{"id":"fad795fb-7401-4641-b9ca-eb5e2fcd4d73","arxiv_id":"2601.22274","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A server-side proximal anchor on FedAvg provably limits forgetting and gives a task-uniform convergence rate of O(sqrt(E/(NT))) for federated domain-incremental learning under partial participation.","lead":"SPECIAL, a new federated learning method, adds one server-side anchor to the standard FedAvg algorithm and is claimed to stop accuracy loss on earlier tasks when new domain-shift tasks arrive, while keeping communication costs unchanged. The paper also presents the first convergence-rate guarantee for federated domain-incremental learning with partial client participation, where only a subset of clients join each round.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary 1's O(sqrt(E/(NT))) rate is not implied by Theorem 2: its learning rates give a leading term of order 1/sqrt(N E T), and Eq. (9)'s residual contains a learning-rate-independent K B^2 floor.","rationale":"The reader's verdict of REJECT is justified, but the single most load-bearing concern is not the epsilon-alignment assumption named as the reader's weakest_assumption. That assumption is genuinely strong and empirically unverified, and the negative BWT values in Table 1 do suggest it may fail. However, the convergence-rate claim is more decisively invalidated by an internal algebraic inconsistency: Corollary 1's learning-rate choices do not produce the stated O(sqrt(E/(NT))) rate from Theorem 2, and Theorem 2's own residual Psi contains a learning-rate-independent K B^2 term that prevents any vanishing bound. This is an internal inconsistency, not a disagreement with consensus, and it does not depend on how the epsilon-alignment condition behaves in practice. The reader's rationale does mention the Corollary 1 step-size problem, so I mark partial agreement: same final verdict and partially overlapping reasoning, but the weakest_assumption field points elsewhere. I would keep the verdict at REJECT; the theoretical headline claims are unsupported even before considering the empirical plausibility of the BKT assumption.","tokens_in":35304,"tokens_out":9652,"duration_ms":90545,"concrete_test":"Recompute Corollary 1 from Eq. (9) by symbolic substitution: plug gamma_L = lambda/(sqrt(K T) E L) and gamma_G = sqrt(N E)/((K-1) lambda L) into both the leading coefficient and every term of Psi, then check the three step-size inequalities of Theorem 2 on the paper's own benchmark configuration (e.g., K=4, N=4, E=5, lambda=0.25, L=1, T=20). If the resulting leading term is O(sqrt(K)/sqrt(N E T)) rather than O(sqrt(E/(N T))) and Psi >= K B^2, the claimed rate fails. This check requires no new experiments, only algebra.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Corollary 1 is unsupported by the paper's own Theorem 2. Substituting gamma_L = lambda/(sqrt(K T) E L) and gamma_G = sqrt(N E)/((K-1) lambda L) into the denominator 2(1+lambda)(1-1/K) E gamma_G gamma_L T of Eq. (9) gives gamma_G gamma_L E T = 2(1+lambda)(1-1/K) sqrt(N E T)/(sqrt(K) (K-1) L^2), so the vanishing term is O(sqrt(K)/sqrt(N E T)) with lambda, L, K fixed, not O(sqrt(E)/(sqrt(N) sqrt(T))); the E-dependence is inverted. The same substitution also violates Theorem 2's step-size constraints for small T: gamma_G <= 1/(K-1) requires sqrt(N E) <= lambda L, and gamma_G gamma_L <= (1+lambda)/(3 E L) requires 3 sqrt(N E)/(sqrt(K T) (K-1) L) <= 1+lambda, which fails for small T. More fundamentally, Psi in Eq. (9) contains the additive term K B^2 with no factor of gamma_L or gamma_G, so the stationarity bound cannot decrease below roughly 2 K B^2/(1-1/K) as T grows. Any rate that vanishes with T, in particular O(sqrt(E/(N T))), is therefore inconsistent with the theorem's own bound. The claimed task-uniform convergence rate is an algebraic artifact of Corollary 1, independent of any empirical judgment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies federated domain-incremental learning (FDIL) under partial participation and proposes SPECIAL, a server-side proximal anchor added to FedAvg. Two theoretical contributions are claimed: Theorem 1 gives a backward-knowledge-transfer bound on the increase of prior-task