{"id":"79540fb7-edcf-4670-b919-8746c542542e","arxiv_id":"2509.06219","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"MCIGLE combines aligned multimodal graph features, Fourier feature extraction, and recursive least squares to reduce forgetting in exemplar-free class-incremental graph learning.","lead":"This paper proposes MCIGLE, a model for learning new classes over time from graphs with text and image features, without storing old training examples. It combines feature alignment, Fourier-based feature extraction, and recursive least squares, reporting accuracy and forgetting numbers on four public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"C-RLS non-forgetting mechanism is mathematically unsupported: Eq. 9–10 are not a valid batch recursive least-squares update, so the main claim lacks a mechanism.","rationale":"The reader's weakest assumption correctly identifies the C-RLS update as the load-bearing mechanism: the paper claims that recursive updates achieve joint-training-level performance without historical data, and the empirical results are the only evidence offered for this. My stress-test sharpens this concern by showing that the equations as written are not even a valid batch RLS update: the gain formula in Eq. 10 has incompatible dimensions for N_k > 1 and omits the required matrix inverse, and the forgetting factor β is not shown to be part of any well-defined objective. The paper provides no derivation, no code, and no hyperparameters, so the numerical claims in Table 1 and Figures 2–3 are not independently checkable. The malformed optimal transport equation in Eq. 4 and the omission of the recursive autocorrelation/cross-correlation update definitions further reduce confidence, but they are secondary. The proposed synthetic test would settle the central question: if Eq. 9–10 do not reproduce the exact solution of a β-weighted joint objective, then the non-forgetting property is not established and the paper's main claim loses its mechanism. This does not change the reader's recommendation to reject, since the evidence is insufficient as presented.","tokens_in":11067,"tokens_out":4429,"duration_ms":49493,"concrete_test":"Run a numerical two-phase experiment with synthetic data. Phase 1: X_1 (N_1×d), Y_1. Phase 2: X_2, Y_2. Apply Eq. 9–10 exactly as written (with any natural reading of the denominator, e.g., treating it as scalar or as matrix inverse) and compare to the exact minimizer of the joint objective J(W)=β||Y_1−X_1W||²_F + ||Y_2−X_2W||²_F + γ||W||²_F. If the two updates disagree for N_2>1, the claimed joint-training equivalence is false. Repeat with β=1 and γ=0 to isolate whether the issue is the forgetting factor or the batch dimension. Report phase-1 accuracy after the update in all cases.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central non-forgetting claim (Section 2.3: recursive updates 'without storing historical data and achieving performance comparable to joint training') is not supported by the equations given. The objective in Eq. 8 is written only for the current phase (X_{M,k}, Y^k_train); the autocorrelation/cross-correlation updates and the exact loss being minimized online are never specified. Eq. 9–10, as stated, are not a valid recursive least-squares update for a batch of N_k samples. For the usual forgetting-factor objective sum_i β^{k-i}||Y_i−X_iW||² + γ||W||², the gain is P_{k−1}X_k^T(βI + X_kP_{k−1}X_k^T)^{-1}, a d×N_k matrix with an N_k×N_k inverse; Eq. 10 instead writes a scalar denominator β + XΦ^{-1}X^T. Unless N_k=1, the dimensions do not match, and no inverse is indicated. Moreover, because β<1 exponentially down-weights old data and no old labels are available, the recursion cannot be the exact minimizer of an objective that preserves old classes. If this fails, MCIGLE's non-forgetting module is an online linear head trained only on the current phase, and the low forgetting reported in Table 1 has no mechanism. Table 1 also lacks error bars, task setup, code, and hyperparameters, so the empirical claim cannot be independently checked.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MCIGLE, a framework for exemplar-free class-incremental graph learning on multimodal data. It combines a multimodal feature processing module based on optimal transport, a Fourier-based periodic feature extraction module, a 'Non-Forgetting Mainstream Module' using Concatenated Recursive Least Squares (C-RLS), and a residual fitting enhancement module. The framework is evaluated on four public datasets (COCO-QA, VoxCeleb, SNLI-VE, AudioSet-MI) against five baselines, with tables and figures reporting accuracy and forgetting metrics. The central claim is that C-RLS recursively updates weights without storing historical data and achieves performance comparable to joint training, thereby