w_i^t - w^t||/||w^t||. Substituting the client best response tau_i^* = gamma * omega_i / (2 c_i) gives U_server = gamma * sum_i omega_i + (gamma^2 / 2) * sum_i omega_i / c_i - t * gamma^2. Setting dU_server/dgamma = 0 yields gamma^* = (sum_i omega_i) / (2t - sum_i omega_i / c_i). The paper instead claims gamma^* = omega_i * c_i / (2 t c_i - omega_i), using a single client and dropping both sums. That expression does not maximize the leader's actual utility except in the degenerate case of one homogeneous client. Consequently Lemma 3, the claimed Stackelberg equilibrium, and the incentive-compatible characterization in Section IV.C do not follow from the stated model. Algorithm 1 line 17 merely says 'Update the decay factor gamma' with no rule, and the experiments fix tau_i = 10 rather than tau_i^*, so no part of the empirical evaluation exercises the derived equilibrium. The central claim that FLamma provides an incentive-compatible mechanism therefore rests on an invalid derivation; the reported accuracy/variance gains may reflect a gamma-scaled FedAvg variant, but that is a different claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FLamma, a federated learning framework in which a server acts as a Stackelberg leader by setting a decay factor gamma, while clients act as followers who choose their number of local epochs tau_i to maximize a utility that combines contribution-dependent reward with a quadratic cost. The authors derive a closed-form client best response tau_i^* and a server-optimal gamma^*, state lemmas on concavity, Nash equilibrium, and individual rationality, and provide a FedAvg-style convergence bound. They evaluate FLamma on MNIST, FashionMNIST, and CIFAR-10 under IID and non-IID splits, comparing against FedAvg, FedProx, q-FFL, and an incentivization baseline, and report higher accuracy and substantially lower accuracy variance.","tokens_in":15959,"tokens_out":5683,"duration_ms":57294,"significance":"If the equilibrium analysis were correct, the paper would contribute a principled incentive mechanism for fairness in non-IID federated learning, with a falsifiable prediction about accuracy-variance reduction. The experimental results, taken at face value, show large variance reductions across three datasets. Credit is due for a clearly described experimental setup and for identifying limitations in Section VII. However, the significance is contingent on the game-theoretic derivation, and the central derivation does not survive scrutiny: the closed-form gamma^* is algebraically inconsistent with the stated server utility, the implemented algorithm does not specify the gamma update, and the experiments fix tau_i instead of using the derived equilibrium. The contribution is therefore not established in its current form.","major_comments":[{"comment":"The derivation of gamma^* drops the sum over clients. From the server utility in Section IV.B, substituting tau_i^* = gamma * omega_i / (2 c_i) gives U_server = gamma * sum_i omega_i + (gamma^2 / 2) * sum_i omega_i / c_i - t * gamma^2, whose maximizer is gamma^* = (sum_i omega_i) / (2t - sum_i omega_i / c_i), not the per-client expression gamma^* = omega_i * c_i / (2 t c_i - omega_i) displayed after the derivative. The displayed substitution also omits the summation sign. Consequently Lemma 3, the claimed Stackelberg equilibrium, and the incentive-compatible characterization in Section IV.C do not follow from the stated model except in the degenerate single-client case.","section":"Section IV.C, server utility"},{"comment":"The best-response formula tau_i^* = gamma * omega_i / (2 c_i) is obtained as an unconstrained interior maximum, but omega_i = 1 - ||w_i^t - w^t|| / ||w^t|| is not guaranteed to be nonnegative under non-IID data. When omega_i < 0, the candidate tau_i^* is negative and outside any reasonable strategy set, the client utility is decreasing for tau_i >= 0, and the optimal strategy is a boundary point, not the interior solution. The IR guarantee in Lemma 5 is then not automatic: with a positive lower bound tau_min, a client with omega_i < 0 can have negative utility at tau_min, and the proof's deviation to tau_i = 0 is unavailable if 0 is not in the strategy space. The manuscript provides no clipping, projection, or alternative treatment of negative contributions.","section":"Section IV.C, Eq. (1) and Lemma 5"},{"comment":"The convergence analysis is internally inconsistent with Algorithm 1. Line 12 of Algorithm 1 uses the effective learning rate eta * gamma, but the recurrence in the proof is written for a plain eta step, with gamma appearing only as a multiplier of the gradient-divergence term. Assumption 5 bounds E[sum_k p_k ||w_t - w_t^k||^2] by 4 eta^2 (tau_max - 1)^2 G^2 without the gamma^2 factor that a gamma-scaled local update would introduce, and the final bound in Eq. (2) contains terms (e.g., rho * xi * gamma_max / 2 * M and the expression for B) that do not follow from the proof's algebra, which ends with eta^3 terms. Therefore