{"id":"90c2fffb-f22e-494f-be9c-21fb35187cd8","arxiv_id":"2504.17528","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"TACO assigns client-specific correction coefficients based on gradient magnitude and direction to reduce over-correction in non-IID federated learning, improving accuracy, rounds, and wall-clock time.","lead":"Federated learning struggles when different devices have very different data. This paper proposes TACO, which gives each device a correction strength matched to its own gradient, and reports faster, more accurate training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TACO's core proof is invalid: Assumption 2 cannot yield E||Delta_i|| <= (mu_i/c_i)||grad f||, and Corollary 2's proportionality does not minimize the displayed Y_t, so the claimed over-correction analysis and alpha_i optimality are unsupported.","rationale":"TACO's empirical study is substantial: eight datasets, multiple baselines, ablations, scalability, and time-to-accuracy measurements. Those results may well support TACO as a practical heuristic. The advertised identity of the paper, however, is the 'first convergence analysis that reveals the root cause of over-correction' and the theoretical optimality of the tailored coefficients. That identity requires Theorem 1 and Corollary 2 to be valid. The appendix's key bound is a Jensen violation; Corollary 2's claimed minimization is inconsistent with the product form of Y_t; and Lemma 1 is not the recursion implied by Eq. (9). Each of these is internal to the proof and independent of empirical success or of any field consensus. Consequently, the central theoretical claim is unsupported, and the paper as written overstates its contribution. This does not impugn the authors or the experiments; it means the theoretical pillar, which is a major advertised contribution, needs repair before the convergence and optimality claims can be accepted. The reader's REJECT verdict is therefore appropriate; a corrected proof and released code could change that assessment.","tokens_in":19243,"tokens_out":9684,"duration_ms":86138,"concrete_test":"Perform the following analytic re-derivation. First, for any client with Delta_i = +/-M with probability 1/2, E[Delta_i] = 0, so Assumption 2 holds trivially while E||Delta_i|| = M; this directly falsifies the appendix inequality for any M > 0. Second, recompute the minimization in Corollary 2 from the displayed Y_t: with x_i = (1-alpha_i^t), y_i = mu_i/c_i, and X = Sum x_i, Y = Sum y_i, Y_t = C X^2 Y^2, whose partial derivative with respect to x_i is 2C X Y^2 for every i; hence any x with the same sum gives the same Y_t, and no allocation proportional to y_i is singled out. If both checks reproduce these conclusions, Theorem 1 and Corollary 2 do not justify TACO's design.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is in the Appendix proof of Theorem 1. After bounding the correction term, the proof states 'Using Eq. (11)/Eq. (12), we have E||Delta_i^{t-1}|| <= mu_i/c_i ||grad f(w^{t-1})||.' This does not follow. Assumption 2 bounds the cosine similarity between grad f and the expectation E[Delta_i]; it gives ||E[Delta_i]|| <= (mu_i/c_i)||grad f||, not the expectation of the norm. By Jensen's inequality, E||Delta_i|| >= ||E[Delta_i]|| and can be much larger, so the Y_t term that quantifies over-correction is not established. Independently, even accepting Y_t, Corollary 2 is arithmetically false: Y_t = C (Sum_i (1-alpha_i^t))^2 (Sum_i mu_i/c_i)^2. Since the second factor is independent of the correction coefficients, minimizing over (1-alpha_i) with the stated constraint Sum_i (1-alpha_i^t) >= sigma only forces the sum to sigma; it imposes no per-client proportionality (1-alpha_i) proportional to mu_i/c_i. Thus TACO's tailored alpha_i is not shown optimal by the paper's own bound. A further inconsistency is that Lemma 1 (Delta^{t+1} = tildeDelta^t + (1-alpha^t)Delta^t) does not follow from the weighted aggregation rule in Eq. (9), which multiplies each Delta_i^t by alpha_i^{t+1} before summing; the lemma uses only the average alpha^t.