acceptodds
Under review as a conference paper at ICLR 2027

Co-Evolving Rehearsal States for Federated Class-Incremental Learning under Asynchronous Class Exposure

Abstract

Federated class incremental learning (FCIL) must retain earlier classes while clients continue learning from different class histories. At the same round, a class may be current for one client, historical for another, and unseen by a third. We call this disagreement asynchronous class exposure (ACE). Existing methods preserve history through replay, prototypes, feature memories, or parameter constraints, but they often treat stored evidence as fixed while the model representation continues to change. As a result, historical evidence can bypass trainable modules, use incompatible encoder coordinates, or become stale over time. We address this gap with Projected Rehearsal Orchestration (PRO), which rehearses current and remembered evidence before the adapter and rebuilds available class statistics under the aggregated global encoder. When fresh examples are unavailable, PRO-MAX estimates class local feature movement from current calibration data and combines client estimates using support weights to maintain the remaining memories. Across six image, text, and graph benchmarks, PRO and PRO-MAX improve retention under heterogeneous exposure, with PRO-MAX achieving the highest reported mean final accuracy among the evaluated methods on all six streams. We further show that the gains are not explained by feature sampling alone. Instead, effective rehearsal depends on keeping historical evidence trainable, compatible with the current representation, and updated as learning evolves.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
federated class-incremental learning, asynchronous class exposure, heterogeneous, continual learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_w1pkOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper addresses Federated Class-Incremental Learning (FCIL), where at the same global round the same class can simultaneously be current, historical, or global-only across participating clients. The authors name this Asynchronous Class Exposure (ACE), and propose treating FCIL as co-evolution of a global model and a shared rehearsal state. Their PRO (Projected Rehearsal Orchestration) decomposes each round into READ–OPTIMIZE–WRITE: balanced pre-adapter rehearsal plus fresh WRITE under the aggregated global encoder to eliminate coordinate mismatch; PRO-MAX adds kernel-regression estimation of conditional class-mean displacement when no fresh WRITE is available. PRO-MAX achieves the highest mean Final Average Accuracy on six image / text / graph benchmarks (CIFAR-100, TinyImageNet, THUCNews-10, CLINC150, Cora, ogbn-arxiv). The theoretical appendix (App. B) gives six propositions and one theorem with explicitly stated scope; ablations isolate placement + refresh from feature-spread and replay fidelity.

Critical strengths
  1. Conceptual reframing of FCIL around ACE and co-evolution. Naming ACE (current / historical / global-only) and treating FCIL as joint evolution of (\Theta_r^g, M_r) rather than "model with frozen memory" gives a clean lens on a fragmented literature. The three failure modes (“bypass adapter” / “coordinate mismatch” / “staleness”) map one-to-one onto READ–OPTIMIZE–WRITE, making the design choices feel necessary rather than ad-hoc.

  2. Methodologically tight and modality-agnostic. PRO is closed-form on the server; PRO-MAX MOVE is a small, inspectable kernel-regression step that explicitly does not attempt covariance recovery; the per-class memory interface is reused unchanged across CIFAR-100 / TinyImageNet / THUCNews-10 / CLINC150 / Cora / ogbn-arxiv; PRO-MAX is the strongest reported method on every one of the six streams, and ablations (Fig. 3 and Tables 3–4) isolate placement + refresh from feature-spread and replay fidelity.

Critical weaknesses
  1. Replay-object budget is not matched. Table 9 makes the asymmetry explicit (HR / FedCBDR / TARGET / MFCL get 100 replay objects vs. PRO / PRO-MAX at 20) and acknowledges it ("object-count configurations, not matched-byte, matched-information, or matched-compute regimes"), but does not run the controlled experiment that would be needed to fully separate the algorithmic contribution from the resource asymmetry — e.g. HR / FedCBDR / TARGET / MFCL reduced to 20 replay objects, or PRO / PRO-MAX expanded to 100 projected pseudo-features.