the claimed FedAvg-comparable convergence rate is not established by the supplied argument.","section":"Section V, Theorem and proof"},{"comment":"The experimental evaluation does not exercise the proposed equilibrium. Section VI.B states that clients train for ten local iterations and that the contribution is updated every ten global rounds, while Algorithm 1 line 17 says only 'Update the decay factor gamma' with no rule, and no experimental values for gamma or tau_i^* are reported. Thus Table I and Figs. 2-3 can at most demonstrate properties of a gamma-scaled FedAvg variant, not of the Stackelberg mechanism whose gamma^* is derived in Section IV.C. Section VII's admission that 'the optimal tuning of this decay factor requires non-negligible experimentation' further confirms that the closed-form equilibrium is not what is implemented and evaluated.","section":"Sections VI.B, VI.C, and Algorithm 1"}],"minor_comments":[{"comment":"The server utility formula is displayed with ambiguous parentheses, making it unclear whether tau_i is inside or outside the factor gamma * (...); omega_i is also introduced parenthetically after the summation, so it should be defined before the utility expression.","section":"Section IV.B"},{"comment":"The client utility is written as U_i(gamma, tau_i, tau_{-i}) but contains no dependence on tau_{-i}; either remove the redundant argument or explain how other clients' strategies enter the utility.","section":"Eq. (1)"},{"comment":"The text says the contribution is updated every 10 global rounds, while Algorithm 1 recomputes contributions every round (lines 3-4); this discrepancy should be reconciled.","section":"Section VI.B vs. Algorithm 1"},{"comment":"Assumption 3 appears to be stated but is not used in the proof; either use it explicitly or remove it to avoid confusing the reader about which assumptions are load-bearing.","section":"Section V"},{"comment":"The experimental description reports only that each experiment was run 3 times and does not give hyperparameters such as batch size, learning rate schedule, local data partitioning details, or the gamma schedule used, which are necessary for reproducibility; the Table I caption also contains a grammatical typo ('across CIFAR10, and FMNIST, and MNIST').","section":"Section VI.A and Table I"}],"recommendation":"reject","confidential_remarks":"The core theoretical claim is not salvageable by local edits: the gamma^* derivation must be redone, the gamma update rule must be specified, and the experiments must be run with the specified mechanism before the Stackelberg-equilibrium claim can be evaluated. If the reported accuracy and variance gains are robust, a substantially revised paper with a corrected game-theoretic analysis and reproducible code could be reconsidered. The paper also sits in a crowded Stackelberg-incentive FL literature, so the positioning would need to be sharpened after the derivation is fixed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper has a promising empirical recipe — a gamma-scaled FedAvg with contribution-based client selection — but the theory that is supposed to justify it does not hold together, and the experiments never actually run the mechanism they analyze.\n\nWhat's genuinely new is the specific combination: a Stackelberg game where the server commits to a decay factor gamma, clients respond with local epoch counts, and the client utility includes a contribution term based on the distance between the local and global models. The decay-factor idea is a reasonable way to balance early high-contributor influence with later fairness. The experiments are extensive: three datasets, IID and non-IID splits, four baselines, and the reported accuracy/variance gains are large. If those numbers hold up, the method is a simple and cheap fairness intervention worth knowing about.\n\nBut the central derivation is wrong. The server utility in Section IV.B is a sum over clients. When you substitute the client best response and differentiate with respect to gamma, you get a closed form involving the sums of omega_i and omega_i/c_i. The paper instead drops both sums and reports a formula that depends on a single client. That is not a minor slip: Lemma 3, the claimed Stackelberg equilibrium, and the incentive-compatibility story all ride on it. Also, the convergence proof in Section V uses the standard FedAvg eta in the recurrence, while Algorithm 1 scales the local update by eta*gamma; the two are not the same, so the theorem as stated does not match the algorithm. And the experiments fix tau_i = 10 local epochs for all clients — they don't let clients choose their best response at all. So the reported results are for a heuristic that is loosely inspired by the game, not for the mechanism the paper actually solves. The paper's own Limitations section concedes the decay factor is a tuned hyperparameter, which is honest but undercuts the equilibrium claim further.