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TACO, a federated learning algorithm that assigns per-client, time-varying correction coefficients α_i^t, used both in local updates and in server-side aggregation. The authors claim to identify a previously hidden 'over-correction' phenomenon caused by uniform correction coefficients in existing FL methods, and they support this with a convergence analysis (Theorem 1 and Corollaries 1–2) as well as experiments across eight datasets and several baselines. The central theoretical claim is that the error term Y_t in the bound is minimized when (1−α_i^t) ∝ μ_i/c_i, which motivates the particular form of α_i^t in Eq. (7).","tokens_in":19693,"tokens_out":6557,"duration_ms":58242,"significance":"If the theoretical results were sound, TACO would be a useful lightweight contribution to non-IID federated learning: it requires no auxiliary parameters, adds little per-round computation, and the time-to-accuracy comparison is a valuable evaluation lens. The experimental section is broad, comparing six baselines across eight datasets and including ablation, scalability, sensitivity, and freeloader-detection studies. However, the load-bearing theoretical arguments are not valid: a key inequality in the appendix is unjustified, the main recursion in Lemma 1 is inconsistent with the aggregation rule, and the optimality statement in Corollary 2 does not follow from the displayed bound. The empirical breadth does not compensate for the unsupported central claims, since the paper's novelty is advertised as the first convergence analysis that reveals the root cause of over-correction.","major_comments":[{"comment":"The step 'Using Eq. (11)/Eq. (12), we have E||Δ_i^{t−1}|| ≤ μ_i/c_i ||∇f(w^{t−1})||' is invalid. Assumption 2 bounds the inner product and cosine similarity involving the expectation E[Δ_i^{t−1}]; it yields a bound on ||E[Δ_i^{t−1}]||, not on E||Δ_i^{t−1}||. Jensen's inequality gives the reverse direction, E||Δ_i^{t−1}|| ≥ ||E[Δ_i^{t−1}]||, and the expectation of the norm can be much larger. This step is load-bearing because it is the point where the term Y_t is introduced in Theorem 1, and Corollary 1 inherits this error.","section":"Appendix, proof of Theorem 1"},{"comment":"The claimed optimization is not correct. In the expression Y_t = (L^2 G^2)/(K^2 N^4 η_l^2) (Σ_i (1−α_i^t))^2 (Σ_i μ_i/c_i)^2, the second factor is independent of the correction coefficients. Minimizing over (1−α_i^t) subject to Σ_i (1−α_i^t) ≥ σ only forces the sum to σ; any allocation with that sum gives the same Y_t. Thus the statement that (1−α_i^t) ∝ μ_i/c_i is not derived from the bound, and the claimed optimality of the tailored coefficients in Eq. (7) is unsupported.","section":"Corollary 2"},{"comment":"Lemma 1 (Δ^{t+1} = tildeΔ^t + (1−α^t)Δ^t) does not follow from the aggregation rule in Eq. (9) when the coefficients α_i^{t+1} differ across clients. Expanding Δ_i^t from Eq. (8) with γ = 1 gives terms of the form α_i^{t+1}(1−α_i^t)Δ^t inside the summation, and the outer factor (Σ_i α_i^{t+1})/N in Eq. (9) does not reduce this to the average coefficient α^t = (1/N)Σ_i α_i^t. Consequently Lemma 2 and the subsequent bound based on z^{t+1} = z^t − η_g tildeΔ^t are not justified.","section":"Lemma 1 and Eq. (9)"}],"minor_comments":[{"comment":"The claimed improvement range '2.76%–58.68%' is not clearly defined; please specify the comparison basis (final accuracy after fixed rounds, or rounds to target accuracy).","section":"Section V-B"},{"comment":"The x-axis labels in Fig. 5(a) and 5(b) are compressed ('FedAvg/FGFedProxScaffold'), making the figure difficult to read.","section":"Fig. 5"},{"comment":"The timing of the correction coefficients should be stated explicitly: Eq. (7) computes α_i^t using Δ_i^{t−1} and the previous round's average gradient, while Eq. (9) uses α_i^{t+1} computed after receiving Δ_i^t; the relationship to the local update in Eq. (8) is not immediately transparent.","section":"Section IV-A"},{"comment":"The proof sets γ = 1 'for brevity', but γ is a tunable hyperparameter in the experiments; a sentence explaining why the analysis extends to general γ would help.","section":"Appendix, proof of Theorem 1"}],"recommendation":"reject","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"TACO is two papers in one. The empirical half is solid and worth taking seriously: per-client correction coefficients from gradient norm and cosine similarity, tailored aggregation, and a freeloader filter are all sensible, and the experiments span eight datasets, several baselines, and both round-to-accuracy and time-to-accuracy. The ablation study shows the correction term is the main driver, and the time-to-accuracy comparison makes a fair point that STEM-type methods buy rounds at the cost of client compute. That part deserves a serious look.