acceptodds
Withdrawn from ICLR 2027

Knowledge-Aware Evolution for Task-Free Streaming Federated Continual Learning with Arbitrary Class Overlap

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated Continual Learning (FCL) leverages inter-client collaboration to balance new knowledge acquisition and old knowledge retention, assuming explicit task boundaries. However, existing FCL methods struggle in streaming scenarios where ephemerally accessible data chunks exhibit arbitrary class overlap and lack task identifiers, inducing knowledge conflicts of the same class and preventing continual inference on local cumulative classes. To address this, we propose FedKACE with three synergistic components: 1) a gradient-responsive replay scheme that utilizes the ratio of squared L2 gradient norms of buffer to new samples to achieve client-specific balance between acquisition and retention; 2) a holistic buffer maintenance strategy that retains boundary-critical samples to enhance knowledge retention under class overlap; 3) an adaptive local-to-global inference model switching mechanism for continual inference on local cumulative classes. Experiments under diverse class overlap settings, along with theoretical analysis, demonstrate the effectiveness of FedKACE.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Federated Learning, Federated Continual Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_HUVeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies task-free streaming federated continual learning, where clients receive short-lived data chunks with arbitrary class overlap and need to perform inference on all previously seen classes. The authors propose FedKACE, which combines a gradient-responsive replay scheme to balance new and old knowledge, a buffer maintenance strategy to retain important samples, and an adaptive mechanism to switch between local and global models for inference. Experiments on CIFAR100 and ImageNet100 under different class overlap settings show that FedKACE outperforms the evaluated baselines in average accuracy. The paper also provides theoretical analyses to support the proposed framework.

Critical strengths
  1. The paper addresses a meaningful and challenging problem in federated continual learning by considering streaming data with arbitrary class overlap and no explicit task boundaries. This setting relaxes several common assumptions in existing FCL methods and is relevant to practical scenarios where data access and storage are limited. The problem is also clearly motivated and formulated.

  2. The proposed framework combines three components to address knowledge retention, buffer management, and inference model selection in a coherent way. The experiments cover two datasets and several class overlap settings, with FedKACE consistently outperforming the evaluated baselines. The paper also includes ablation studies to examine the contributions of different components, providing useful empirical evidence for the proposed approach.

Critical weaknesses
  1. Insufficient justification for the adaptive inference model switching criterion. In Section 4.3, the authors decide when to switch from the local model to the global model based on the change in the gap between the global model's accuracy and prediction confidence. However, a decreasing gap does not necessarily mean that the global model is improving, or that it has become better than the local model. For example, the gap could decrease even when the global model's accuracy drops. Moreover, the proposed criterion does not directly compare the two models, and the switch is irreversible even as new data continue to arrive. This raises concerns about whether the proposed criterion can reliably identify the right time to switch.

  2. Limited justification for the gradient-responsive replay weighting strategy. In Section 4.1, the authors use the ratio of squared gradient norms between buffered and new samples to balance knowledge retention and acquisition. However, gradient magnitude alone may not reliably reflect knowledge forgetting or conflicts between old and new knowledge. For example, large gradients can also result from difficult samples or distribution shifts, while gradient directions are important for identifying conflicts between updates. In addition, the paper does not clearly explain why the squared gradient norm ratio is preferred over other gradient-based weighting strategies.

  3. Limited baseline comparisons for the proposed task-free streaming setting. The experiments mainly compare FedKACE with traditional federated continual learning methods, many of which rely on explicit task boundaries or task identifiers. The authors adapt these methods by using FL rounds as task identifiers, but it is unclear whether this provides a sufficiently fair comparison, given the different assumptions of these methods. Although task-free online continual learning methods are discussed in the related work, their federated adaptations are not evaluated. As a result, the current comparisons may not fully demonstrate the advantages of FedKACE in the proposed task-free streaming setting.