\n\nWho gets value from this? A reader working on incentive mechanisms for FL might mine the related work and the experimental setup. But the core contribution needs major revision: fix the gamma-star derivation, state the actual update rule, reconcile the convergence analysis with the algorithm, and run the game with client-chosen epochs. As it stands, it is a promising working-paper idea with a solid empirical pilot, not a theory-backed result.\n\nI would not accept it as is, but I would send it to serious peer review rather than desk reject. The flaws are fixable in principle, and the empirical direction is worthwhile. Worth a careful referee.","headline":"Nice empirical recipe for a gamma-scaled FedAvg, but the Stackelberg theory that is supposed to justify it collapses on inspection, and the experiments never test the mechanism they analyze.","tokens_in":16527,"tokens_out":1927,"would_cite":false,"duration_ms":19875,"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":"FLamma, an adaptive gamma-based Stackelberg game, claims to make federated learning fairer and more accurate by decaying dominant clients' influence while clients choose local epochs to maximize utility.","keywords":["federated learning","Stackelberg game","incentive mechanism","fairness","non-IID data","decay factor","convergence analysis","accuracy variance"],"falsifier":"Run FLamma on a non-IID split and measure $\\omega_i$ for every selected client each round. If any client has $\\|w_i^t - w^t\\| > \\|w^t\\|$, the paper's best-response derivation and IR proof do not apply, so the reported equilibrium story fails for that client. A direct check would report the fraction of selected clients with negative $\\omega_i$ during training.","tokens_in":15447,"feed_emoji":"⚖️","tokens_out":7617,"duration_ms":64163,"temperature":0.7,"pith_summary":"Federated learning breaks down when a few clients dominate the global model, especially when data is non-IID, because accuracy varies sharply across participants. FLamma attacks this by modeling the server–client interaction as a Stackelberg game: the server leads with a decay factor $\\gamma$ that shrinks the influence of high-contributing clients as training progresses, and each client follows by choosing the number of local epochs $\\tau_i$ that maximizes its own utility $U_i$. The paper claims this restores fairness—measured as lower variance of per-client accuracy—while keeping global accuracy and convergence at or above the level of FedAvg, FedProx, q-FFL, and a contribution-based incentive baseline, reporting gains such as a 120.93% accuracy improvement over q-FFL on non-IID CIFAR-10 and variance reductions of 85–99% relative to FedAvg. A convergence theorem (Eq. (2)) is provided to argue that the decay mechanism tightens the error bound instead of loosening it.","feed_headline":"Server-set decay rule lifts federated learning accuracy and fairness","feed_subtitle":"A server-set decay factor shrinks dominant clients' influence, cutting per-client accuracy variance by up to 99%.","key_machinery":"The load-bearing object is the Stackelberg game with decay factor $\\gamma$: the server leads by setting $\\gamma$, each client follower replies with the best response $\\tau_i^* = \\gamma \\omega_i / (2 c_i)$ that maximizes the concave utility $\\gamma \\omega_i \\tau_i - c_i \\tau_i^2$, and the server's utility, after substitution, is concave in $\\gamma$ with a closed-form maximizer. Repeated play drives the system toward the equilibrium, and the same $\\gamma$ scales the local gradient updates ($w_k^t \\leftarrow w_k^t - \\eta \\gamma \\nabla F_k(w_k^t)$), which is what lets the mechanism control both client influence and the divergence term in the convergence bound of Eq. (2). The contribution measure $\\omega_i$ is the quantity that connects geometric closeness of local and global models to an economic \"contribution.\"","core_discovery":"The paper's central claim is that fairness in federated learning can be engineered through the equilibrium of a two-level game rather than through a hand-designed reweighting rule. The server announces a decay factor $\\gamma \\in [0,1]$; client $i$ then chooses its local iteration count $\\tau_i$ to maximize $\\gamma\\,\\omega_i\\,\\tau_i - c_i \\tau_i^2$, where $\\omega_i = 1 - \\|w_i^t - w^t\\|/\\|w^t\\|$ measures how close the client's local model is to the global model. Concavity gives the unique best response $\\tau_i^* = \\gamma \\omega_i / (2 c_i)$, and substituting these responses into the server's utility yields an optimal $\\gamma^*$; iterating this play drives the system to a Stackelberg equilibrium in which each client is individually rational. As $\\gamma$ decays over rounds, initially dominant clients lose influence, the accuracy distribution across clients flattens, and the global model improves or holds steady. Experiments on MNIST, FashionMNIST, and CIFAR-10 under both IID and non-IID splits show FLamma with lower accuracy variance than FedAvg, FedProx, q-FFL, and the incentive baseline, alongside