  2. Scope and directness of the ACE measurement. The headline setting (40/50/60) conflates three interventions — task-support, label coverage and task-order perturbation — rather than directly realising ACE (two clients disagreeing about the role of the same globally exposed class in the same round). Appendix D is transparent ("severe heterogeneity in the retained tables denotes this reference setting, not a measured ACE value"), but the headline claim "transfer under ACE" is not directly tested. Theoretical results are per-round with no convergence / generalisation analysis of the federated training procedure (App. B.6).

Questions for authors
  1. Could you re-run the heterogeneous CIFAR-100 comparison under matched replay budgets (e.g. HR / FedCBDR / TARGET / MFCL with 20 replay objects and PRO / PRO-MAX with 100 projected pseudo-features, or matched FLOP / matched bytes)? This would let the reader separate the proposed algorithmic choice from the asymmetry in Table 9.

  2. Could you add a control that exposes an explicit ACE rate — e.g. per round, the fraction of globally exposed classes that two participating clients assign to different roles (current vs. historical vs. global-only) — and report FAA / AF / CTA as a function of that ACE rate? This would let the community verify the ACE-centric narrative more directly than the bundled 40 / 50 / 60 setting does.

  3. Could you provide a short ablation on the size of the calibration set, the choice of K for KNN, and the bandwidth \tau_{c,y} (median heuristic vs. cross-validated scale) in PRO-MAX MOVE, so that the gap between PRO and PRO-MAX can be attributed cleanly to the bandwidth / weight scheme rather than to the extra forward passes?

7 comments
Reviewer_w1pkOpenReview ↗

The statement of AI documentation

I have filled out the AI documentation. The review was independently assessed by me. An LLM was only used for language polishing; all technical evaluation and the final score are my own judgment. There is no mismatch between my human assessment and the AI-assisted drafting. I will actively join the discussion with authors and the AC during the discussion period.

We are grateful for your detailed comments. We respond to each question below.

Q1. Matched replay budgets

A1. Thank you for suggesting this comparison. To address the budget difference, Table R11 compares 20 and 100 replay objects per historical class, participating client, and local epoch. The original settings use 100 for TARGET, MFCL, HR, and FedCBDR and 20 for PRO and PRO-MAX (paper Table 11). Each method retains its architecture and initialization. Loss weights for each source and optimizer updates are controlled separately.

Table R11. FAA at 20 and 100 replay objects (paper Table 3).

MethodFAA, 20 objectsFAA, 100 objects
TARGET27.1527.86
MFCL28.7129.42
HR31.4732.18
FedCBDR38.2538.96
PRO43.2843.99
PRO-MAX46.1246.83

PRO and PRO-MAX exceed FedCBDR by 5.03 and 7.87 FAA points at either count (Table R11). The ordering is unchanged, but these are mean comparisons. Results for individual seeds and paired uncertainty are unavailable for this control.

Beyond replay count, paper Table 4 compares methods using the same architecture and training policy. At 20 objects, this raises FedCBDR to 43.93 FAA. PRO is then 0.65 lower and PRO-MAX 2.19 higher. Section 4 distinguishes these margins from matching only object counts. Equal counts still allow differences in information content, storage, and FLOPs because images and feature vectors require different resources.

Q2. Realized ACE and performance

A2. Thank you for requesting explicit ACE rates. Eq. (21) measures role disagreement for a pair of participating clients and a globally exposed class. The 40/50/60 configuration describes the sampler settings. It is not itself an ACE measurement.

Using this definition, Table R12 reports mean ACE from our exposure simulation with fixed ownership and participation while changing order.

Table R12. Stream-average ACE from the exposure simulation (paper Table 45).

Order perturbationStream-average ACE
0%0.33815
30%0.25249
60%0.22996
90%0.22304

If more clients assign a class to the global-only role, disagreement can fall despite greater order perturbation. However, these ACE values do not come from the training sweep, so they cannot provide the requested relationship between ACE and FAA, AF, or CTA in the same runs.

