{"id":"7ec0aa60-7944-4dc3-a273-8e61f7c19d9f","arxiv_id":"2607.19384","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SUM projects client and task adaptation vectors to remove directional interference during server aggregation, improving federated class-incremental learning accuracy without client-side changes.","lead":"A server-side method that orthogonalizes client and task updates during federated aggregation, removing overlapping or conflicting directions before merging, reports large accuracy gains on federated class-incremental benchmarks. If genuine, it offers a way to improve distributed continual learning without changing client-side training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported FAA may rely on oracle task-ID selection: Appendix A.3 admits the inference module assumes task identity, but main experiments never validate a router, so the 22% FCIL gain and >Joint results may be task-incremental, not class-incremental.","rationale":"The weakest condition for the central claim is the evaluation protocol. The paper's central assertion is empirical: up to 22% improvement over prior FCIL methods and >Joint performance. That assertion is only meaningful if SUM is evaluated as a class-incremental model without task IDs. Section 4.4 builds one inference model per task, and Appendix A.3 explicitly admits the assumption. The main text omits how FAA is computed; it only describes the task splits and reports FAA. The theorem issue is secondary: even if Theorem 1's proof is not valid, the empirical results could still support the method; conversely, if the evaluation uses oracle task IDs, the empirical results are not comparable to single-model baselines. This is therefore the most load-bearing concern. It can be settled by a task-agnostic inference evaluation. The reader already identified this as the weakest assumption; I agree. The appropriate verdict remains conditional until that test is run.","tokens_in":33527,"tokens_out":6996,"duration_ms":67065,"concrete_test":"Implement SUM with the same backbone and CIFAR-100 β=0.05 split, but replace oracle per-task module selection with a task-agnostic router: for each test sample, compute the logit/softmax score of all task masks (or use a single head over all classes after merging) and report end-to-end FAA. Compare this number to the reported 96.32 and to LoRM/Joint under identical protocol. If the task-agnostic FAA is substantially lower (e.g., >5 points) or no longer beats Joint, the central claim is not supported for class-incremental inference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that SUM achieves up to 22% FAA improvement and even surpasses centralized joint training—depends on how FAA is computed. Section 4.4 constructs one inference model per task (theta_infer^(k) = theta_pre + m_k ⊙ s_k·tau_uni), and Appendix A.3 explicitly states: 'The inference-ready module assumes that the relevant task adaptation vectors can be selected at inference time.' The main text never specifies how task identity is obtained in the reported numbers; it only says '10 sequential class-split tasks' and reports FAA after the last task. If evaluation used oracle task IDs to activate the correct module for each task's test split, SUM is being compared under a task-incremental/multi-head protocol, while baselines such as EWC, LwF, CCVR, and LoRM are single models evaluated class-incrementally. Under strict class-incremental evaluation (no task IDs, one classifier over all seen classes), the per-task masks cannot be selected, and the reported near-zero forgetting (Fig. 10) and >Joint performance would not be expected. The proposed mitigation—a key-value retrieval mechanism in Appendix A.3—is not implemented or evaluated in the main experiments. Thus the headline numbers may reflect the task-selection premise rather than the geometric surgery itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SUM (Surgery & Merge), a server-side aggregation framework for Federated Class Incremental Learning (FCIL). Updates from clients and tasks are treated as adaptation vectors in parameter space; spatial SUM projects each client update against the others within a round, and temporal SUM projects the current task vector against previously stored task vectors. The refined vectors are then merged, and a compact inference module is built via sparsification, sign-consensus, and element-wise masking. The authors claim up to 22% FAA improvement over prior FCIL methods, robustness to unreliable clients, and no additional client-side overhead. A theorem is stated asserting that SUM merging reduces the loss relative to standard merging under smoothness and convexity assumptions, with a proof built on a transfer of the PCGrad guarantee.","tokens_in":33906,"tokens_out":3706,"duration_ms":40482,"significance":"If the empirical claims hold under a fair FCIL protocol, the work is significant: it addresses a real problem (spatio-temporal interference in FCIL) with a purely server-side mechanism, preserves standard client communication, and reports consistent gains across vision and language benchmarks. The ablation study and robustness analyses are useful. However, the paper currently