Questions for authors
  1. Could the authors further explain why a decreasing gap between accuracy and prediction confidence indicates that the global model has become better than the local model? It would be helpful to show how this gap relates to the actual performance difference between the two models over FL rounds, and compare the proposed criterion with a simple performance-based switching strategy

  2. Could the authors clarify the motivation for using the squared gradient norm ratio, and how it relates to knowledge forgetting or conflicts? It would also be helpful to compare this choice with alternative strategies, such as the regular gradient norm ratio or methods that consider gradient directions, to better demonstrate the benefit of the proposed design.

  3. Could the authors clarify how the task-dependent baselines were adapted to the proposed setting and whether these adaptations ensure fair comparisons? It would be helpful to include at least one task-free online continual learning method adapted to federated learning, under the same memory and training budgets, to further validate the effectiveness of FedKACE.

3 comments

We sincerely thank the reviewer for reading our manuscript and providing valuable comments. Below, we present our point-by-point responses to the comments.


For Question 1,We thank the reviewer for the careful observation and constructive comments. We would like to clarify that the original intention of this switching criterion is not to directly compare the absolute performance of the two models in each round. Instead, it is designed to detect whether the global model has crossed the underfitting stage and achieved a good "prediction calibration" state on the client's local buffer, thereby determining its "usability" for the client for inference.

Regarding the concern that "the gap could decrease when accuracy drops," we highlight the truncation property of GAP=max⁡(0,ACC−PROB)GAP = \max(0, ACC - PROB). When a model overfits the local buffer, its accuracy (ACC) may stagnate or drop, but prediction confidence (PROB) surges abnormally. Consequently, PROB>ACCPROB > ACC forces GAP to truncate directly to 0, rather than "slowly decreasing." Instead, ΔGAP<0\Delta GAP < 0 captures the global model's monotonic transition from "underfitting" (ACC>PROBACC > PROB) to "well-calibrated" (PROB steadily approaching ACC) in early FL rounds, indicating it has fully adapted to the local distribution.

Regarding the rationality of the "unidirectional irreversible switching" and the robustness of "decreasing for two consecutive rounds,", it stems from Streaming FCL constraints. With tiny buffers, local models inevitably suffer catastrophic forgetting from incoming classes, while global models monotonically enhance generalization via aggregation. Thus, once well-calibrated, unidirectional switching optimally counters local degradation. Meanwhile, requiring "ΔGAP<0\Delta GAP < 0 for two consecutive rounds" smoothly filters outliers, preventing misjudgments from single-round fluctuations and ensuring robust decisions. We have confirmed this through ablation studies (AS1 - AS3).

Finally, regarding the "simple performance-based switching strategy" suggested by the reviewer, we need to point out its fatal flaws under the current setting. In the Streaming FCL setting, clients do not have an independent validation set, and all buffer data are involved in the training of the local model. Therefore, the ACC measured by the local model on the buffer suffers from severe systematic bias (optimistic estimation/overfitting). Directly comparing this biased ACC with the global model would almost always result in the local model "winning," which would cause the switching logic to completely fail and make it impossible to determine when the global model is truly usable.


For Question 2,We sincerely thank the reviewer for the careful observation and fully understand your concerns about whether gradient magnitude alone can reliably reflect knowledge forgetting or conflicts between old and new knowledge, as well as your questions about why the squared gradient norm ratio is preferred over other gradient-based strategies. We provide detailed clarifications and explanations for these critical issues below.

First, we would like to clarify that the weight λ\lambda is not a direct measure of "knowledge forgetting" or "gradient direction conflicts," but rather a proxy for the relative update pressure of the replay constraint versus the new data constraint. The ratio of squared gradient norms essentially corresponds to the gradient energy ratio, which is more sensitive to the relative changes in replay and task constraints. When the replay gradient is relatively larger than the task gradient, it indicates that the loss sensitivity of the model on old samples is relatively stronger; thus, increasing the replay weight strengthens the retention constraint for old knowledge. Conversely, decreasing the weight frees up parameter capacity for new knowledge acquisition. Moreover, this relative ratio inherently resists noise: global perturbations (e.g., distribution shifts) affect both losses and cancel out in the ratio, ensuring the weight only responds to single-sided imbalances for dynamic equilibrium.