loss, and Theorem 2 together with Corollary 1 claims the first task-uniform non-convex convergence rate O(sqrt(E/(NT))) for FDIL with partial participation. Experiments on Digit-10, VLCS, PACS, and DN4IL compare SPECIAL to memory-based and memory-free baselines in terms of ACC and BWT, with ablations over the proximal weight, communication rounds, local epochs, and heterogeneity level.","tokens_in":35593,"tokens_out":11214,"duration_ms":98304,"significance":"If the advertised rate were valid, SPECIAL would be a useful memory-free baseline with a rigorous task-uniform guarantee, and the BKT theorem would be a meaningful extension of centralized continual-learning theory. The paper is clearly written, the algorithm is simple, and the experiments are reasonably thorough. However, the central convergence claim does not follow from the paper's own analysis: the proposed step sizes produce the wrong dependence on E and T, the step-size constraints are violated, and the residual in Theorem 2 has a non-vanishing floor. The BKT result rests on a strong alignment condition that is unverified and appears inconsistent with the reported negative BWT on all four datasets. These are load-bearing issues for the main theoretical claims.","major_comments":[{"comment":"The advertised rate O(sqrt(E/(NT))) does not follow from Theorem 2. Substituting the stated gamma_L = lambda/(sqrt(KT)EL) and gamma_G = sqrt(NE)/((K-1)lambda L) into the denominator 2(1+lambda)(1-1/K)E gamma_G gamma_L T of Eq. (9) gives E gamma_G gamma_L T = sqrt(N E T)/(sqrt(K)(K-1)L^2), so the first term in Eq. (9) is O(sqrt(K)/sqrt(N E T)), with E in the denominator, not O(sqrt(E/(NT))). The step-size constraints of Theorem 2 are also violated by these choices: gamma_G <= 1/(K-1) requires sqrt(NE) <= lambda L, and gamma_G gamma_L <= (1+lambda)/(3EL) requires 3 sqrt(N) sqrt(E/T)/((K-1)L sqrt(K)) <= 1+lambda, which fails for small T. More fundamentally, the residual Psi in Eq. (9) contains the additive term (2/(1-1/K)) K B^2, with no factor of gamma_L, gamma_G, or 1/T. Consequently the right-hand side of Eq. (9) cannot go below this positive floor as T grows, so no vanishing rate, in particular no O(sqrt(E/(NT))) rate, can be concluded from Theorem 2. This is not a typo: the same per-round K B^2 residual is present in the summed bound in Appendix E.3, so the claimed Corollary 1 is an algebraic artifact of the present analysis.","section":"Section 4.3, Corollary 1 and Eq. (9)"},{"comment":"The epsilon-alignment condition is imposed for every communication round tau, every local epoch e, and every client m, requiring every new-task stochastic gradient to be positively correlated with the aggregate earlier-task gradient at the task-K start. This is a very strong per-sample assumption. It is not verified in the experimental section, and Table 1 reports negative BWT for SPECIAL on all four datasets, suggesting the condition does not hold in the benchmarks. As a result, Theorem 1 does not apply to the empirical evaluation, and the statement in Section 5 that the empirical BWT trends 'match' the theory is not justified. The paper should either verify the alignment condition, test a proxy for it, or state clearly that the BKT guarantee is conditional on an assumption that is not met in the reported experiments.","section":"Section 4.2, Theorem 1"}],"minor_comments":[{"comment":"In the description of Algorithm 1, 'stochastic-gradient decent' should read 'stochastic-gradient descent'.","section":"Section 3"},{"comment":"The word 'roud' should be 'round'.","section":"Assumption 5"},{"comment":"'Bellow' should be 'Below'.","section":"Section 4.4"},{"comment":"The caption says results are demonstrated on 'three datasets' but lists four datasets; it should say 'four datasets' or correct the list.","section":"Figure 3 caption"},{"comment":"The notation table defines T as 'Number of global epochs', but the paper uses T for communication rounds and E for local epochs; this should be corrected to avoid confusion.","section":"Table 3"},{"comment":"The main results paragraph says SPECIAL attains the highest ACC on 'all three datasets', but Table 1 reports four datasets; it should say 'all four datasets'.","section":"Section 