lacks code and error bars, and the evaluation protocol for inference-time task selection is not clearly specified. The theoretical guarantee, as stated, is not sound because the proof transfers a PCGrad result that applies only to conditional projection (negative inner products) to the unconditional projection in Eq. (5). These issues must be resolved before the central claims can be accepted.","major_comments":[{"comment":"The evaluation protocol for FAA is not specified with respect to task identity. Appendix A.3 states that the inference-ready module 'assumes that the relevant task adaptation vectors can be selected at inference time,' and the main text does not describe how this selection is performed in Tables 2–3. If oracle task IDs are used to choose the correct module for each test sample, SUM is evaluated in a task-incremental/multi-head setting while baselines such as EWC, LwF, CCVR, and LoRM are single-model class-incremental learners. This would make the reported >Joint results and the 22% improvement an artifact of the protocol rather than of the geometric surgery. The authors must report exactly how inference was conducted, and provide results under strict class-incremental evaluation (or with the proposed retrieval mechanism actually implemented).","section":"§5.1, Tab. 2, Appendix A.3"},{"comment":"Theorem 1 as stated claims L(θSum) ≤ L(θstandard) unconditionally for convex losses with L-Lipschitz gradients. The proof maps Eq. (5) to a PCGrad update, but PCGrad's guarantee in Yu et al. (2020) holds only for the conditional projection used by PCGrad, which subtracts pairwise components only when the inner product is negative. SUM's Eq. (5) subtracts projection components for all nonzero inner products, including positive ones. Appendix B.5 implicitly acknowledges this by downgrading the result to a 'local descent-bound interpretation' under a signal–interference condition, but the main theorem still states an unconditional inequality. This is a load-bearing theoretical claim and must be corrected: either state and prove a conditional theorem with explicit assumptions, or remove the theorem and present the analysis as a heuristic. Note also that for two identical vectors, Eq. (5) can","section":"§4.1, Theorem 1, Appendix B.4"},{"comment":"No error bars or standard deviations are reported for the main FAA results. Given the large claimed improvements and the fact that SUM has several per-dataset hyperparameters (z_thr, λ_S, k_pct in Tables 12–19) that appear to be tuned per distribution level, it is important to establish that the gains are not due to hyperparameter selection or run-to-run variance. At minimum, report means and standard deviations over multiple seeds, and state the number of runs. This is especially important because Table 7, which does include errors, shows a large variance (±9.6) for SUM under unreliable-client settings.","section":"§5.1, Tables 2–3"},{"comment":"The projection in Eq. (5) is applied to all pairwise interactions, including positive (aligned) ones. In federated learning, clients working on the same task are expected to share a common signal direction; removing that shared direction from every client's update may discard useful task information and can even reverse the aggregate update when multiple clients are aligned (for n equal vectors, each refined vector becomes (2−n)v). The paper's own ablation in Fig. 3 shows that positive-only projection underperforms the no-surgery baseline. This does not invalidate the empirical results, but it means the method's success cannot be attributed to the projection alone without a more careful analysis of when aligning versus decorrelating helps. The authors should clarify what property of the update distribution makes full orthogonalization beneficial, and ideally provide an experiment isolati","section":"§4.1, Eq. (5)"}],"minor_comments":[{"comment":"The term 'Final Averaged Accuracy (FAA)' is used throughout but never formally defined. Please provide the exact formula (e.g., average accuracy over all tasks after the last task, using which inference rule).","section":"§5.1"},{"comment":"The last displayed equation before Eq. (30) states the stochastic deviation is 'O(η_eff,i) in norm', while the text says O(η_eff,i^2) in mean-square. This is confusing; please be consistent about the quantity being bounded.","section":"Appendix B.3"},{"comment":"The figure appears garbled in the PDF (many literal '/uni0000...' sequences), making the ablation curves unreadable. The source figure must be embedded properly.