For performance, the training sweep in paper Table 29 provides separate evidence. From 0% to 90% perturbation, FAA falls by 8.02, 2.92, and 2.66 points for FedCBDR, PRO, and PRO-MAX. AF rises from 30.96 to 46.92, 28.64 to 36.58, and 24.22 to 31.24, respectively. These results support lower sensitivity to ordering. Relating these performance results to ACE still requires role histories and metrics from the same runs, with labeled sample identities, ownership, and counts fixed. Order also changes recency, so this intervention cannot isolate ACE as a causal factor.

For the theoretical question, Proposition 1 separates class coverage and weighting from feature mismatch.

∣Fcum−F~c∣≤B∥ρcrep−1/J∥1+LuDc.|F^{\mathrm{cum}}-\widetilde F_c|\le B\|\rho_c^{\mathrm{rep}}-\mathbf1/J\|_1+L_uD_c.

In this bound, uniform rehearsal over every exposed class removes the first term even when local roles differ. The remaining feature mismatch motivates WRITE and MOVE. Appendix I explains these operations at a fixed stage without establishing convergence or generalization of the full training process. ACE describes the exposure setting and is not itself the error term in this bound.

Q3. Calibration size, neighbors, and bandwidth

A3. Thank you for suggesting these controls. Table R13 varies K with 128 calibration samples. Table R14 varies calibration size with K=5K=5.

Table R13. Sensitivity to K with 128 calibration samples (paper Table 35).

KFAAAF
143.9230.86
345.3827.42
546.1226.18
1045.9426.74
2044.2629.62

Table R14. Sensitivity to calibration size with K=5K=5 (paper Table 37).

Calibration samplesFAAAF
1643.8431.04
3244.9628.72
6445.6827.36
12846.1226.18
25646.2426.02

In the neighborhood sweep, intermediate values of K give the strongest means. Increasing calibration from 128 to 256 adds 0.12 FAA and reduces AF by 0.16 points. This supports the selected setting but does not establish statistical equivalence between nearby settings.

To examine the transport estimate itself, Table R15 tests the contribution of the estimated movement.

Table R15. Mean FAA and AF for transport controls (paper Table 15).

VariantFAAAF
PRO43.2831.54
Zero transport43.3231.50
Shuffled class42.8633.06
Uniform weights44.9028.66
PRO-MAX46.1226.18

PRO-MAX exceeds zero, shuffled, and uniform transport by 2.80, 3.26, and 1.22 FAA points. Zero transport stays close to PRO despite retaining the additional code path. These controls support the value of estimating movement for each class.

However, to isolate the effect of extra forward passes, a control must match forward calls, normalization updates, and random draws while suppressing both local and server movement. Because the current code description does not verify all these details, the control does not fully isolate this effect.

Finally, the comparison between median and cross-validated bandwidths is not reported. Uniform weighting also changes the estimator and may change aggregation support. Its 1.22-point gap therefore does not isolate bandwidth. A bandwidth comparison must keep the kernel family, neighbors, calibration inputs, and aggregation rule fixed while changing how the scale is selected. Appendix J.4 states that these controls support the tested transport setting without establishing that the median heuristic is optimal.

Reviewer_mBSuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper addresses the problem of asynchronous class exposure in federated class-incremental learning, where clients encounter different classes at different stages of training. The proposed methods (PRO and PRO-MAX) perform an initial label-free warm-up at clients. Then, PRO maintains a global model and a shared class-level memory on the server, consisting of the mean, standard deviation, and sample count of feature representations for each observed class. During local training, clients combine features extracted from current data with synthetic historical features sampled from the shared memory to train the adapter and classifier, mitigating catastrophic forgetting. After model aggregation, clients with available class examples recompute and upload class statistics using the updated global encoder, keeping the memory aligned with the evolving representation. PRO-MAX extends PRO by addressing classes for which fresh examples are unavailable; in this case, clients estimate feature displacement using current calibration samples.

Critical strengths
  • The paper addresses a relevant problem in federated class-incremental learning, considering heterogeneous class exposure across clients.

  • The experimental evaluation is extensive, covering multiple datasets and modalities, with significant improvements over existing baselines.

Critical weaknesses
  • Limited methodological novelty: The approach appears to be primarily an engineering contribution, combining several mechanisms already explored in previous work. The fundamental novelty beyond their integration is not sufficiently clear, also due to the limited discussion of related work in the main paper.