higher or comparable accuracy.","pith_inferences":["If the contribution measure is clipped or rectified to be non-negative, the same Stackelberg construction could be extended to other client actions, such as participation probability or local batch size, with the server's $\\gamma$ controlling the rate at which influence equalizes.","The reported gains may mix two effects: the game's epoch choice and the server's contribution-based client selection. An ablation that fixes client selection to uniform random while keeping the $\\gamma$ mechanism would isolate the fairness effect of the decay factor itself.","A negative-$\\omega_i$ case is not just a theoretical corner: it is likely to occur precisely in the non-IID regime the paper studies, so a practical deployment would need a projection or clipping rule to keep $\\tau_i^*$ in the feasible interval.","The variance metric is computed across clients' test accuracies on their own local distributions; under extreme label skew, equalizing this variance may trade against worst-case per-class accuracy, which the paper does not report."],"forward_implications":["FLamma's Stackelberg design gives each client a rational, self-interested reason to choose a finite number of local epochs, so fairness emerges from incentives rather than from a fixed aggregation rule.","The variance reductions on non-IID datasets (99.09% on FMNIST, 85.03% on CIFAR-10 relative to FedAvg) suggest that a single server-side decay parameter can substantially flatten the accuracy distribution across heterogeneous clients.","The convergence bound of Eq. (2) implies that the decay factor does not sacrifice the FedAvg-style convergence rate; at $\\gamma = 1$ the bound reduces to the standard setting, and smaller $\\gamma$ only shrinks the divergence term.","Because the method only adds a server-side $\\gamma$ update and a client-side epoch choice to standard FedAvg, it can be layered onto existing FL systems without changing the aggregation rule or communication protocol."],"supporting_citations":[{"why":"Defines the FedAvg aggregation and training loop that FLamma builds on and compares against.","marker":"[1]"},{"why":"Supplies the FedAvg convergence analysis template that the paper adapts to include the decay factor.","marker":"[46]"},{"why":"Introduces q-FFL, the fairness-oriented baseline whose accuracy and variance FLamma claims to outperform.","marker":"[43]"},{"why":"Introduces FedProx, the heterogeneity-focused baseline used as a comparison.","marker":"[52]"},{"why":"Describes the contribution-based incentive mechanism used as the incentive baseline.","marker":"[16]"},{"why":"Provides the fixed-point theorem used to argue existence of a Nash equilibrium in the client sub-game.","marker":"[45]"}],"fun_headline_variants":["Stackelberg game balances federated learning accuracy","Game-theoretic fairness boosts federated learning","Server decay rule evens out federated client accuracy","FLamma: game-driven fairness for federated learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All of the game-theoretic guarantees assume the contribution $\\omega_i = 1 - \\|w_i^t - w^t\\|/\\|w^t\\|$ is never negative; with the non-IID data the paper targets, a client's local model can easily be farther from the global model than the global model's own norm, making $\\omega_i < 0$, the best response $\\tau_i^*$ negative, and the individual-rationality guarantee in Lemma 5 void for that client.","fun_headline_variants_meta":{"raw":{"variants":["Stackelberg game balances federated learning accuracy","Game-theoretic fairness boosts federated learning","Server decay rule evens out federated client accuracy","FLamma: game-driven fairness for federated learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000673,"raw_usage":{"total_tokens":3101,"prompt_tokens":1020,"completion_tokens":2081,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":2021}},"tokens_in":636,"tokens_out":2081,"duration_ms":13739,"temperature":1.0,"reasoning_tokens":2021,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:07:58.118002+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FLamma on a non-IID split and measure $\\omega_i$ for every selected client each round. If any client has $\\|w_i^t - w^t\\| > \\|w^t\\|$, the paper's best-response derivation and IR proof do not apply, so the reported equilibrium story fails for that client. A direct check would report the fraction of selected clients with negative $\\omega_i$ during training.","supporting_citations":[{"cited_title":"Incentive mechanism for federated learning with random client selection,","cited_arxiv_id":null,"evidence_quote":"Describes the contribution-based incentive mechanism used as the incentive baseline."},{"cited_title":"Existence and uniqueness of equilibrium points for concave n-person games,","cited_arxiv_id":null,"evidence_quote":"Provides the fixed-point theorem used to argue existence of a Nash equilibrium in the client sub-game."}],"review_version":1}