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is plausible and the server-only property is attractive, but I am concerned that the evaluation protocol may not be a fair class-incremental comparison. If the inference-ready modules are selected using ground-truth task IDs, the comparison against single-model baselines is invalid. The theory also needs to be either fixed or explicitly downgraded to a heuristic. I would be willing to reconsider after seeing the revised protocol and concrete evidence on task-agnostic inference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The core idea is genuinely clean: treat client updates and accumulated task vectors as adaptation vectors in the same parameter space, and at the server do Gram-Schmidt-style removal of components that overlap with other updates before merging. That gives you a purely server-side method with zero client overhead and no replay, which is a real practical advantage for FCIL. The empirical sweep is broad — vision and language, multiple heterogeneity levels, architectural robustness, ablation of each component, hyperparameter sensitivity, efficiency table. The gains over LoRM and others are large and consistent, and the sharpness/margin analysis helps explain why projection could beat plain joint training. This deserves credit.\n\nThe soft spots are in two places, both load-bearing. First, Theorem 1 is not proven as stated. The proof maps SUM to a PCGrad-style update and then inherits PCGrad's descent guarantee, but PCGrad's guarantee is conditional — it projects only when gradients conflict, and the proof uses that condition. SUM subtracts the projection for every pair, including positively aligned vectors. If two updates point the same way, SUM discards their shared component entirely, and your own positive-only ablation (Fig. 3) shows that removing positively aligned components hurts. So the unconditional claim L(SUM) ≤ L(standard) cannot be right in general; Appendix B.5 quietly downgrades it to a 'descent-bound interpretation' under a signal–interference condition that never appears in the theorem statement. That mismatch needs to be fixed, either by proving a real bound with the actual algorithm or by stating the theorem with its true assumptions.\n\nSecond, the inference protocol is under-specified. Appendix A.3 admits that the inference-ready module assumes the relevant task adaptation vectors can be selected at test time. The main text never says how that selection is performed in the reported FAA numbers. If the numbers use oracle task IDs to pick the correct mask for each task's test split, SUM is being evaluated under a task-incremental/multi-head protocol while baselines like EWC and LwF get a single class-incremental model. That comparison would be unfair, and the headline 22% gain would not survive a strict class-incremental evaluation without a validated router. The key–value retrieval idea is proposed as a mitigation but not evaluated. This is an empirical protocol issue, not a circularity issue — the method isn't fitting to test labels — but it has to be clarified before the results can be trusted.\n\nMinor issues: no error bars on the main tables, no code/data despite the strong claims, and the server-side O(TD) storage is honestly discussed. The citation pattern looks fine and the writing is clear.\n\nBottom line: this is a serious paper with a plausible idea and broad experiments, but the theory and the evaluation protocol both need substantial work. I'd send it to a careful referee — likely major revision — and I'd want to see code, error bars, and a straight answer on how task identity is handled before believing the >Joint numbers. I'd bring it to a reading group for the method discussion, but I wouldn't cite it yet.","headline":"Clean server-side projection idea for FCIL with large reported gains, but the theory is overclaimed and the inference-time task-selection protocol is under-specified; worth a careful referee, not worth believing yet.","tokens_in":34329,"tokens_out":4450,"would_cite":false,"duration_ms":43416,"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":"Update surgery beats joint training in federated continual learning.","keywords":["federated learning","class-incremental learning","catastrophic forgetting","adaptation vectors","model merging","gradient surgery","server-side aggregation","multi-task learning"],"falsifier":"Run the same FCIL protocol but evaluate inference with a single shared model (e.g., the spatially aggregated global model without per-task module selection) on test samples whose task labels are concealed; if the accuracy gain over baselines largely disappears, then the improvements hinge on task-selection rather than on the surgery operator itself.","tokens_in":33410,"feed_emoji":"✂️","tokens_out":3178,"duration_ms":27854,"temperature":0.7,"texified_at":"2026-08-05T21:33:09.490288+00:00","pith_summary":"This paper tries to establish that the two intertwined sources of forgetting in Federated Class Incremental Learning (FCIL)—client heterogeneity (spatial) and sequential tasks (temporal)—can be handled together by a single server-side operation: projecting each update vector to remove components aligned with other updates before merging. The authors argue that standard aggregation naively sums updates whose directional interference causes redundant and destructive parameter changes, and that surgically stripping those coupled components yields a cleaner global model. They claim this improves final accuracy by up to 22% over prior FCIL methods across vision and language benchmarks, and even exceeds centralized joint training on several datasets, all without modifying client training. If true, this would make interference-aware aggregation a cheap, drop-in replacement for plain averaging in continual federated systems.