{"id":"50a36b03-c5ab-4331-b168-32b8fff6efbb","arxiv_id":"2412.18926","paper_version":2,"verdict":"REJECT","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A federated class-incremental learning method that uses condensed synthetic exemplars, produced by gradient and feature matching plus a shared VAE, to reduce catastrophic forgetting.","lead":"This paper proposes ECoral, a method for federated class-incremental learning that replaces stored real images with condensed synthetic exemplars to reduce catastrophic forgetting. It combines dataset condensation, a shared variational autoencoder, and knowledge distillation, and reports accuracy gains over baselines on CIFAR-100, Tiny-ImageNet, and Caltech-256.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported CIFAR-100 50-task numbers are mutually inconsistent: abstract says 32.42%, Section 6 says 91.00% average and 64.40% final, and Figure 3's y-axis cannot display 91%. The central long-term superiority claim is therefore not supported by the reported evidence.","rationale":"The paper's central assertion is an empirical one: ECoral's condensed exemplars outperform sampled real exemplars across task sequences, including long-horizon settings. The most load-bearing vulnerability is therefore not primarily the theoretical transfer assumption highlighted in the reader's weakest_assumption, but the fact that the reported numbers cannot all be true simultaneously. The abstract and Section 6 give contradictory 50-task accuracies, and Figure 3 cannot visually contain the stated average. Table 1's sub-blocks also appear shifted, with ECoral's Aincre_last at sigma=0.5 equal to its Aavg at sigma=0.2. If these numbers are not reproducible, then the empirical comparison underpinning the central claim collapses, regardless of whether the condensation objective is theoretically sound. The reader's rationale already notes these inconsistencies, so there is substantial agreement on the destination, but the reader's formal weakest_assumption focuses elsewhere; I therefore mark agreement as partial. I recommend keeping the reader's REJECT verdict rather than altering it, because this concern strengthens rather than redirects the rejection. The concrete test of obtaining raw per-task accuracies is decisive: either the reported values can be recovered from the raw curves, in which case the manuscript has a severe reporting error that still undermines the current version, or they cannot be recovered, in which case the central claim lacks stated evidence. Either way, the paper as written does not support its headline conclusion.","tokens_in":16830,"tokens_out":4604,"duration_ms":41579,"concrete_test":"Obtain the raw per-task accuracy arrays for ECoral and each baseline for the 50-task CIFAR-100 experiment (from the authors or an independent reimplementation), recompute Aavg and Alast using the definitions in Section 5.4, and compare the results against the abstract's 32.42% and Section 6's 91.00%/64.40%. If the recomputed Aavg is not equal to 32.42% and also not equal to 91.00%, or if neither final accuracy value is reproducible from the same per-task curves, then the reported long-term superiority of ECoral is not currently supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that ECoral's condensed exemplars outperform sampled real exemplars in federated class-incremental learning—rests on the experimental tables and figures. That evidence is internally inconsistent, making the claimed superiority unverifiable as presented. In the abstract, ECoral is reported to achieve 32.42% on CIFAR-100 with 50 tasks, while Section 6, in discussing Figure 3, states that 'in the more challenging 50-task setup, ECoral maintains a strong performance with an average accuracy of 91.00%, outperforming BiC (90.50%) and iCaRL (90.00%). By the final task, ECoral achieves 64.40% accuracy.' These two accounts differ by roughly 59 accuracy points for the same setting. Moreover, the 50-task panel of Figure 3 has a y-axis maximum of 80%, so neither 91.00% average nor 64.40% final accuracy can be read from the plotted curves. This is not a matter of metric ambiguity: both values are labeled as average accuracy and final accuracy, respectively. Additional inconsistencies appear in Table 1. In the A subtable at sigma = 0.5, ECoral's Aavg is 49.17 and Alast is 27.97; in the Aincre subtable, ECoral's Aincre_last at sigma = 0.5 is 43.52, which equals the Aavg value reported for sigma = 0.2 (43.52). This suggests that the sub-table columns are shifted relative to their headers. Because the numeric basis for the headline improvement of 5.32% and 10.29% cannot be pinned down, the empirical side of the central claim fails before one even reaches the theoretical question of whether gradient matching on a re-initialized local condensation model (Eqs. 3-5) transfers to the FedAvg global model. No code, random seeds, or error bars are provided, so an independent reader cannot reconstruct which number is correct. The authors also state in the conclusion that ECoral's advantage decreases on more complex datasets, which further weakens the general claim of seamless integration and enhancement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Exemplar-Condensed federated class-incremental learning (ECoral), a replay-based method that replaces sampled real exemplars with condensed synthetic exemplars generated via gradient matching, feature matching, and a meta-knowledge contrastive loss, supported by a shared VAE for client-wise feature disentanglement. The claimed contribution is that ECoral improves accuracy and reduces forgetting compared to existing FCIL baselines, while also being privacy-friendly because it stores condensed data instead of raw images. The manuscript reports experiments on CIFAR-100, TinyImageNet, and Caltech-256 under non-IID settings, with multiple evaluation metrics and an ablation study.","tokens_in":17191,"tokens_out":2807,"duration_ms":25866,"significance":"The conceptual direction—using dataset condensation rather than sample selection for replay in federated class-incremental learning—is timely and potentially useful, especially for privacy and memory-constrained settings. If the empirical claims were reliable, the method would offer a clear improvement over existing rehearsal baselines and could be integrated into other approaches. However, the current manuscript contains severe internal inconsistencies in the reported experimental numbers, which makes the central empirical claim unverifiable as written. The paper also relies on an unexamined assumption that locally condensed exemplars remain effective after FedAvg aggregation. The idea is worth pursuing, but the evidence presented does not currently support the headline conclusions.","major_comments":[{"comment":"The reported 50-task CIFAR-100 numbers are mutually contradictory. The abstract states ECoral achieves 32.42% accuracy with 50 tasks, while Section 6 states that in the 50-task setup ECoral achieves an average accuracy of 91.00% and a final-task accuracy of 64.40%, outperforming BiC (90.50%) and iCaRL (90.00%). These are about 59 percentage points apart for the same setting. Furthermore, the rightmost panel of Figure 3 has a y-axis maximum of 80%, so neither 91.00% average nor 64.40% final accuracy can be read from the plotted curves. This is a load-bearing inconsistency because the paper's central claim of long-term superiority is based on this experiment. The authors must correct the abstract, Section 6, and Figure 3 so that all reported values agree, and they should explain the source of the discrepancy.","section":"Abstract and Section 6 (CIFAR-100 50-task results)"},{"comment":"Table 1's sub-tables (A, Aincre, Aa) appear misaligned relative to their headers, which casts doubt on the claimed improvements. For example, in the A sub-table at σ=0.5, ECoral's Aavg is 49.17 and Alast is 27.97, but in the Aincre sub-table, ECoral's Aincre_last at σ=0.5 is 43.52, which exactly equals the Aavg value reported for σ=0.2 (43.52). This suggests that the columns have been shifted, making it impossible to determine the true values for each metric and condition. Since the quantitative claims of improvement (e.g., 5.32% and 10.29% over baselines) are computed from these numbers, the table must be restructured with clearly matching metrics, and the authors should verify that each reported value is placed under the correct header.","section":"Table 1 (CIFAR-100, 10-task results)"},{"comment":"The core mechanism of ECoral is gradient matching performed on a local condensation model that is re-initialized whenever new classes appear and is trained only on real local images. The paper asserts that the resulting condensed exemplars will also be effective for the global model after FedAvg aggregation, but no analysis or control experiment is provided to justify this transfer. Given that the empirical results are currently inconsistent, the authors should at least include a sanity check—for example, a comparison of ECoral's condensed exemplars against random real exemplars under identical memory budgets in a single-client and a multi-client setting—to demonstrate that the condensation step actually improves global accuracy rather than merely matching local gradients. This would address the load-bearing assumption that locally condensed exemplars survive aggregation.","section":"Section 4.1, Eqs. (3)-(5)"}],"minor_comments":[{"comment":"There is a typo in the phrase 'Exemplar-Condensed federated class-increment al learning' where 'incremental' is split across a line break; this should be corrected to a single