  • Potential unfair advantages over baselines: It is unclear whether competing methods use the same encoder–adapter–classifier architecture and whether the shared self-supervised warmup disproportionately benefits PRO and PRO-MAX. The contribution of these design choices is not sufficiently isolated, raising questions about whether the reported improvements originate from the proposed mechanisms.

  • Privacy concerns with sparse class availability: When a client has only one sample of a class, the uploaded class mean corresponds exactly to its feature representation, potentially exposing information about individual training examples. The paper does not adequately investigate this privacy issue.

Questions for authors

Q1. The proposed methods rely on an encoder–adapter–classifier architecture, where historical feature rehearsal explicitly updates the adapter. However, I am not able to understand from the paper whether all the baselines use such architecture. Could the authors clarify which architectures are used by the competing baselines and, if the encoder–adapter–classifier architecture is not a standard choice, isolate its contribution to the reported improvements?

Q2. Could the authors report results without warmup for PRO, PRO-MAX, and the strongest baselines?

Q3. Could the authors discuss on the feasibility (in real environments) of this warmup phase and whether during this phase it is possible that clients see data from "future" tasks?

Q4. The proposed approach combines feature-space rehearsal, class-level statistics, and representation-drift compensation, which have been explored in previous continual and federated learning methods (as far as I can see from related work). What are the method novelty compared to existing approaches such as AF-FCL and Hybrid Rehearsal, besides the integration of these mechanisms?

7 comments

We appreciate your valuable feedback. Our responses follow below.

Q1. Architecture and training fairness

A1. Thank you for this remark. To address this concern, we separate architecture from historical supervision. Original baseline configurations are not fully tabulated. Table R7 uses 20 replay objects and PRO's adapter, encoder schedule, normalization, and proximal policy with FedCBDR replay.

Table R7. Architecture and training comparison at 20 replay objects (paper Table 4).

ConfigurationFedCBDR FAAPRO FAAPRO-MAX FAA
Original settings38.2543.2846.12
Same architecture and training43.9343.2846.12

Under this control, FedCBDR improves by 5.68 points. PRO is then 0.65 below it and PRO-MAX 2.19 above it. These results indicate that architecture and optimization explain much of the original margins. Several factors change together, so this does not isolate adapter capacity. The PRO-MAX margin concerns the complete system.

Separately, the placement control tests whether historical losses reach the adapter. PRO obtains 43.28 FAA and 31.54 AF versus 38.64 and 41.38 for detached features replayed after the adapter. The differences are +4.64 FAA and -9.84 AF (paper Tables 13 and 14). Replay before the adapter lets historical losses constrain it, whereas detached features replayed after it bypass this gradient path. Section 4 distinguishes this result from the architecture comparison.

Q2. Performance without warmup

A2. Thank you for requesting this comparison. We remove SSL warmup but keep the same architecture, training policy, 20 replay objects, and encoder schedule (Table R8).

Table R8. FAA with and without warmup at 20 replay objects (paper Table 4).

MethodWarmup FAANo-warmup FAADrop
FedCBDR43.9338.255.68
PRO43.2836.187.10
PRO-MAX46.1239.027.10

In this setting, PRO and PRO-MAX lose 1.42 points more than FedCBDR. Their margins accordingly change from -0.65 to -2.07 and from 2.19 to 0.77, respectively. Warmup benefits them more under this schedule. Although PRO-MAX retains a small mean advantage, paired uncertainty is unavailable, so we do not claim statistical separation. As noted in Section 4, encoder training remains brief. Longer training would change both learning opportunity and computation.

Q3. Warmup feasibility and future data

A3. Thank you for asking about practical access. Yes, main warmup uses each client's full unlabeled local pool, including inputs from later tasks. This models institutions holding unlabeled data before annotations arrive. Appendix D.3 distinguishes this from inputs arriving with each task.

To examine more limited access, Table R9 restricts warmup to inputs at local position 0.

Table R9. FAA with full-pool and position-0-only warmup (paper Table 21).