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":4080,"prompt_tokens":720,"completion_tokens":3360,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":720,"completion_tokens_details":{"reasoning_tokens":2731}},"feed_headline":"Update surgery beats joint training in federated continual learning","feed_subtitle":"Removing aligned components from client and task updates at the server prevents forgetting without extra client work.","key_machinery":"The load-bearing identity is the projection-based surgery operator: for each adaptation vector $v_i$, compute $\\hat{v}_i = v_i - \\sum_{j\\neq i} \\frac{v_i \\cdot v_j}{\\|v_j\\|^2} v_j$, then aggregate with $\\hat{v}_i$ instead of $v_i$. This operator is applied to client updates within a round (Spatial SUM) and recursively to task vectors that share a common pretrained base (Temporal SUM), making the temporal version a sequential orthogonalization. The same refinement is supported by Theorem 1, which states that under convexity, Lipschitz-smooth losses and a common base, the refined merge has loss no higher than the naive sum. The inference-ready module then compresses the refined task vectors into a unified direction plu","core_discovery":"The paper's central discovery is that spatio-temporal catastrophic forgetting can be reduced by treating client and task updates as adaptation vectors and removing, at aggregation time, each vector's components that are aligned with the others (Eqs. 5 and 9). Spatial SUM applies this within a round to client updates; causal online temporal SUM applies it sequentially across task vectors anchored to a common base model. The refined vectors are then merged, and compact inference modules are built via sparsification, sign consensus, and task-specific masks. Empirically, SUM reports the highest Final Averaged Accuracy on six vision and two language benchmarks, outperforming all baselines and som","pith_inferences":["The strongest practical gains depend on the assumption that the correct task-specific module can be selected at inference; the paper does not report results under a fully class-incremental protocol where task identity is unknown, so the headline numbers may overstate the benefit in truly task-agnostic deployment.","The core surgery operator could be adapted to other settings that merge parameter updates, such as model merging of independently trained models or federated personalization, where interfering directions are known to hurt.","Because the refinement is purely server-side, one could test it as a drop-in addition to existing FL and FCIL pipelines by simply replacing the aggregation rule, making the claimed improvement directly reproducible."],"forward_implications":["If the claim holds, any federated continual system can adopt interference-aware aggregation without changing client code, replay buffers, or extra communication.","SUM's improvement over the centralized joint baseline suggests that directional conflict resolution during aggregation can compensate for the loss of central data access.","The reported O(TD) server memory can be compressed with top-k/low-precision storage, so the method remains tractable for long task sequences.","The method's robustness to corrupted clients suggests that projection surgery also filters malicious update directions, not just heterogeneity."],"fun_headline_variants":["Server-side vector surgery curbs federated forgetting","Trim aligned updates to stop forgetting in federated continual learning","Geometric update surgery boosts federated continual learning","No client extra work: server aligns updates to defeat forgetting","Remove aligned components from updates to prevent cross-task interference"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The inference-ready module assumes that the correct task (or set of tasks) can be identified for each test input; if task identity is unavailable at test time, the reported accuracy gains may not materialize because the evaluation protocol becomes task-incremental rather than class-incremental.","fun_headline_variants_meta":{"raw":{"variants":["Server-side vector surgery curbs federated forgetting","Trim aligned updates to stop forgetting in federated continual learning","Geometric update surgery boosts federated continual learning","No client extra work: server aligns updates to defeat forgetting","Remove aligned components from updates to prevent cross-task interference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1233,"prompt_tokens":757,"completion_tokens":476,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":400}},"tokens_in":501,"tokens_out":476,"duration_ms":4888,"temperature":1.0,"reasoning_tokens":400,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T09:14:13.678892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same FCIL protocol but evaluate inference with a single shared model (e.g., the spatially aggregated global model without per-task module selection) on test samples whose task labels are concealed; if the accuracy gain over baselines largely disappears, then the improvements hinge on task-selection rather than on the surgery operator itself.","supporting_citations":[],"review_version":1}