\n\nThe theoretical half does not hold up. The advertised 'first convergence analysis revealing the root cause of over-correction' is built on lemmas that contradict the algorithm. Eq. (9) uses client-specific alpha_i^{t+1} in the aggregation, so Delta^{t+1} is not generally equal to tildeDelta^t + (1-alpha^t)Delta^t with an average alpha. Lemma 2 then discards both the correction and the tailored aggregation, so the proof is for a different recursion. The appendix's key step E||Delta_i^{t-1}|| <= (mu_i/c_i)||grad f|| does not follow from Assumption 2, which bounds only ||E[Delta_i]||; Jensen's inequality goes the other way. And even if Y_t were as stated, Corollary 2 is arithmetic: Y_t = C( sum_i (1-alpha_i) )^2 ( sum_i mu_i/c_i )^2, so the constraint sum_i (1-alpha_i) >= sigma only fixes the first factor; it says nothing about proportionality to mu_i/c_i. The optimality claim is not derived.\n\nThe over-correction framing is useful as a label, and the paper correctly credits [11] for prior evidence of instability, but the theory does not establish that uniform coefficients cause the failure or that TACO's alpha_i is optimal. Since the theory is a major advertised contribution, the paper in its current form should be rejected. The empirical method could be resurrected with a corrected proof and released code, and I would send it to peer review because the empirical study is substantive and the design question is practical. Any referee should be asked to check the appendix line by line before engaging with the convergence claims.","headline":"Solid empirical study of a lightweight adaptive-correction FL method, but the advertised convergence analysis is invalid and the optimality corollary is arithmetically wrong.","tokens_in":20158,"tokens_out":4018,"would_cite":false,"duration_ms":36071,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TACO claims that per-client correction coefficients eliminate the over-correction caused by uniform coefficients in federated learning, and supports this with a convergence analysis and experiments on eight datasets.","keywords":["federated learning","non-IID data","over-correction","client drift","client-specific correction","convergence analysis","time-to-accuracy","freeloader detection"],"falsifier":"Check whether $E\\|\\Delta_i^{t-1}\\| \\le (\\mu_i/c_i)\\|\\nabla f(w^{t-1})\\|$ holds on a two-client example that satisfies Assumption 2: take both client directions close to the global gradient but give client 1 a much larger local-gradient magnitude than client 2; if the bound fails while the stated cosine-similarity assumption holds, the $Y_t$ term that Corollary 2 minimizes is an artifact of the proof rather than a real over-correction effect. An experimental counterpart is to compare TACO with FedProx whose $\\zeta_i$ is tuned per client; matching accuracy would indicate the uniform coefficient, not the correction direction, is what matters.","tokens_in":19045,"feed_emoji":"🔄","tokens_out":9903,"duration_ms":85274,"temperature":0.7,"pith_summary":"This paper tries to establish that a hidden 'over-correction' effect, not the absence of correction, explains why many federated-learning fixes for non-IID data (FedProx, Scaffold, STEM, FedACG) often do worse than plain FedAvg. The proposed remedy, TACO, gives each client its own correction coefficient, computed from the size and direction of that client's local gradient relative to the global gradient, and uses the same coefficient as an aggregation weight. Because clients with more skewed local data are corrected harder and weighted less, the paper argues, the over-correction introduced by uniform coefficients disappears. This would matter because non-IID data is the standard obstacle in federated learning, and the paper's design adds little client-side computation, so better accuracy could come without sacrificing training speed.","feed_headline":"Per-client corrections beat uniform ones in federated learning","feed_subtitle":"TACO sets each client's correction by gradient size and direction, cutting over-correction and wall-clock time.","key_machinery":"The central object is the client-specific correction coefficient $\\alpha_i^t$ in Eq. (7), built from the