enabling exemplar-free class-incremental learning.","tokens_in":11481,"tokens_out":3950,"duration_ms":48949,"significance":"The problem addressed—exemplar-free class-incremental learning on multimodal graph-structured data—is timely and relevant, and the paper targets a real gap in the literature. The proposed architecture combines several plausible components, and the experimental section attempts a broad comparison. However, the significance is conditional on the validity of the C-RLS non-forgetting mechanism, which is the paper's core contribution. Since the mathematical formulation of C-RLS is internally inconsistent and the empirical results are not reproducible from the information given, the paper does not currently establish its central claim. No code, proofs, or detailed experimental setups are provided, so the reported gains cannot be independently verified.","major_comments":[{"comment":"The C-RLS update is dimensionally invalid for the stated setting. Eq. (10) writes the gain k_k with a scalar denominator β + X_{M,k}Φ_{k-1}^{-1}X_{M,k}^T, but for a batch of N_k samples this term is an N_k×N_k matrix and the gain should involve its inverse; the given expression is only valid for N_k=1. The appeal to Sherman–Morrison is also not justified for batch updates. More importantly, the claimed non-forgetting property is not derived: since β is introduced as a forgetting factor, β<1 exponentially down-weights earlier data, and no old labels or features are available to correct the update. The assertion that the recursion 'achiev[es] performance comparable to joint training' is therefore unsupported and in tension with the equations as written. This is the load-bearing mechanism of the paper.","section":"§2.3, Eqs. (8)–(10)"},{"comment":"The optimal transport alignment objective is not a valid OT program. In Eq. (4), the cost term (1−λ1)L(x_i, y_j, x_i′, y_j′)P_{i′j′} depends on P twice, and the summation over four index sets i,i′,j,j′ is not a standard transport cost. No marginal constraints on P are specified, the regularization term εW(P) is undefined, and no algorithm for computing P* is given. As a result, the 'multimodal feature processing' module is not concretely specified and cannot be implemented or evaluated from the text.","section":"§2.1, Eq. (4)"},{"comment":"The empirical evaluation lacks essential reproducibility information. Table 1 reports no error bars, number of runs, or statistical significance tests. The task schedule (number of phases, class order, number of classes per phase), dataset splits, graph construction, and hyperparameters (λ1, λ2, γ, β, learning rates, number of layers) are not provided. Without this information, the claimed improvements over baselines cannot be distinguished from noise, and the ablation figures (Figures 2 and 3) do not have labeled axes or numeric values. A rigorous comparison is necessary to support the paper's central empirical claim.","section":"§3, Table 1"},{"comment":"The residual fitting module's 'phase-wise label exclusivity' is asserted without justification. Zero-padding the residual with 0_{N0:k−1×dy,k−1} presupposes a consistent label dimension across phases, but in class-incremental learning new classes are added and the label space changes. The claim that this construction 'prevents error propagation from earlier phases' is not derived, and the mechanism by which the compensation stream improves old-class retention is unclear. This weakens the completeness of the framework description.","section":"§2.4, Eqs. (11)–(13)"}],"minor_comments":[{"comment":"The acronym MCIGLE is never expanded in the text.","section":"Abstract"},{"comment":"The framework figure contains garbled mathematical expressions and unlabeled components; it is not useful for understanding the architecture.","section":"Figure 1"},{"comment":"The captions mention 'TA, UCL, DGR' while the text describes 'knowledge distillation, parameter regularization, memory replay'; these should be aligned.","section":"Figure 3"},{"comment":"The statement that 'the only exception is Acc on AudioSet-MI, where CavRL performs better, which is expected given its specialization' is not a scientific justification and should be supported by analysis or removed.","section":"§3"},{"comment":"Some references lack venues or are clearly preprints; please ensure all citations are complete and accurate.","section":"References"}],"recommendation":"reject","confidential_remarks":"The paper is likely a preliminary draft; the core algorithmic contribution is not mathematically well-defined, and the empirical section lacks the detail needed for verification. Even though the problem is relevant, the current manuscript does not meet the standard for publication in a serious journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"MCIGLE is a pipeline paper: it bolts together a graph feature module, an OT alignment step, Fourier features, and a recursive least squares head, then reports numbers on four datasets. The parts are mostly borrowed, the pipeline is new as a combination, and the benchmark table is new. I'll give credit for that: the problem is real, and the multi-channel design with a residual compensation stream is a reasonable way to attack multimodal exemplar-free CIL on graphs. The ablations, though only described in prose, suggest the authors actually tried substitutions for the Fourier encoder and the RLS head. The citation to DS-AL for phase-wise label exclusivity is appropriate.