5"},{"comment":"The denominator notation $(K-1)tEMNLB^2$ is ambiguous; parentheses should be added, e.g., $((K-1) t E M N L B^2)$, to clarify that all quantities multiply.","section":"Eq. (7)"}],"recommendation":"reject","confidential_remarks":"The advertised convergence rate is unsupported by the paper's own Theorem 2, and the residual floor in Eq. (9) means the rate cannot be rescued by a minor correction within the present framework. The BKT theorem additionally relies on an unverified and empirically implausible alignment condition. I recommend rejection: the central theoretical contributions are not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I just read arXiv:2601.22274. The punchline: the central theoretical claims don't survive contact with the paper's own equations. The algorithm—a server-side proximal anchor on FedAvg—is simple and the experiments are reasonable, but the advertised task-uniform convergence rate is an artifact of the corollary's algebra, and the BKT bound rests on an assumption the authors don't verify.\n\nWhat's genuinely new: the FDIL setting with partial participation is under-theorized, and the paper correctly identifies that existing FCL analyses focus on the last task. The proposed SPECIAL is a clean one-line modification to FedAvg, and the empirical study covers four domain-shift benchmarks with sensible baselines and ablations. If all you take away is the empirical claim that a server-side anchor beats a client-side proximal term, that's a useful data point.\n\nThe soft spots are serious. In Corollary 1, substituting the stated gamma_L and gamma_G into the vanishing term of Theorem 2 gives an O(1/sqrt(N E T)) term, not O(sqrt(E/(NT)))—the E-dependence is inverted. Worse, the chosen gamma_G violates Theorem 2's own condition gamma_G <= 1/(K-1) unless sqrt(N E) <= lambda L, which isn't true for the paper's typical hyperparameters. And the residual Psi in Eq. (9) contains a K B^2 term with no factor of gamma_L or gamma_G. That means Theorem 2's bound has a floor that does not go to zero as T grows. So even if the step-size algebra were fixed, the bound does not establish convergence to stationarity. The K B^2 floor is not a minor technical annoyance; it kills the claim that SPECIAL achieves a O(1/sqrt(T)) rate.\n\nThe BKT bound (Theorem 1) similarly sits on an epsilon-alignment condition that must hold every round, every local epoch, every client, with respect to the gradient of the prior-task sum evaluated at the task start. That's a strong, per-sample correlation requirement. The paper doesn't test it, and the uniformly negative BWT values in Table 1 suggest it doesn't hold in practice. The theorem may be true as a conditional statement, but its relevance to the experiments is unestablished.\n\nThe related work is cited appropriately, and I didn't see evidence of circular reasoning or self-citation padding. The writing is clear.\n\nWho is this for? Practitioners looking for a memory-free heuristic for federated domain shift might get something from the empirical part. Theorists shouldn't take the rates at face value. I'd send it to peer review—the problem is important enough and the empirical method is worth a serious look—but I'd expect the reviewers to require a corrected theory or a careful retraction of the rate claims. As it stands, I would not cite the theoretical results.","headline":"The headline rate doesn't follow from the paper's own theorem and the BKT bound rests on an unverified strong assumption, but the algorithm and experiments are worth a look.","tokens_in":36118,"tokens_out":6319,"would_cite":false,"duration_ms":53673,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SPECIAL's one server-side proximal anchor guarantees backward transfer and a task-uniform convergence rate of $O(\\sqrt{E/(NT)})$ for federated domain-incremental learning under partial participation.","keywords":["federated domain-incremental learning","server-side proximal anchor","backward knowledge transfer","task-uniform convergence","partial participation","non-convex optimization","memory-free continual learning","FedAvg"],"falsifier":"On a synthetic federated domain-incremental problem with two tasks whose gradients are