magnitude of the client's accumulated local gradient $\\|\\Delta_i^{t-1}\\|$ divided by the sum of all clients' magnitudes, multiplied by the clipped cosine similarity between $\\Delta_i^{t-1}$ and the average global gradient. It enters the local update $w_{i,k+1}^t = w_{i,k}^t - \\eta_l(g_{i,k}^t + \\gamma(1-\\alpha_i^t)\\Delta^t)$ and the server aggregation $\\Delta^{t+1} = \\frac{1}{K\\eta_l}\\frac{\\sum_{i\\in\\mathcal{N}}\\alpha_i^{t+1}\\Delta_i^t}{\\sum_{i\\in\\mathcal{N}}\\alpha_i^{t+1}}$. The same coefficient thus simultaneously decides how strongly a client's local update is pulled toward the global model and how much weight that client's gradient gets in aggregation, which is the mechanism that turns the paper's over-correction diagnosis into a trainable algorithm.","core_discovery":"On the paper's own terms, the central discovery is that the error bound for federated optimization contains a correction-induced term $Y_t = \\frac{L^2G^2}{K^2N^4\\eta_l^2}\\big(\\sum_{i\\in\\mathcal{N}}(1-\\alpha_i^t)\\sum_{i\\in\\mathcal{N}}\\frac{\\mu_i}{c_i}\\big)^2$ that grows when all clients share one correction coefficient. TACO chooses $\\alpha_i^t$ from $\\|\\Delta_i^{t-1}\\|$ and its cosine similarity to the global gradient, and Corollary 2 states that the optimal choice sets $(1-\\alpha_i^t)$ proportional to $\\mu_i/c_i$, which minimizes $Y_t$. The paper claims this is the first convergence analysis to pinpoint uniform correction coefficients as the root cause of over-correction, and it reports that the same $\\alpha_i^t$ doubles as a freeloader detector because clients that merely re-upload the previous global gradient receive unusually high $\\alpha$ values. Empirically, TACO is reported to exceed all six baselines on eight datasets in round-to-accuracy and to reach target accuracy with 25.6–62.7% less client computation time than FedAvg.","pith_inferences":["Beyond the paper's own claims, a testable extension is that other drift-correction families (FedDyn, FedDC, FedNova) should also improve when their uniform regularizer or weighting is replaced by a per-client $\\alpha$-style term, since the paper only demonstrates the swap for FedProx and Scaffold.","The paper's time-to-accuracy re-evaluation predicts that communication-round comparisons favor algorithms with heavy local computation; an independent benchmark reporting both metrics could reveal how much of the reported TACO advantage is speed rather than accuracy.","If the over-correction diagnosis is right, tuning a uniform coefficient more carefully does not remove the problem; it only picks which clients to overcorrect. A direct check would compare TACO against FedProx with per-client $\\zeta_i$ chosen by grid search, since matching performance would suggest the uniform-coefficient story is not the whole explanation.","The freeloader test relies on lazy clients behaving predictably (re-uploading the previous global gradient), so a plausible evasion is to copy a randomly selected honest client's gradient; whether TACO's $\\alpha$ signature still exposes that variant is an open empirical question."],"forward_implications":["Uniform-coefficient correction methods can be read as a single flawed family: FedProx, Scaffold, STEM, and FedACG all enlarge the error term $Y_t$ when their shared coefficient does not match each client's skew, which the paper argues explains their instability and convergence failures on SVHN and CIFAR-10.","Replacing a baseline's uniform coefficient with TACO's tailored coefficients yields the paper's reported gains (FedProx on SVHN and Scaffold on CIFAR-10), suggesting the correction coefficient, not the surrounding method, carries much of the non-IID benefit.","TACO reaches target accuracy in fewer communication rounds than every baseline on the tested datasets, and cuts client-side computation time by 25.6–62.7% relative to FedAvg.","The convergence rate becomes $O\\big(\\sqrt{L/T} + \\sqrt[3]{Y/T^2}\\big)$, and when all $\\alpha_i^t$ are equal it reduces to the existing FedCM rate, so the analysis nests prior uniform-coefficient results as a special case.","Freeloaders that re-upload the previous global gradient receive high $\\alpha_i^t$ and can be expelled with threshold $\\kappa$ over $\\lambda$ rounds, with the reported operating region $\\kappa\\in[0.6,0.8]$, $\\lambda=T/5$ giving 100% true positives and 0% false positives on