Second, as you rightly pointed out, large gradients in a short time window can indeed result from the sampling of difficult samples or local distribution shifts. To mitigate this, instead of using single-batch instantaneous gradients, we compute the next epoch's replay weight via epoch-level running averages of both losses. This smoothing mechanism effectively filters out gradient spikes, ensuring a robust λ\lambda evolution that accurately reflects the macroscopic relative learning states of old and new knowledge.

Regarding the alternative strategies considering gradient directions you suggested, while effective in traditional CL, they face semantic and computational failures under our task-free, arbitrary-overlap streaming setting. Semantically, without task boundaries, old and new samples of the same class often yield synergistic rather than conflicting gradients, rendering directional conflict signals invalid. Computationally, full-model gradient projection requires storing all parameter gradients, which is prohibitive for edge clients. Our ablation (AS5) confirms this increases FLOPs by 2,780× to 35,600×. Thus, the output-layer gradient proxy remains the only feasible compromise balancing theoretical rigor and practical scalability.

Finally, regarding why the "squared" gradient norm is preferred over the "regular" gradient norm, this is not an arbitrary heuristic design but a mathematical necessity dictated by strict optimization theory. The regular L2 norm is non-differentiable (non-smooth) when the gradient approaches zero, which violates the fundamental assumptions in optimization theory. Conversely, the squared norm's smoothness ensures the quotient mapping is Lipschitz continuous near the equilibrium, a prerequisite for the Banach fixed-point theorem. Crucially, only the squared norm, combined with active parameter masking, strictly derives the "gradient energy balance" (yielding λ∗=1\lambda^*=1 from ∥∇Ltask∥22=∥∇Lrep∥22\|\nabla L_{task}\|_2^2 = \|\nabla L_{rep}\|_2^2). This proves convergence to a perfectly balanced equilibrium, which non-squared norms cannot guarantee.


For Question 3,We sincerely thank the reviewer for the insightful comments. Regarding your concerns about the baseline comparisons and the fairness of the adaptations, we provide detailed clarifications and supplementary explanations below.

First, regarding the adaptation of Task-dependent FCL Baselines, we uniformly provided them with the FL round index as the explicit task identifier. The fairness of this adaptation is validated across settings. Specifically, under the O=0O=0 setting, adjacent FL rounds have strictly disjoint class sets, which perfectly aligns with the "mutually exclusive class sets across tasks" assumption inherent to traditional FCL methods, making this comparison absolutely fair. Conversely, under O=5O=5, the class window remains fixed for 5 consecutive rounds, simulating the "static class sets" assumption in OCL. For partial overlaps (O=4,2O=4, 2), arbitrary class overlap breaks traditional FCL assumptions. As shown in Table 1, traditional methods suffer significant performance degradation here, precisely demonstrating that their reliance on explicit task boundaries is an inherent limitation in streaming scenarios. Highlighting this gap is necessary, as overcoming this exact limitation is the core motivation for proposing FedKACE.

Second, we sincerely appreciate this constructive suggestion for including at least one task-free online continual learning method adapted to federated learning, we have supplemented a federated adaptation of a task-free OCL method DEMD[1], namely F-DEMD, as a new baseline to further validate the effectiveness of FedKACE.

[1] F. Ye and A. G. Bors, Online Task-Free Continual Learning via Dynamic Expansional Memory Distribution. CVPR2025


We thank the reviewer for patiently reading through this entire response, and we kindly invite the reviewer to examine the newly uploaded revised manuscript.

Reviewer_qdx7OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates a federated continual learning (FCL) problem. The proposed FedKACE consists of three modules: GRRS, which adjusts replay weights based on the gradient norm ratio between new and old data; HBMS, which maintains a memory buffer by considering both informativeness and diversity; and an inference mechanism that switches unidirectionally from the local model to the global model based on the difference between buffer accuracy and the predicted probabilities assigned to the ground-truth classes. The authors report performance improvements over existing baselines on CIFAR100 and ImageNet100 and attempt to theoretically justify the proposed modules through three theorems.