MethodFull-pool FAAPosition-0 FAADrop
FedCBDR38.9638.100.86
PRO43.2842.700.58
PRO-MAX46.1245.350.77

Under this restriction, PRO and PRO-MAX retain margins of 4.60 and 7.25 points. This control restricts unlabeled warmup inputs and retains labeled incremental data, while still using SSL and the original settings. By contrast, A2 removes warmup after matching architecture and training, explaining the different FedCBDR values.

However, clients can begin with different tasks, so local position 0 does not exclude every globally future class. The result supports robustness under this local restriction. Section 4 distinguishes warmup on the full pool, warmup at position 0, and no warmup.

Q4. Novelty beyond existing methods

A4. Thank you for these references. AF-FCL uses flow-based assessment and generative replay. Hybrid Rehearsal combines latent exemplars, reconstruction, and generation. Section 2 and Appendix A explain how our policy trains and maintains shared class evidence.

Within our framework, the policy places rehearsal before the adapter, rebuilds available moments under the aggregated encoder, and applies server MOVE only without completed WRITE. Fresh WRITE takes priority because it uses updated coordinates. Table R10 reports sequential controls.

Table R10. Mean FAA and AF for sequential rehearsal controls (paper Table 13).

ConfigurationFAAAF
FedProto+BR39.8438.34
PRO-Point42.9632.10
PRO43.2831.54
PRO-MAX46.1226.18

The FAA increments are 3.12, 0.32, and 2.84 points, so feature sampling alone does not explain the full gain. The benefit also depends on configuration. PRO-MAX without spread reaches 43.46 versus 46.12 at the default (paper Table 34).

For comparison, HR obtains 32.18 FAA and 49.76 AF at its declared operating point (paper Table 1). This comparison evaluates complete methods, so it does not isolate one mechanism or match all resources. Our novelty claim therefore concerns the tested training and maintenance policy. Comparisons using the same settings for all relevant predecessors remain incomplete.

Q5. Privacy with sparse classes

A5. Thank you for raising the singleton case. When a client has one example of a class, its uploaded mean equals that example's feature representation. Keeping raw inputs local therefore provides no formal privacy guarantee. In particular, a variance floor or noise added during later replay cannot protect a mean already disclosed. This exposes an individual feature, although it does not establish unique reconstruction of the raw input.

For this reason, Section 5 and Appendix J.4 limit the privacy statement to communication of class statistics. We do not provide counts of singleton uploads or contributors, or an attack evaluation. Secure aggregation, minimum-count suppression, and differential privacy offer different protections. However, these are not established properties of the evaluated system. A stronger privacy claim would therefore require an implemented mechanism and a stated threat model.

Reviewer_mBSuOpenReview ↗

Assessment of the Rebuttal and Remaining Concerns

Assessment of the rebuttal and remaining concerns

I appreciate the authors' substantial efforts to address my concerns and the additional experiments included in the revised manuscript.

Q1–Q2 (Architecture and warmup). The new controlled comparisons are valuable. However, they also confirm that architecture and training choices explain a substantial portion of the original performance improvements. Under matched conditions, PRO performs slightly worse than FedCBDR, while PRO-MAX retains a smaller advantage (+2.19 FAA points). Without warmup, this advantage decreases further (+0.77 points). Moreover, the new comparisons do not provide uncertainty estimates (as far I can see), and equivalent architectural controls are not reported for other strong baselines. Therefore, my original concerns about the attribution and robustness of the performance improvements are only partially resolved.

Q3 (Warmup feasibility). I appreciate the clarification regarding access to unlabeled future-task data and the additional position-0-only experiment. I consider this concern largely addressed, provided that the corresponding assumptions and limitations remain explicit in the manuscript.

Q4 (Novelty). The revised discussion better distinguishes the proposed approach from related work, and the new ablations help establish the usefulness of the individual design choices. However, my concern regarding methodological novelty remains. In particular, feature rehearsal, memory reconstruction, and drift compensation have substantial precedents, including closely related drift-estimation mechanisms. The main contribution therefore appears to be in their coordination within a federated learning framework. While this integration may be useful, the experimental evidence does not yet convincingly establish its advantages over the closest existing approaches under comparable architectures and training conditions.

