{"id":"6a934ece-21fd-451c-9be2-c651abcb8094","arxiv_id":"2504.13691","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MEGA pre-trains a GNN with MAML-style second-order gradients over simulated task sequences plus knowledge distillation and replay, improving graph few-shot class-incremental learning without using query sets during incremental training.","lead":"This paper proposes a stricter evaluation protocol for graph few-shot class-incremental learning that forbids the use of query sets during incremental training, and introduces MEGA, a meta-learning framework that pre-trains graph neural networks on simulated task sequences to reduce forgetting. The authors report accuracy gains on four graph datasets, but the comparison baselines are adapted by the authors and code is not yet released.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The support-only incremental protocol may still leak query information through full-graph message passing; whether MEGA works without transductive query access is untested.","rationale":"The reader correctly identified the paper's weakest assumptions but ranked base-to-novel transfer first. I agree transfer is unproven, yet it is partially mitigated by the fact that base and novel classes are drawn from the same dataset and feature distribution. The transductive-leakage issue is more load-bearing because it attacks the paper's primary claimed contribution: a rigorous setting that excludes query sets during incremental training. If the full graph is used in the forward pass, then query node features and topology are part of the training computation, and the setting is not support-only in any information-theoretic sense. This also affects all baseline comparisons, since the baselines may or may not have been run under the same transductive conditions; the paper does not say. The concern is not that the authors are dishonest, but that an unspecified implementation detail could overturn the central claim. The proposed concrete test is inexpensive and would settle the matter. Secondary issues from the reader's review remain valid: the abstract's improvement range (2.60% to 7.43% on the first novel task) does not match Table III for DBLP (Task 1 improvement is +1.83%), the Reddit results lack error bars, and the authors themselves note that the adapted GFSCIL baselines 'preclude equitable performance comparison.' These issues reinforce a conditional verdict but do not replace the transductive-leakage concern as the central one.","tokens_in":22160,"tokens_out":6661,"duration_ms":74465,"concrete_test":"Retrain MEGA under an inductive support-only protocol on at least Amazon-Clothing and Cora-Full: for each incremental task, construct the computation graph using only support nodes and their K-hop neighbors, with all current and future query/test nodes excluded from message passing (or, equivalently, mask query-node features and remove query-node edges during training). Compare the resulting accuracy against Table III. If any Task t accuracy drops by more than the reported standard deviation relative to the transductive setting, then the support-only claim fails and the comparison protocol needs revision; if accuracy is unchanged, the transductive-leakage concern is refuted.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central novelty is a GFSCIL setting that 'excludes query sets' during incremental training, yet the method description never specifies whether the GNN forward pass is inductive or transductive. The implementation details in Section IV-A state only the network architecture and optimizer; no mention is made of masking query nodes, sampling support-only subgraphs, or otherwise preventing message passing from unlabeled query/test nodes. In standard graph node classification, a GNN layer aggregates features from all neighbors, so if the full adjacency matrix is used during incremental fine-tuning, representations of support nodes are computed using the features and topology of query nodes, and gradients flow through those computations. This means the model is not 'trained exclusively on the novel support sets' as asserted in the Problem Definition (Section III-A) and Figure 2; rather, it is trained with transductive access to the query nodes' input information, even if their labels are not used in the loss. The claimed distinction from HAG-Meta and GEOMETER, which the paper frames as eliminating query-set leakage, then reduces to withholding query labels rather than genuinely excluding query sets. If MEGA's gains come partly from this transductive information, the headline comparison under the 'more rigorous' setting is not a fair test of the support-only protocol, and the method's contribution is substantially weakened. This concern is concrete and resolvable, but the paper as written leaves it open.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces MEGA, a model-agnostic meta-learning framework for graph few-shot class-incremental learning (GFSCIL). The authors first propose a stricter protocol in which incremental training uses only novel support sets, without fine-tuning on novel query sets, to remove query-set leakage. MEGA has two main components: MCTF, a meta-training algorithm that samples ordered base-task sequences and updates initial parameters via the gradient of cumulative query losses through inner-loop adaptation (Eq. 2), and KDSIR, a continual-learning loss module combining knowledge distillation with per-class single-instance replay. The paper evaluates MEGA against GFSL, GCL, and adapted GFSCIL baselines on Amazon-Clothing, DBLP, Cora-Full, and Reddit, reporting that MEGA outperforms all baselines and that MCTF improves existing GCL methods when embedded.","tokens_in":22389,"tokens_out":9842,"duration_ms":81847,"significance":"If the empirical results hold, the paper makes two useful contributions: it highlights the query-set leakage in prior GFSCIL protocols and proposes a model-agnostic way to