\n\nThe soft spots are serious. The optimal transport objective in Eq (4) is not a valid OT program: the second term is quadratic in P, with P appearing on both sides, and no marginal constraints or transport polytope are defined. More importantly, the C-RLS update in Eqs (9)-(10) is dimensionally wrong for a batch of N_k samples: the gain should be a d x N_k matrix with an N_k x N_k inverse, not a scalar denominator. As written, it reduces to an online linear head trained on current-phase data. With a forgetting factor beta < 1, old data are explicitly down-weighted and no old labels are available, so the claimed equivalence to joint training has no derivation; the paper simply asserts non-forgetting. The stress-test note is right that this is load-bearing. Table 1 has no error bars, no task split details, no hyperparameters, and no code, so the headline empirical claim cannot be independently checked.\n\nWho is this for? Maybe a reader wanting a quick map of modules that could be combined in this area. Not anyone who wants a trustworthy method. It deserves a thorough rewrite, not a quick accept. I would not cite it in its current form, and I would not bring it to reading group except as a cautionary exercise in spotting dimensionally invalid math.\n\nFor peer review: I would desk reject in the current state, or at most send it back with a request for code, error bars, and a corrected derivation of the RLS update. As is, it is not ready for serious refereeing.","headline":"An integration of existing modules with fresh benchmark numbers, but the central non-forgetting claim rests on an invalid RLS update and unverifiable experiments.","tokens_in":11877,"tokens_out":1945,"would_cite":false,"duration_ms":23010,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MCIGLE claims that an exemplar-free, multimodal graph learner can add new classes without storing old data and still match joint training on accuracy and forgetting.","keywords":["class-incremental learning","exemplar-free","multimodal learning","graph neural networks","recursive least squares","catastrophic forgetting","optimal transport","continual learning"],"falsifier":"Train MCIGLE on a sequence of classes, then retrain the same architecture on all classes jointly from scratch; if joint training clearly beats MCIGLE on old-class accuracy, the claimed joint-training equivalence fails. Also, run a variant with the feature extractor unfrozen after phase one; if forgetting spikes, the method's non-forgetting property depends on freezing.","tokens_in":10975,"feed_emoji":"🔄","tokens_out":4477,"duration_ms":45731,"temperature":0.7,"pith_summary":"This paper proposes MCIGLE, a method for class-incremental learning on multimodal graphs that does not store any old training samples. It claims that by aligning text and visual features and updating a linear classifier through recursive least squares, the model can add new classes while keeping accuracy on old classes high. Experiments on four public datasets report higher accuracy and lower forgetting than several baselines. The authors argue the approach approaches the performance of training on all data at once.","feed_headline":"Recursive least squares lets graphs learn new classes without old data","feed_subtitle":"Aligns text and vision features, updates a linear head recursively, and reports lower forgetting on four datasets.","key_machinery":"Concatenated Recursive Least Squares (C-RLS): a recursive update rule for a linear classifier that maintains an inverse autocorrelation matrix via the Sherman-Morrison formula, adjusting weights with a gain term as new phases arrive. Because it only needs current-phase features and labels, it avoids storing old data. The residual fitting module adds a nonlinear correction stream trained on the current phase's prediction error.","core_discovery":"The central claim is that exemplar-free class-incremental graph learning can be made practical by combining four modules: a multimodal feature processor that aligns visual and textual node features via optimal transport, a Fourier-based periodic feature extractor, a non-forgetting linear head updated with Concatenated