word.","section":"Abstract"},{"comment":"The number of classes per task for TinyImageNet is inconsistent: Section 5.2 states 10 tasks with 20 classes per task, while Table 2's caption says 'Tiny-ImageNet with 10 tasks (10 classes per task).' Please clarify the correct task/class split.","section":"Section 5.2 and Table 2 caption"},{"comment":"The cosine similarity formula is not formatted clearly; the denominator appears to be missing the multiplication sign and the temperature parameter τ is not defined in the surrounding text. Please rewrite the equation and define τ explicitly.","section":"Equation (8)"},{"comment":"Several figure labels contain 'uni00A0' artifacts (e.g., 'Task/uni00A0Step' in Figures 3 and 4). These should be replaced with proper spaces or non-breaking space characters.","section":"Figure labels"},{"comment":"The ablation table header (A, G, F, C, K) is not self-explanatory; the component names are only given in the caption. Consider adding a legend or using full component names in the header for readability.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistencies in the headline numbers (abstract vs. Section 6 vs. Figure 3) are severe enough that I would ask the authors to provide the raw experimental logs or rerun the experiments before the paper can be considered further. The issue is not a minor presentational glitch; it affects the interpretability of every quantitative claim. I would also encourage the editor to request the code or at least a detailed experimental configuration so that the results can be reproduced. The conceptual idea is interesting, but the reported evidence must be made internally consistent and verifiable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is worth a look: instead of sampling exemplars for rehearsal in federated class-incremental learning, condense them via gradient matching, add feature matching and a contrastive compensation loss, and use a shared VAE to generate features for unseen classes. That is a real extension of dataset condensation into the federated continual setting, and the design choices are reasonable. The ablation, if the numbers are right, suggests each component adds something. So there is a salvageable method underneath.\n\nThe problem is that the empirical support collapses on inspection. The 50-task CIFAR-100 result is reported three ways: the abstract says 32.42%, Section 6 says average accuracy 91.00% and final accuracy 64.40%, and Figure 3's y-axis stops at 80% so neither 91 nor 64 can be read from the plot. These are labelled as the same setting. That is not a minor typo; it is the central scalability claim. Table 1 has a similar issue: in the Aincre subtable at sigma=0.5, ECoral's Aincre_last is 43.52, which is exactly the Aavg value listed for sigma=0.2, suggesting the subtable columns are shifted relative to their headers. No code, seeds, or error bars are provided, so an independent reader cannot reconstruct which number is correct. The conclusion also admits the advantage shrinks on more complex datasets, which undercuts the \"seamlessly integrated\" claim in the abstract.\n\nThe theoretical worry the reader raised — that gradient matching on a re-initialized local condensation model may not transfer to the FedAvg global model — is legitimate but not proven wrong; the paper simply does not analyze it. That lack of analysis is a soft spot, but it is secondary to the numerical inconsistencies.\n\nWho should read this? Researchers working on federated continual learning or dataset condensation. A serious referee could push the authors to resolve the internal contradictions, provide code and seeds, and re-run the 50-task experiment. The method may survive that process, but as written the current version does not support its own headline claims. My recommendation: send it to peer review with the expectation of major revision, not desk reject — the idea is timely and the flaws are fixable.","headline":"A plausible method for federated class-incremental learning, but the headline numbers contradict each other and the empirical case is not verifiable as written.","tokens_in":667,"tokens_out":713,"would_cite":false,"duration_ms":22615,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that federated clients can replace stored real images with condensed synthetic exemplars and forget less, reporting 49.17% average accuracy on CIFAR-100 and 27.97% on the last task.","keywords":["federated learning","class-incremental learning","dataset condensation","exemplar replay","catastrophic forgetting","non-IID data","knowledge distillation","federated continual learning"],"falsifier":"Run a controlled federated class-incremental experiment in which the global model is frozen during the condensation phase, then compare ECoral's condensed exemplars against an equal-sized set of