orthogonal, the $\\epsilon$-alignment premise of Theorem 1 is violated by construction; if SPECIAL still shows the earlier-task loss decreasing beyond what Eq. (7) allows, or if the loss increase exceeds the right-hand side, then Eq. (7) is not a valid certificate in that regime. Concretely, track the cosine similarity between $\\nabla f_{1:K-1}(\\theta^0_K)$ and each local stochastic gradient and check the inequality round by round.","tokens_in":35052,"feed_emoji":"🔄","tokens_out":10720,"duration_ms":86383,"temperature":0.7,"pith_summary":"This paper introduces SPECIAL, a federated continual learning algorithm for the setting where tasks are distinct input domains with the same label space and only some clients participate each round. Its central thesis is that a single server-side proximal anchor—a weighted average of the round's aggregated update and the previous global model—is enough to guarantee two properties that had been missing for this setting. First, training on a new task cannot increase the loss on any earlier task beyond a drift-controlled term that shrinks as rounds, local epochs, and per-round participation grow (Theorem 1). Second, the algorithm converges on the cumulative objective across all tasks at rate $O(\\sqrt{E/(NT)})$ under partial participation, matching single-task FedAvg in communication efficiency (Theorem 2 and Corollary 1). If these guarantees hold, memory-free federated continual learning can be deployed on drifting domains without replay buffers or task-specific heads.","feed_headline":"Server anchor gives first all-task rate for federated domain shifts","feed_subtitle":"One quadratic blend toward the previous model protects old domains while matching FedAvg's speed.","key_machinery":"The central mechanism is the server-side proximal anchor: after averaging the $N$ participating clients' update vectors, the server solves the quadratic problem $\\min_u (\\|u-\\bar\\theta^{t+1}_i\\|^2 + \\lambda\\|u-\\theta_{i-1}\\|^2)$, whose closed form is $\\theta^{t+1}_i = (\\bar\\theta^{t+1}_i + \\lambda\\theta_{i-1})/(1+\\lambda)$. This one-step blend is the entire difference from FedAvg. It converts the previous global model into a compact memory that stabilizes the task-initial point, giving a uniform within-task drift bound $\\|\\theta^t_i-\\theta^0_i\\|^2 \\le \\gamma_G^2\\gamma_L^2E^2B^2/\\lambda^2$ that is independent of round and task. That drift control is what lets the proof transfer single-task FedAvg-style convergence to the cumulative objective and keeps the backward-transfer correction shrinking with $t$, $E$, and $N$.","core_discovery":"The paper's discovery is that the placement of a proximal regularizer—on the server rather than on each client—turns a one-line modification of FedAvg into a principled federated domain-incremental learning method. At the end of each round the server replaces the plain aggregate $\\bar\\theta^{t+1}_i$ with $\\theta^{t+1}_i = (\\bar\\theta^{t+1}_i + \\lambda\\theta_{i-1})/(1+\\lambda)$, a quadratic blend toward the final model of the previous task. With this anchor, Theorem 1 bounds the expected loss on earlier tasks by their starting loss plus a vanishing correction of order $2\\epsilon^2\\sigma_L^2\\|\\nabla f_{1:K-1}(\\theta^0_K)\\|^2 / ((K-1)tEMNLB^2)$ under an $\\epsilon$-alignment condition, and Theorem 2 gives a task-uniform non-convex stationarity bound whose best achievable rate is $O(\\sqrt{E/(NT)})$. The rate explicitly separates stochastic and client variance ($\\sigma_L$, $\\sigma_G$) from inter-task drift ($\\sigma_T$).","pith_inferences":["Editorial inference: the update rule $\\theta^{t+1}_i = (1/(1+\\lambda))\\bar\\theta^{t+1}_i + (\\lambda/(1+\\lambda))\\theta_{i-1}$ is an exponential moving average across tasks, so SPECIAL can be viewed as continuously decaying old-task knowledge at a fixed rate; this suggests a testable rule for setting $\\lambda$ from the measured cosine alignment between tasks—larger alignment, smaller $\\lambda$, and","Editorial inference: Theorem 1's $\\epsilon$-alignment condition can be checked at runtime by storing only the vector $\\nabla f_{1:K-1}(\\theta^0_K)$ and computing cosine similarities against incoming local gradients; because the paper reports