Recursive Least Squares (C-RLS), and a residual fitting network that compensates for the linear head's underfitting. The C-RLS module maintains recursive autocorrelation and cross-correlation matrices with a forgetting factor and uses the Sherman-Morrison formula to avoid storing historical data. The paper reports that MCIGLE consistently outperforms five baselines across four","pith_inferences":["The forgetting factor beta is a key hyperparameter; its optimal value likely depends on task similarity, and sweeping it may reveal a capacity-forgetting tradeoff the paper does not analyze.","The C-RLS update's 'joint-training equivalence' is asserted, not derived; a derivation or counterexample would clarify whether the method is truly non-forgetting or just a well-tuned online linear head.","The framework could be extended to other modalities or node-level tasks like link prediction, provided the alignment step generalizes.","If the feature extractor is allowed to update after the first phase, the C-RLS guarantee likely breaks; freezing the backbone may be essential to the method."],"forward_implications":["If the claim holds, exemplar-free CIL can match joint training in graph node classification without memory overhead.","Multimodal graph data (text and visual) can be handled in one incremental framework without replay.","The method gives a practical recipe for privacy-sensitive continual learning where old data cannot be kept.","The Fourier extractor reduces parameter count compared to CNNs or MLPs while capturing periodic patterns.","The residual stream can correct linear-head underfitting, improving performance on complex samples."],"supporting_citations":[{"why":"Supplies the recursive least squares update and Sherman-Morrison formula used in the Non-Forgetting Mainstream Module.","marker":"[8]"},{"why":"Defines the class-incremental learning setting and the challenge of learning without historical data.","marker":"[13]"},{"why":"Provides the survey-level framing of class-incremental learning and its variants.","marker":"[25]"},{"why":"Dual-stream analytic learning baseline that the residual compensation stream extends or competes against.","marker":"[27]"},{"why":"Multimodal graph learning context and the neural architecture search approach that motivates graph-oriented continual learning.","marker":"[3]"},{"why":"Fourier analysis network used in the Periodic Feature Extraction Module.","marker":"[7]"}],"fun_headline_variants":["No old data? No problem for graph learning","Graphs master new classes, zero memory of old ones","Multimodal graphs learn new classes without old data","Recursive least squares frees graphs from memory constraints"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The recursive least squares update is assumed to preserve old-class knowledge even though it is only fitted to current-phase data with a forgetting factor that down-weights old information and no old labels to correct it.","fun_headline_variants_meta":{"raw":{"variants":["No old data? No problem for graph learning","Graphs master new classes, zero memory of old ones","Multimodal graphs learn new classes without old data","Recursive least squares frees graphs from memory constraints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1141,"prompt_tokens":600,"completion_tokens":541,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":344,"completion_tokens_details":{"reasoning_tokens":478}},"tokens_in":344,"tokens_out":541,"duration_ms":6311,"temperature":1.0,"reasoning_tokens":478,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:50:48.065050+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train MCIGLE on a sequence of classes, then retrain the same architecture on all classes jointly from scratch; if joint training clearly beats MCIGLE on old-class accuracy, the claimed joint-training equivalence fails. Also, run a variant with the feature extractor unfrozen after phase one; if forgetting spikes, the method's non-forgetting property depends on freezing.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the recursive least squares update and Sherman-Morrison formula used in the Non-Forgetting Mainstream Module."},{"cited_title":"In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion","cited_arxiv_id":null,"evidence_quote":"Defines the class-incremental learning setting and the challenge of learning without historical data."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"Dual-stream analytic learning baseline that the residual compensation stream extends or competes against."},{"cited_title":"In: Proceedings of the ACM Web Conference 2022","cited_arxiv_id":null,"evidence_quote":"Multimodal graph learning context and the neural architecture search approach that motivates graph-oriented continual learning."}],"review_version":1}