Critical strengths
  1. Compared with settings that rely on explicit task boundaries, this paper simultaneously considers short-term data accessibility, class recurrence over time, and differences in class trajectories across clients, while requiring inference over all classes previously encountered locally at every round. These constraints help advance FCL beyond static task partitioning toward more dynamic application scenarios.

  2. Table 1 shows that FedKACE outperforms all listed baselines across both datasets and all four overlap configurations. This combined approach demonstrates some empirical value under the current experimental protocol.

Critical weaknesses
  1. The objective is defined as L(θ,λ)=Ltask(θ)+λLrep(θ)L(\theta,\lambda)=L_{\mathrm{task}}(\theta)+\lambda L_{\mathrm{rep}}(\theta). When the replay loss is positive, fixing θ\theta makes the objective monotonically increasing in λ\lambda. Thus, maximizing over λ\lambda should yield the maximum of its feasible domain. However, the appendix derives λ∗=1\lambda^*=1 from a gradient-ratio fixed point without proving that it solves the stated maximization problem or specifying that max⁡Ωλ=1\max\Omega_\lambda=1. Therefore, the weight-update fixed point cannot be directly equated with a minimax saddle point of Equation (8).

  2. On page 20, the authors infer ∇Q(θ)→0\nabla Q(\theta)\to0 and ∇2Q(θ)→0\nabla^2Q(\theta)\to0 from Q(θ)→Q(θ∗)=1Q(\theta)\to Q(\theta^*)=1 and continuity. This is generally invalid, as convergence of a function does not imply convergence of its derivatives to zero. Moreover, the Hessian lower bound in Equation (41) should yield an inner-product lower bound in Equation (43), but the paper states an upper bound and uses it to establish contraction. These errors undermine Lemma 1 and Theorem 1.

  3. After canceling the positive learning rate, the second condition in Equation (62) becomes LQL2+HQ≤μ1(δ)/4L_QL_2+H_Q\leq\mu_1(\delta)/4. This condition cannot be satisfied merely by reducing the learning rate. Nevertheless, the paper converts it into a learning-rate upper bound and claims that contraction follows as the learning rate approaches zero. The required relationship among these constants must be established rather than relying solely on Robbins–Monro step-size conditions.

  4. All clients use 500 new samples per round, yielding ν=nmin⁡/nmax⁡=1\nu=n_{\min}/n_{\max}=1. However, Equations (195) and (218) contain 1/(1−ν)1/(1-\nu) without a separate derivation for ν=1\nu=1. For 0<ν<10<\nu<1, the long-term limit is 3/ν2≥33/\nu^2\geq3, looser than the trivial upper bound of 1 on total variation (TV) distance. This provides little support for the claimed decay of the heterogeneity term.

  5. For Equations (97)–(98), V>0V>0 and m>1m>1 imply δlb>δub\delta_{\rm lb}>\delta_{\rm ub}, contradicting the stated ordering δub≥ropt≥δlb\delta_{\rm ub}\geq r_{\rm opt}\geq\delta_{\rm lb}. Moreover, normalized logits lie on the unit sphere, which has zero dd-dimensional Lebesgue measure in Rd\mathbb R^d. Thus, the positive-volume argument is invalid.

  6. On page 32, the authors claim that maximizing IDV is equivalent to maximizing the distance term because the information term does not directly optimize the kernel condition number. However, the information term can still change candidate rankings. Furthermore, DS measures distances to the fixed old buffer rather than the incrementally constructed new set. Therefore, the greedy max-min approximation guarantee does not directly apply, and the claim that CDV ranking strictly improves the final set's minimum pairwise distance remains unsupported.

  7. Equation (116) claims a lower bound on the spectral radius of random sampling based on expected minimum distance without proof. Gershgorin's circle theorem does not directly establish the condition-number lower bound required by Equation (117). Additionally, the assumption that relevant radii are smaller than 1, required by expressions such as Equation (115), is not established. Hence, the claimed strict spectral advantage Cκ<1C_\kappa<1 remains unproven.