FMNIST."],"supporting_citations":[{"why":"Provides the vanilla FedAvg baseline, the aggregation rule, and the comparison target for all other methods.","marker":"[2]"},{"why":"FedProx is the main uniform-regularization baseline whose over-correction the paper diagnoses and then fixes by inserting TACO-style coefficients.","marker":"[6]"},{"why":"FoolsGold supplies the aggregation-calibration baseline with no local correction, used to separate aggregation effects from correction effects.","marker":"[7]"},{"why":"Scaffold is the uniform-momentum baseline whose convergence failures are attributed to over-correction and whose coefficient swap is tested in Fig. 6.","marker":"[8]"},{"why":"STEM is the two-sided momentum baseline whose uniform $\\alpha^t$ exemplifies over-correction and whose extra client computation motivates TACO's lightweight design.","marker":"[9]"},{"why":"FedACG is the latest state-of-the-art baseline that TACO must beat, combining regularization and momentum with uniform coefficients.","marker":"[10]"},{"why":"The experimental study reporting instability and mediocre performance of these algorithms supplies the empirical motivation for re-evaluating them.","marker":"[11]"},{"why":"Provides the client-level momentum convergence lemma and rate template that Corollary 1 adapts to tailored coefficients.","marker":"[19]"}],"fun_headline_variants":["Tailored corrections fix federated learning's over-correction","Per-client tuning beats blanket fixes in federated learning","TACO: stopping over-correction with bespoke client updates","Over-correction in FL? Tailor each client's gradient fix","TACO cuts over-correction, saves up to 62.7% client time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof's main conclusion depends on assuming that the expected size of a client's local gradient is no more than a fixed multiple of the size of the true global gradient, yet the assumptions stated in the paper only constrain the direction of the expected gradient, not its size; the over-correction term in the error bound comes from exactly this unsupported step.","fun_headline_variants_meta":{"raw":{"variants":["Tailored corrections fix federated learning's over-correction","Per-client tuning beats blanket fixes in federated learning","TACO: stopping over-correction with bespoke client updates","Over-correction in FL? Tailor each client's gradient fix","TACO cuts over-correction, saves up to 62.7% client time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000765,"raw_usage":{"total_tokens":3439,"prompt_tokens":1035,"completion_tokens":2404,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":651,"completion_tokens_details":{"reasoning_tokens":2323}},"tokens_in":651,"tokens_out":2404,"duration_ms":17694,"temperature":1.0,"reasoning_tokens":2323,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:38:26.007870+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check whether $E\\|\\Delta_i^{t-1}\\| \\le (\\mu_i/c_i)\\|\\nabla f(w^{t-1})\\|$ holds on a two-client example that satisfies Assumption 2: take both client directions close to the global gradient but give client 1 a much larger local-gradient magnitude than client 2; if the bound fails while the stated cosine-similarity assumption holds, the $Y_t$ term that Corollary 2 minimizes is an artifact of the proof rather than a real over-correction effect. An experimental counterpart is to compare TACO with FedProx whose $\\zeta_i$ is tuned per client; matching accuracy would indicate the uniform coefficient, not the correction direction, is what matters.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Provides the vanilla FedAvg baseline, the aggregation rule, and the comparison target for all other methods."},{"cited_title":"Stem: A stochastic two-sided momentum algorithm achieving near-optimal sample and communication complexities for federated learning,","cited_arxiv_id":null,"evidence_quote":"STEM is the two-sided momentum baseline whose uniform $\\alpha^t$ exemplifies over-correction and whose extra client computation motivates TACO's lightweight design."},{"cited_title":"Communication-efficient federated learn- ing with accelerated client gradient,","cited_arxiv_id":null,"evidence_quote":"FedACG is the latest state-of-the-art baseline that TACO must beat, combining regularization and momentum with uniform coefficients."}],"review_version":1}