randomly sampled real exemplars: if the condensed examples do not yield higher average accuracy and lower forgetting when the local gradient-matching model is decoupled from the global model, the central transfer claim is falsified.","tokens_in":16610,"feed_emoji":"🖼️","tokens_out":6063,"duration_ms":51337,"temperature":0.7,"pith_summary":"This paper claims that in federated class-incremental learning, a client can store fewer, denser training images by distilling them into condensed exemplars, and that these synthetic exemplars preserve more of the original training signal than an equal number of randomly selected real images. The proposed framework, ECoral, combines gradient matching, feature-relationship matching, and a contrastive compensation loss to make the exemplars informative, and it uses a shared disentanglement model to reduce the harmful heterogeneity that non-IID data introduces into the condensation process. The authors report that on CIFAR-100 with 10 tasks, ECoral reaches 49.17% average accuracy and 27.97% last-task accuracy, beating the strongest baseline iCaRL, which reaches 43.85% and 21.76%. If the claim holds, it would let federated systems rehearse past knowledge with less stored data and stronger privacy, since the stored exemplars are synthetic rather than raw images.","feed_headline":"Synthetic exemplars beat real replays in federated class-incremental learning","feed_subtitle":"ECoral's condensed memory images lift CIFAR-100 average accuracy to 49.17%, topping iCaRL's 43.85%.","key_machinery":"ECoral's load-bearing component is the condensed exemplar set $\\mathcal{M}_{\\text{cond}}$, a fixed-size set of synthetic images produced per client by three losses: gradient matching ($\\mathcal{L}_{\\text{cond}}$) aligns the gradients induced by condensed and real mini-batches in a local condensation model; relationship matching ($\\mathcal{L}_{\\text{rel}}$) keeps the feature relationships among condensed images consistent with those among real images; and meta-knowledge compensate matching ($\\mathcal{L}_{\\text{MKCL}}$) pulls each condensed feature toward unbiased class prototypes and away from others. The prototypes come from FINCH clustering on features generated by a Shared-VAE, a disentangled generative model that each client updates with global parameters, letting clients produce features for local and unseen classes. An adjustable memory allocation re-balances slots between old and new classes at each task. This machinery is supervised by knowledge distillation from the previous task's model, and the whole local update is aggregated with FedAvg.","core_discovery":"The central claim is that replacing sampled real exemplars with condensed synthetic exemplars in each client's fixed memory improves both average accuracy and retention in federated class-incremental learning. ECoral condenses local data online by matching the gradients that real and condensed images induce in a local condensation model, matching feature relationships between condensed and real images, and pulling condensed features toward unbiased class prototypes while pushing them away from other classes. A Shared-VAE, updated with global parameters, generates disentangled features for local and unseen classes, and FINCH clustering turns those features into prototypes; this is meant to counteract the meta-information heterogeneity that arises when each client condenses from a different non-IID slice. The paper reports consistent gains over seven baselines across CIFAR-100, TinyImageNet, and Caltech-256, with the largest margins under the most skewed non-IID partition, and it positions ECoral as a memory component that can be combined with existing replay methods.","pith_inferences":["We infer that ECoral's advantage should shrink as the local condensation model's gradient updates diverge from the global model's update path; a direct test would be to freeze the global model during the condensation phase and measure the resulting accuracy gap.","The privacy argument is stronger than the evidence: condensed images are called indecipherable, but dataset condensation can still encode identifiable cues about source images, so a formal privacy analysis would be needed to support the claim.","The paper notes its advantage decreases on more complex datasets, so we infer the method will be most attractive for small-image benchmarks and may need stronger distillation losses for high-resolution or fine-grained domains.","A testable extension is to shuffle class order across clients, since the current client-increment strategy follows a fixed old, in-between, and new grouping, and to check whether the condensation losses remain aligned with the global model."],"forward_implications":["A fixed per-client memory can rehearse more class knowledge when its slots hold condensed