Privacy. I appreciate the explicit acknowledgment of the potential disclosure of individual feature representations through singleton class statistics. The revised manuscript appropriately limits its privacy claims, although the actual exposure risk remains unquantified.

Overall, the rebuttal represents a meaningful improvement. However, my two main concerns regarding experimental fairness and methodological novelty remain only partially resolved.

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

This paper studies FCIL where, at the same round, a class can be current for one client, old for another and unseen for a third (they call it asynchronous class exposure, ACE). PRO keeps a diagonal Gaussian per class before the adapter, replays it with class-balanced sampling, and recomputes the stats with the aggregated encoder at task boundaries. PRO-MAX adds a MOVE step that shifts old class means using drift estimated from current data on each client, averaged at the server. Experiments are on six image, text and graph benchmarks, and PRO-MAX is best on all of them.

Critical strengths
  1. The per-client view of class roles (current / historical / global-only) is a useful way to think about heterogeneous FCIL.
  2. Good set of controls, especially the zero / shuffled / uniform transport variants for MOVE and the replay-source swap in Table 4.
  3. 5 seeds, three modalities, and low overhead compared to the generative replay baselines.
Critical weaknesses
  1. Novelty is limited. Gaussian feature replay is well known (PASS, FeTrIL, and in FL, FCLPF (Yoo & Park, ACCV 2024) and HGP (Salami et al., 2024)). MOVE is basically SDC with K-NN and averaging across clients. ADC (CVPR 2024) and EFC (ICLR 2024) do drift compensation from current data and are not cited. The theory in App. I is mostly generic Lipschitz/triangle inequality bounds and doesn't guide the design.

  2. ACE is never measured. ACE(r) is defined in eq. 21 but no values are reported, and the support sweep also removes data, so heterogeneity and data size are confounded. PRO-MAX gains about the same over PRO in the aligned setting (+2.86) as in the heterogeneous one (+2.84), so MOVE does not seem specific to ACE.

  3. I think the frozen SSL encoder explains a lot of the results. PRO loses only 0.58 FAA with roughly 5x fewer labels, and with 80% support it is actually better than its own aligned result (44.26 vs 43.86). The encoder trains only 2 of 20 rounds per task, and it's not said if baselines use the same schedule. The warmup also sees future-task inputs from the local pool. A frozen encoder + NCM baseline, and baselines with the same freezing, would help a lot.

  4. There is no pretrained-model FCIL methods (PILoRA, LANDER, HGP), which is where most recent work is. THUCNews-10 (5 tasks x 2 classes) is easy, and on Cora the std is around 4, so the gain over MOTION is within noise. It is also unclear what WRITE re-encodes. If clients keep old inputs, that's exemplar storage and the comparison to exemplar-free methods isn't fair.

  5. The appendix says several times that the "supplied record" does not specify key things (permutation rules, whether MMD is squared, whether the optional adapted bank was on in the main runs). The compute ledger also doesn't match the per-run times: TinyImageNet lists 790 GPU-h, but Table 38 implies about 1430 h. No code is released.

  6. In eq. 13, σ=max⁡(s−μ⊙μ,ϵ)\sigma = \max( s - \mu \odot \mu , \epsilon ) is a variance, but eq. 4 uses it as a std, so a square root is missing. In eq. 7 the mixing weight comes from batch counts drawn from the same mixture, which is circular. The main text has a lot of hedging, so it's hard to tell what is claimed. Typos: "We call this phenomenon as", "class aware replay", "modality specific", and "class incremental" vs "class-incremental".

Questions for authors
  1. What data does WRITE re-encode, only current data or stored past inputs? How often does a class get MOVE instead of WRITE?
  2. Can you report the realized ACE(r) values, and run a sweep that changes ACE with the total labeled data kept fixed?
  3. Do the baselines use the same encoder freezing? What does a frozen encoder + NCM get?
  4. Could you compare with ADC, FCLPF or HGP, and FedProj (Chunduru et al., TMLR 2026)? FedProj deals with forgetting global knowledge for classes missing locally in non-IID FL, which is close to the global-only case here.
3 comments

Thank you for your thoughtful comments. We address each point below.