  8. Equation (153) assumes that HBMS's empirical risk deviation is at most CκC_\kappa times that of a random buffer, but Theorem 2 does not establish this relationship. The preceding RKHS Rademacher bound primarily depends on tr⁡(K)\operatorname{tr}(K), which equals the sample size for a normalized Gaussian kernel and does not automatically decrease with the condition number. This gap undermines Lemma 2 and the subsequent regret analysis.

  9. Equations (212) and (215) provide upper bounds on local and global regret, respectively, but Equation (216) equates the difference between these bounds with the actual regret difference. In general, given a≤Aa\leq A and b≤Bb\leq B, neither a−b=A−Ba-b=A-B nor a−b≤A−Ba-b\leq A-B necessarily holds. This invalidates the subsequent analysis in Equation (217) and the conclusions regarding model switching.

  10. All clients use 500 new samples per round, yielding ν=nmin⁡/nmax⁡=1\nu=n_{\min}/n_{\max}=1. Yet Equations (195) and (218) contain 1/(1−ν)1/(1-\nu) without handling this case separately. For 0<ν<10<\nu<1, the long-term limit 3/ν2≥33/\nu^2\geq3 also exceeds the trivial TV-distance bound of 1, weakening the claimed heterogeneity analysis.

  11. According to Section 5.1 and Appendix A.1, each client processes 100×5×100=50,000100\times5\times100=50{,}000 distinct training samples, exactly matching the size of the CIFAR100 training set. Thus, under the reported protocol, every client eventually sees the entire training set rather than maintaining mutually exclusive federated data partitions.

Questions for authors

as above

3 comments

We thank the reviewer for the careful reading of our paper and for the extremely detailed comments on our theoretical derivations, which have been very helpful in improving our manuscript. We have rewritten the entire theoretical part of the paper.

During proof reconstruction, we found that after resolving the argument errors raised in your comments 5, 6, 7, and 8, Theorem 2 only holds with high probability within a specific kernel bandwidth range, which our data-dependent bandwidth cannot guarantee. Moreover, it fails to establish a connection with the regret bound, rendering the recursive bound in Lemma 2 redundant. Thus, we removed both Theorem 2 and Lemma 2.

The revision changes the theoretical framework as follows.

Removals: the old Theorem 2 (spectral efficiency of HBMS) and the old Lemma 2 (recursive regret bound for local training) are removed together with their proofs; five assumptions (old 4, 6, 8, 9, 12) are removed accordingly; and the theoretical claims regarding HBMS are withdrawn, leaving only empirical evidence as support.

Additions: one assumption (new 14).

Revisions: one assumption (11).

Rewritings: the statements and proofs of the two retained theoretical modules are rewritten, namely the old Lemma 1 and Theorem 1 (rewritten as new Lemma 1 and new Theorem 1), and the old Lemma 3 and Theorem 3 (rewritten as new Lemma 3 and new Theorem 3).

The revised theory section contains 2 theorems, 2 lemmas, and 9 assumptions. The point-by-point response below explains how these changes specifically resolve each of the issues you raised.


In response to Comment 1 (Minimax Saddle Point), we fully agree this comment, and we have removed all two-sided minimax equalities and the "ideal saddle point / minimax equilibrium" statements from Lemma 1 and Theorem 1, restating the conclusion as a gradient-balanced fixed point satisfying the coupled stationarity condition, the weight identity, and a one-sided local variational inequality, whose existence and local uniqueness are established by a Banach fixed-point argument; the equivalence claim "weight-update fixed point = minimax saddle point" has been removed from the paper.


In response to Comment 2 (Invalid Derivative Limits & Inner-product Bound Direction), we fully agree, and in the proof of the new Lemma 1 we have removed in its entirety the Hessian-perturbation derivation that relied on this invalid inference, and instead we directly establish an inner-product lower bound with the correct direction, using only the Lipschitz continuity of the gradient-quotient mapping Q, Assumption 1, and the revised Assumption 11(strong convexity in active parameter space).