negative backward transfer on all four datasets, a direct measurement would either confirm that the condition fails in practice or reveal that the bound is lo","Editorial inference: the momentum-like interpretation suggests a natural extension to federated class-incremental learning, where the anchor would need to separate representation drift from classifier drift; the same proof structure might apply if the inter-task drift constant $\\sigma_T$ is replaced by a class-specific drift measure."],"forward_implications":["Under the alignment condition of Theorem 1, training on a new task increases the loss on any earlier task by at most a term that shrinks with more rounds, more local epochs, and more participating clients.","SPECIAL reaches a task-uniform non-convex convergence rate of $O(\\sqrt{E/(NT)})$ under partial participation, matching single-task FedAvg when measured per participating update.","The bound's residual separates optimization noise ($\\sigma_L$, $\\sigma_G$) from inter-task drift ($\\sigma_T$), so the choice of proximal weight $\\lambda$ and the round/epoch trade-off can be matched to the expected amount of domain shift.","All of this is achieved without replay buffers, synthetic data, or task-specific heads: the server stores only the previous global model, so communication volume and model size stay identical to FedAvg."],"supporting_citations":[{"why":"The FedAvg algorithm that SPECIAL builds on; its client sampling, local-update aggregation, and communication-efficiency baseline are reused unchanged.","marker":"McMahan et al. (2017)"},{"why":"Supplies the partial-participation analysis lemma used to bound client drift under uniform subsampling of N of M clients.","marker":"Yang et al. (2021)"},{"why":"The centralized backward-knowledge-transfer result under cosine gradient alignment that Theorem 1 extends to the federated, partial-participation regime.","marker":"Lin et al. (2022)"},{"why":"Standard non-convex FedAvg convergence assumptions (bounded gradients, smoothness, client heterogeneity) on which the analysis rests.","marker":"Li et al. (2019a)"}],"fun_headline_variants":["Server-side anchor curbs drift in federated domain learning","One-line change to FedAvg gives backward transfer","Memory-free protection of old tasks in federated learning","Quadratic server blend matches FedAvg speed, protects prior tasks","SPECIAL: single server tweak for shifting federated domains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The backward-transfer guarantee in Theorem 1 only holds if, at every local step of every participating client, the new task's gradient points in the same direction as the combined gradient of all earlier tasks at the start of the task; the paper's own experiments show negative backward transfer on all four datasets, which suggests that condition is not satisfied there.","fun_headline_variants_meta":{"raw":{"variants":["Server-side anchor curbs drift in federated domain learning","One-line change to FedAvg gives backward transfer","Memory-free protection of old tasks in federated learning","Quadratic server blend matches FedAvg speed, protects prior tasks","SPECIAL: single server tweak for shifting federated domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000332,"raw_usage":{"total_tokens":1910,"prompt_tokens":1069,"completion_tokens":841,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":685,"completion_tokens_details":{"reasoning_tokens":761}},"tokens_in":685,"tokens_out":841,"duration_ms":8209,"temperature":1.0,"reasoning_tokens":761,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:39:40.019341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic federated domain-incremental problem with two tasks whose gradients are orthogonal, the $\\epsilon$-alignment premise of Theorem 1 is violated by construction; if SPECIAL still shows the earlier-task loss decreasing beyond what Eq. (7) allows, or if the loss increase exceeds the right-hand side, then Eq. (7) is not a valid certificate in that regime. Concretely, track the cosine similarity between $\\nabla f_{1:K-1}(\\theta^0_K)$ and each local stochastic gradient and check the inequality round by round.","supporting_citations":[],"review_version":1}