Q1. WRITE inputs and MOVE frequency

A1. Thank you for asking. WRITE re-encodes examples under the aggregated encoder at task ends. Fresh WRITE overrides MOVE. For the remaining entries, PRO retains their values, while PRO-MAX can move their means. Client retention remains unverified, as noted in Appendix J.3. Server moments alone therefore do not establish exemplar-free clients.

For the frequency question, Table R1 counts class-round pairs in a simulation with current-input WRITE, an empty bank, synchronized positions, and 60% order perturbation. New entries and local alignment are excluded.

Table R1. WRITE and MOVE eligibility in the schedule simulation (paper Table 46).

OutcomeShare (%)
WRITE0.72
Eligible for MOVE9.95
Other89.33

Eligibility need not mean nonzero movement. If the complete feature function changes only in the first 2 of 20 rounds, this schedule bounds nonzero MOVE by 10% and boundary WRITE by 5%. Actual rates require logs.

Q2. Realized ACE and fixed data

A2. Thank you for requesting ACE measurements. The 40/50/60 settings are not ACE. Table R2 reports our exposure simulation with fixed ownership and participation, using 20 clients, 10 participants, 10 tasks, 20 rounds per position, 8 supporting clients per task, and 5 of 10 classes per client.

Table R2. Stream-average ACE from the exposure simulation (paper Table 45).

Order (%)Stream-average ACE ± SD
00.33815 ± 0.00356
300.25249 ± 0.00906
600.22996 ± 0.00837
900.22304 ± 0.00595

If more clients assign a class to the global-only role, disagreement can fall despite greater order perturbation. Separately, Table R3 gives training FAA.

Table R3. FAA under different task-order perturbations (paper Table 29).

Method0%30%60%90%
FedCBDR43.2042.6838.9635.18
PRO43.7643.6443.2840.84
PRO-MAX46.6246.4846.1243.96

FAA losses of 8.02, 2.92, and 2.66 points support lower sensitivity to ordering (Table R3). However, these records are unpaired, so the requested curve needs ACE and performance from the same runs with labeled samples and ownership fixed. MOVE gains of 2.86 aligned versus 2.84 heterogeneous support both settings (paper Table 1), without establishing increasing benefit with ACE.

Q3. Encoder schedule and frozen NCM

A3. Thank you for this remark. Table R4 matches architecture and training at 20 replay objects.

Table R4. Architecture and training comparison at 20 replay objects (paper Table 4).

ConfigurationFedCBDR FAAPRO FAAPRO-MAX FAA
Original38.2543.2846.12
Same architecture and training43.9343.2846.12

In this control, matching uses PRO's adapter, encoder schedule, normalization, and proximal policy with FedCBDR replay. PRO trains its encoder on current inputs for 2 rounds per local task, then freezes it. Rehearsal trains the adapter and classifier. FedCBDR gains 5.68 points. PRO is then 0.65 below it and PRO-MAX 2.19 above it. These differences involve several factors changing together. We examine placement and refresh separately in Table R5.

Table R5. Mean FAA and AF for placement and global refresh controls (paper Tables 13 and 14).

ControlFAAAF
Post-adapter replay38.6441.38
No global refresh40.8636.82
PRO43.2831.54

Placement and refresh add 4.64 and 2.42 FAA points (Table R5). The requested frozen encoder + NCM baseline remains unreported.

Regarding label counts, the review's 0.58-point loss is the change from aligned to heterogeneous in paper Table 1. Support, coverage, and order change together, so it does not isolate label efficiency. Paper Table 21 separately shows a 0.58-point loss from restricting warmup inputs while retaining labeled incremental data. Comparing 44.26 at 80% support with 43.86 aligned also changes coverage and order (paper Tables 28 and 1). Paper Table 28 gives 44.26/43.92/43.28/39.86 at 80/60/40/20% support.