In response to Comment 3 (Learning Rate Condition in Eq. 62), we fully agree, and in the new proof we have explicitly promoted the constant gap guaranteeing contraction (e.g., μ>2L2LQ\mu > 2 L_2 L_{\mathcal{Q}}) to the stated prerequisites of the lemma and the theorem, while reconstructing the learning-rate upper bound in conjunction with the relative gradient bound so that it genuinely constrains the higher-order terms containing the step size, strictly restricting the role of the Robbins-Monro conditions to merely ensuring that the learning rate falls into this admissible interval in the later stages of iteration.


In response to Comment 4 (Singularity at ν=1) and Comment 10 (Trivial TV Bound), we fully agree and have reconstructed the relevant proofs and statements: in Lemma 3, we sharpened the weight-difference bound to (1−ν2)/(νt)(1-\nu^2)/(\nu t) to eliminate the singularity, removed the 1/(1−ν)1/(1-\nu) factor via an integral bound, imposed a min⁡(⋅,1)\min(\cdot, 1) safety cap on ΔTV\Delta^{\mathrm{TV}}, and introduced an instance-dependent overlap rate; in Theorem 3, we completely deleted the original two-stage "decay-dominance" asymptotic narrative and limit claims, replacing them with conditional conclusions based on specific parameter conditions in FL round 1.

In response to Comment 5 (Volume/Packing Argument Contradiction & Zero Lebesgue Measure), Comment 6 (IDV/CDV Greedy Approximation Guarantee),Comment 7 (Spectral Radius Lower Bound & Gershgorin Theorem) and Comment 8 (Eq. 153 CκC_\kappa Scaling & Rademacher Complexity), we fully agree, and we have removed Theorem 2 in its entirety (statement and proof, including the volume/packing argument, the greedy argument, the spectral-radius lower bound, Gershgorin application, radius-assumption derivations and the CκC_\kappa scaling assumption of Eq. (153)), retracted the theoretical claim that HBMS enjoys a strict spectral advantage (CκC_\kappa < CrandC_{rand}), and the theoretical claims of the greedy approximation ratio and the strict improvement by CDV, replaced it with empirical evidence from ablation studies and buffer kernel condition numbers to motivate the design, and simultaneously removed Lemma 2, which consumed CκC_\kappa, so that the regret analysis no longer depends on this constant. Finally, no unproven approximation ratio or strict-improvement claim remains anywhere in the paper, and the deletion incurs no transmission effect on the retained Theorem 1 and Theorem 3.


In response to Comment 9 (Upper Bound Subtraction in Eq. 216), we fully agree, and in the new Lemma 3 we let the client average of the local regrets enter the global regret upper bound as an exact identity (the actual quantity), so that after averaging over clients the local-regret terms on both sides cancel exactly via transposition and the gap upper bound becomes the lemma's statement itself, and the new Theorem 3 directly cites this gap upper bound and replaces the original sign analysis with conditional sign conclusions, deleting the "subtraction of two upper bounds" step entirely.


In response to Comment 11, we thank the reviewer for the rigorous observation and agree that clients eventually access the entire training set. However, this overlooks the paradigm essence of Streaming FCL.

In real edge scenarios, strict spatial exclusivity is unrealistic since devices naturally collect similar data over time. Instead, Streaming FCL focuses on the temporal arrival of ephemeral data chunks, which induce severe intra-client knowledge conflicts and catastrophic forgetting. Accordingly, our metrics strictly evaluate each client's local cumulative classes.

Furthermore, heterogeneous starting points and evolving trajectories cause arbitrary overlaps in local class sets at intermediate FL rounds, providing complementary knowledge for federated collaboration.

Moreover, as shown in Appendix A.4, unlike static FCL, where repeated FL training in each task causes excessive sample exposure, our streaming setting accesses each sample locally only once, bounding global reuse to at most "client count" times.

Finally, Figure 2 illustrates complete accuracy trends across all FL rounds in experiments, enabling direct performance evaluation at any intermediate stage before all data is visited by each client(i.e., t < 100).


We thank the reviewer for patiently reading through this entire response, and we kindly invite the reviewer to examine the newly uploaded revised manuscript; to facilitate verification of the revisions, all retained assumptions and theorems/lemmas keep their original numbering in the revision (the numbers vacated by removed items are not reused, and the newly added assumption is numbered 14), so every number cited in this response can be directly located in the revised manuscript.