images instead of raw samples, because each synthetic image is optimized to carry task-specific gradient information.","Under strong non-IID skew (σ=0.2), ECoral's average accuracy is 43.52%, its largest margin over baselines, suggesting condensation helps most when clients see very different class distributions.","Ablations attribute the gain to all four components: full ECoral adds 12.45 percentage points over replay, with compensation matching alone contributing 5.01 percentage points.","Because condensed exemplars are synthetic and derived only from the client's local data, the method reduces the privacy exposure of rehearsal memory compared with storing raw images.","The method is designed as a plug-in memory module, so existing federated class-incremental approaches that use replay can swap in condensed exemplars without redesigning their training loop."],"supporting_citations":[{"why":"Supplies the gradient-matching objective that condenses exemplars by aligning model updates on condensed and real data.","marker":"[Zhao et al., 2020]"},{"why":"Contributes the online condensation framework with model re-initialization on new classes, adapted in Eqs. 4-5.","marker":"[Gu et al., 2024]"},{"why":"Provides FedAvg, the aggregation rule whose interaction with condensed exemplars is the core setting.","marker":"[McMahan et al., 2017]"},{"why":"Defines the iCaRL baseline and the exemplar-memory and distillation losses that ECoral builds on and compares against.","marker":"[Rebuffi et al., 2017]"},{"why":"Supplies the client-increment strategy (old, in-between, and new groups) used to simulate streaming federated data and the GLFC rehearsal baseline.","marker":"[Dong et al., 2022]"},{"why":"FINCH clustering generates the unbiased feature prototypes used in the meta-knowledge compensate matching loss.","marker":"[Sarfraz et al., 2019]"},{"why":"The beta-VAE disentanglement model becomes the Shared-VAE that generates class features for local and unseen classes to reduce heterogeneity.","marker":"[Higgins et al., 2017]"}],"fun_headline_variants":["Condensed exemplars top federated class-incremental learning","ECoral: synthetic replay wins in federated incremental learning","Condensed memory beats sampled replays in federated learning","Synthetic replay memory lifts federated class-incremental accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that gradients computed on a small local condensation model, which is re-initialized whenever new classes arrive and trained only on real local images, produce exemplars whose training signal still matches the federated global model's update directions after aggregation; the paper does not analyze how re-initialization or non-IID client updates affect that transfer.","fun_headline_variants_meta":{"raw":{"variants":["Condensed exemplars top federated class-incremental learning","ECoral: synthetic replay wins in federated incremental learning","Condensed memory beats sampled replays in federated learning","Synthetic replay memory lifts federated class-incremental accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000525,"raw_usage":{"total_tokens":2504,"prompt_tokens":880,"completion_tokens":1624,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":1555}},"tokens_in":496,"tokens_out":1624,"duration_ms":11891,"temperature":1.0,"reasoning_tokens":1555,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:19:44.169369+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled federated class-incremental experiment in which the global model is frozen during the condensation phase, then compare ECoral's condensed exemplars against an equal-sized set of randomly sampled real exemplars: if the condensed examples do not yield higher average accuracy and lower forgetting when the local gradient-matching model is decoupled from the global model, the central transfer claim is falsified.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"Provides FedAvg, the aggregation rule whose interaction with condensed exemplars is the core setting."},{"cited_title":"Fed- erated class-incremental learning","cited_arxiv_id":null,"evidence_quote":"Supplies the client-increment strategy (old, in-between, and new groups) used to simulate streaming federated data and the GLFC rehearsal baseline."},{"cited_title":"Efficient parameter-free clustering using first neighbor relations","cited_arxiv_id":null,"evidence_quote":"FINCH clustering generates the unbiased feature prototypes used in the meta-knowledge compensate matching loss."},{"cited_title":"beta-vae: Learning basic visual concepts with a con- strained variational framework","cited_arxiv_id":null,"evidence_quote":"The beta-VAE disentanglement model becomes the Shared-VAE that generates class features for local and unseen classes to reduce heterogeneity."}],"review_version":1}