Q4. Novelty, related methods, and theory

A4. Thank you for these references. Section 2 and Appendix A discuss PASS, FeTrIL, SDC, ADC, EFC, and FCLPF. Building on this prior work, our policy links historical adapter gradients, global WRITE, and MOVE without fresh WRITE. PRO-MAX's 46.12 versus 43.32 with zero transport supports it (paper Table 15). However, new accuracy comparisons with ADC, FCLPF, HGP, and FedProj are not reported. HGP, PILoRA, and LANDER use different representations or data access. For the theory, Eqs. (26) and (27) explain the gradient path, while Eq. (57) relates transport error to calibration support. Transport remains approximate without a full-training guarantee.

Q5. Pretrained baselines and benchmark scope

A5. Thank you for raising this concern. Table R6 includes larger streams (FAA, mean ± SD).

Table R6. Selected benchmark results, reported as FAA mean ± SD (paper Table 18).

DatasetComparatorPRO-MAXMean gap
CLINC150 (FedSeIT)59.74 ± 1.1467.84 ± 0.988.10
ogbn-arxiv (MOTION)44.38 ± 2.4651.92 ± 2.187.54
Cora (MOTION)60.14 ± 4.2166.18 ± 3.666.04

PRO-MAX has a higher mean on Cora without established statistical separation. For the other concerns, pretrained comparisons remain unreported (A4), and storage access and its fairness need verification (A1).

Q6. Reproducibility and compute accounting

A6. Thank you for noting these gaps. Appendix J.4 lists unresolved retention, permutation, repair, adapted bank, normalization, and baseline settings. For MMD, Eq. (39) defines the squared form, but its displayed transformation is unverified. At 5 seeds and 2 settings, paper Table 40 gives 1,434 GPU-hours (143.4 × 5 × 2), versus 790 in paper Table 44. Reconciling these totals requires job logs. In addition, a code link and full configurations remain outstanding.

Q7. Equations and writing clarity

A7. Thank you for checking. Eq. (13) includes the square root for standard deviation. For Eq. (7), quotas must precede sampling to avoid circularity. The exact rule still needs confirmation, as noted in Appendix J.4. For presentation, the revision uses “We call this phenomenon asynchronous class exposure,” “class-aware,” “modality-specific,” and “class-incremental” consistently.

Dear Reviewer GsMy, we sincerely thank you for the time and effort you have dedicated to reviewing our work. We would greatly appreciate your thoughts on our response and revised manuscript and would be happy to discuss any remaining questions. We believe this exchange would help us refine the paper and guide further improvements to our work.

Comments on the paper

1 comment

We thank reviewers GsMy, mBSu, and w1pk for their time and thoughtful comments. Our contribution concerns how historical replay trains the adapter and remains useful as the encoder changes. We appreciate your recognition of the class roles, experimental controls, and evaluation across modalities.

The ordering remains unchanged at 20 and 100 replay objects under each method's original settings (Table 3). With the same architecture and training policy, PRO is 0.65 FAA points below FedCBDR and PRO-MAX is 2.19 above it (Table 4). Without warmup, the margins become -2.07 and 0.77 points (Table 4). These mean differences show that architecture, training, and initialization matter. They do not establish statistical separation. Main warmup includes unlabeled inputs from later tasks. Table 21 restricts input access while retaining warmup.

Placement and refresh improve FAA by 4.64 and 2.42 points within the evaluated framework (Tables 13 and 14). PRO-MAX exceeds zero transport by 2.80 points (Table 15). These controls support the training and maintenance policy. The order sweep supports lower sensitivity to ordering (Table 29). However, the ACE measurements from the exposure simulation are unpaired with training metrics (Table 45), so they do not establish that benefits increase with ACE. Sections 2 and 4 and Appendix J clarify this scope and the distinction from earlier methods.

Appendix J records the WRITE access boundary and the retention and implementation details still requiring verification. Communicating class statistics alone provides no formal privacy guarantee. The individual responses give the full comparisons and remaining limitations. We would be happy to clarify further questions during the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.