Although you provided a detailed response, many of the comments I raised above are actually unlikely to lead to a higher score. However, I do appreciate your efforts to ensure the rigor of the paper. I look forward to seeing further improvements and refinements in future papers.

Reviewer_FwNiOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies task-boundary-free streaming federated continual learning with varying degrees of class overlap. It proposes FedKACE, uses sophisticated techniques to design the algorithm, and conducts extensive experiments. The experimental results are also strong. However, the theoretical analysis is extremely difficult to read and falls far short of the standards of academic writing.

Critical strengths

1.The method performs well experimentally. The main text presents a substantial set of experiments, and the figures and tables are clear.

2.The ablation studies are thorough and convincing.

Critical weaknesses

The paper is very poorly written, especially the theoretical sections, which are virtually unreadable. The authors need to carefully rewrite the entire theoretical part. First, the idea of using the regularizer in Equation (4) to constrain the classes appears to conflict with Assumption 11 in the appendix. In addition, the formal statements of the theorems in the main text are very unclear, and regret bounds and other concepts are invoked frequently. The appendix proofs appear to substitute total-variation calculations for many of the bounds that are needed. The proof of Equations (38)–(41) contains an error: on page 31, the dd-dimensional Lebesgue measure of the normalized logits set G⊂Rd\mathcal G \subset \mathbb R^d is zero, making the proof incorrect. The assumptions in the appendix are also highly disorganized and substantially overlap with one another, which makes them very difficult to follow. The appendix proofs feel like a collection of disconnected arguments. Overall, the paper is extremely difficult to read.

Questions for authors

See the Weaknesses.

2 comments

We thank the reviewer for the careful reading and insight comments. We agree that the original theoretical sections suffered from severe flaws in organization, formal statements, and proof derivations, and we accept the recommendation to rewrite the entire theoretical part.

During proof reconstruction, we found that after resolving the argument errors raised in another reviewer qdx7's comments 5, 6, 7, and 8, Theorem 2 only holds with high probability within a specific kernel bandwidth range, which our data-dependent bandwidth cannot guarantee. Moreover, it fails to establish a connection with the regret bound, rendering the recursive bound in Lemma 2 redundant. Thus, we removed both Theorem 2 and Lemma 2.

The revision changes the theoretical framework as follows.

Removals: the old Theorem 2 (spectral efficiency of HBMS) and the old Lemma 2 (recursive regret bound for local training) are removed together with their proofs; five assumptions (old 4, 6, 8, 9, 12) are removed accordingly; and the theoretical claims regarding HBMS are withdrawn, leaving only empirical evidence as support.

Additions: one assumption (new 14).

Revisions: one assumption (11).

Rewritings: the statements and proofs of the two retained theoretical modules are rewritten, namely the old Lemma 1 and Theorem 1 (rewritten as new Lemma 1 and new Theorem 1), and the old Lemma 3 and Theorem 3 (rewritten as new Lemma 3 and new Theorem 3).

The revised theory section contains 2 theorems, 2 lemmas, and 9 assumptions. The point-by-point response below explains how these changes specifically resolve each of the issues you raised.


In response to "the masking mechanism in Equation (4) (which induces zero eigenvalues in the Hessian) conflicts with Assumption 11 (global strong convexity) in the appendix," we fully agree and have made two core revisions: We fully agree and have made two revisions:

  1. For local training (new Lemma 1 and Theorem 1), we restricted the scope of strong convexity in Assumption 11 from the "full parameter space" strictly to the "active parameter subspace and its local neighborhood."

  2. For federated aggregation (new Lemma 3), due to not using masking mechanism during inference, we corrected the strong convexity premise relied upon by Jensen's inequality from the "loss function" to the "risk function", To this end, we introduced a new Assumption 14 (μR\mu_R-strong convexity of the risk function).


In response to "the formal statements of the theorems in the main text are very unclear and regret bounds and other concepts are invoked frequently," we made two categories of revisions:

  1. Self-contained restatements—new Theorem 1 accordingly removes the original two-sided minimax equality and retains only the coupled stationarity condition, the weight identity, and the truncation error bound; and new Theorem 3 is split into two parts—an upper bound on the expected regret gap and a conditional sign conclusion for the first round—with narrative language removed.

  2. Unified definitions and citations—regret is fully defined in place within each result statement that invokes it, while ΔTV\Delta^{\mathrm{TV}} is defined exactly once in new Lemma 3 and directly cited by new Theorem 3 without re-expansion.


In response to "the appendix proofs appear to substitute total-variation calculations for many of the bounds that are needed," we acknowledge that in the original manuscript bounds that should have been established independently—such as the distribution-shift term and the class-prior weight deviation—were substituted by TV calculations, and we made the revisions: we removed the old Lemma 2 and the old Theorem 2, the buffer empirical-risk deviation term and the distribution-shift term, which had been bounded via TV distance plus a decay assumption, both exit the regret chain, so the TV distance is restricted to two uses with independent justification:

  1. Cross-distribution expected-risk differences are bounded directly by the definition of TV distance together with the boundedness of the risk.

  2. The TV distance between the cumulative distributions of any client pair is bounded by a unique global upper bound ΔTV\Delta^{\mathrm{TV}} established via the class-proportion mixture decomposition and recursion.


In response to "the proof of Equations (38)–(41) contains an error and, on page 31, the d-dimensional Lebesgue measure of the normalized logits set G is zero, making the proof incorrect," we fully acknowledge and made two removals:

  1. The old Theorem 2, together with its volume/measure argument on page 31, is removed in its entirety and the claim Cκ<CC_\kappa < C_rand is withdrawn.

  2. The entire block of old Equations (38)–(41) is removed from the proof of new Lemma 1, whose only use of strong convexity now invokes the revised Assumption 11 within the neighborhood of the active parameter subspace rather than being established via Hessian perturbation.

In response to "the assumptions in the appendix are highly disorganized and substantially overlap with one another, which makes them very difficult to follow," we reorganized the assumption system in two aspects:

  1. Removing redundancy—we deleted five assumptions (old 4, 6, 8, 9, 12) old 4 served only the removed old Theorem 2; old 6 lost its consumer with the removal of old Lemma 2; old 8 is no longer consumed in the new proofs (distribution shift is now bounded via the class-proportion recursion); old 9 was never directly consumed by any proof even in the original manuscript (the buffer bound in old Lemma 2 was self-derived); and old 12 and old 13 both belong to second-moment-type assumptions on stochastic gradients while the new proofs consume only the latter;

  2. Separating mixed usage—old 11 originally supported both the loss-space strong convexity for local training and the risk-space strong convexity required by the aggregation Jensen in old Lemma 3; we now restrict it to μ-strong convexity within the active parameter subspace and a local neighborhood (serving only the local training chain), and add new Assumption 14 (νR\nu_R-strong convexity of the risk function on conv(Ω_θ)) exclusively for the aggregation step;


In response to "the appendix proofs feel like a collection of disconnected arguments,", after the removal of the unprovable old Lemma 2 and old Theorem 2—which in the original manuscript served precisely as the bridge connecting the two proof chains (old Lemma 2 consumed the truncation bound of old Theorem 1 and the spectral efficiency constant of old Theorem 2)—do the two proof chains become mathematically independent.

Methodologically, however, the framework remains a closed loop: the buffer maintained by HBMS serves as the data carrier of the replay loss in GRRS; the gradient-balanced convergence of new Theorem 1 guarantees that the local model is reliable in early rounds, supporting the inference model switching mechanism's design of defaulting to the local model early; the regret game between the aggregated global model and the local model over the clients' cumulative data distributions is characterized by new Theorem 3, providing a guarantee for the inference model switch.


We thank the reviewer for patiently reading through this entire response, and we kindly invite the reviewer to examine the newly uploaded revised manuscript; to facilitate verification of the revisions, all retained assumptions and theorems/lemmas keep their original numbering in the revision (the numbers vacated by removed items are not reused, and the newly added assumption is numbered 14), so every number cited in this response can be directly located in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.