import GCL losses into meta-learned initializations. The ablation studies (Tables IV-VII) are a genuine strength: they isolate the contributions of MCTF, L_SIR, and L_KD and show additive gains, and Table VII demonstrates that MCTF can improve three different GCL baselines. The setting itself is a sensible step toward stricter evaluation. However, the main state-of-the-art claim is currently not fully supported: the authors adapt the GFSCIL baselines themselves and report no code, several tabulated differences overlap with standard deviations, and the Reddit results lack error bars. The paper therefore needs additional verification before the headline claims can be accepted.","major_comments":[{"comment":"The manuscript's central claim of a support-only incremental protocol is not operationalized: no statement is made about whether the GNN forward pass during incremental fine-tuning uses the full graph or a support-induced subgraph. Because standard GNN layers aggregate over all neighbors, using the full adjacency leaks query-node features and topology into support representations even if query labels are not in the loss. Please specify the masking or sampling procedure; if the full graph is used, the distinction from HAG-Meta and GEOMETER reduces to withholding labels rather than excluding query sets, and the experimental comparisons need to be redone under a genuinely inductive protocol.","section":"Section III-A and Section IV-A"},{"comment":"The authors adapt Geometer, HAG-Meta, and Mecoin to their new setting, acknowledge that this adaptation 'reveals a partial degradation in their catastrophic forgetting mitigation capabilities, thereby precluding equitable performance comparison,' and yet use these degraded versions in Table III to claim state-of-the-art performance. This is not a controlled comparison: the main SOTA claim is established against baselines that the authors themselves defined and weakened, with no released code to verify the adaptation. Report the original methods in their native settings as an upper reference, or restrict the SOTA claim to the unmodified GCL/GFSL baselines.","section":"Section IV-A and Section IV-B"},{"comment":"Several of the reported improvements are within one standard deviation of the best baseline, and Table XI provides no error bars for Reddit despite the statement in Section IV-A that all accuracy scores are means over five random seeds. For example, Amazon Clothing Task 9 shows MEGA at 49.11±1.75% versus Mecoin at 48.82±1.96%, a 0.29% gap, and Cora-Full Task 1 shows overlapping intervals (52.74±3.89% vs 45.90±6.29%). Add per-task significance tests or paired-seed comparisons, and provide error bars for Reddit, before claiming that MEGA 'consistently outperforms all baselines.'","section":"Table III and Table XI"},{"comment":"The claimed 'improvement of 2.60% to 7.43% in the first novel task' does not match the tables: DBLP Task 1 shows +1.83% in Table III and Reddit Task 1 shows +11.55% in Table XI. Correct the reported range and specify the baseline used for each improvement.","section":"Abstract and Conclusion"}],"minor_comments":[{"comment":"The SIR loss is written as CE(f_θ(x_b), f_θ(y_b)), which is not a valid cross-entropy expression; it should be CE(f_θ(x_b), y_b), with y_b as the label set. The same correction applies to Eq. (7).","section":"Equation (4)"},{"comment":"The text states MEGA consistently outperforms all baselines, but GAT-FT's base accuracy is higher on DBLP (48.08±2.07 vs 46.99±0.94) and Cora-Full (72.64±0.57 vs 58.08±4.09), as the caption itself notes; please reconcile the wording.","section":"Section IV-B vs Table III caption"},{"comment":"The inner-loop step is set to 1, but the complexity analysis in Section III-E and the ablation in Table VIII do not examine the effect of the number of inner steps m on the second-order gradient quality; a short sensitivity study would clarify the role of m.","section":"Table II"},{"comment":"The role of the buffer B_m in Eq. (5) is described only as 'cleared at the end of each epoch'; please clarify how the single-instance replay buffer is populated during meta-training and why the same buffer is not used in the outer-loop objective.","section":"Section III-D"},{"comment":"FTF-ER [45] is the authors' own prior method and serves both as a baseline and as the basis for MCTF-enhanced comparisons in Table VII; a brief independent description of FTF-ER's mechanism would help readers assess potential overlap with KDSIR.","section":"Related work"}],"recommendation":"major_revision","confidential_remarks":"The paper's proposed setting is a meaningful contribution, but the SOTA claim currently rests on author-adapted baselines and on an underspecified protocol (inductive vs transductive message passing). If the authors can release code and provide inductive masking plus significance-aware comparisons, the paper could be publishable. I also note that the large gap between MEGA and adapted GFSCIL baselines on Cora-Full (e.g., +19.72% on Task 1 vs FTF-ER) warrants scrutiny in revision, since it may reflect baseline degradation rather than the method's intrinsic gains."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a legitimate stricter setting and a modular meta-training component that shows internal consistency in ablations, but the 'no query leakage' claim is undercut by the absence of any inductive/transductive specification, and the headline numbers are softer than the abstract suggests.\n\nThe support-only incremental protocol is a real improvement over HAG-Meta/GEOMETER fine-tuning on query sets, and it is clearly described. MCTF is a reasonable adaptation of MAML to a task sequence, and showing that MCTF improves FTF-ER, MAS, and TWP (Table VII) is a useful plug-in result. The ablations in Tables IV-VI are consistent and support the additive value of MCTF and KDSIR. So the engineering story holds up internally.\n\nThe soft spots are real, and the transductive one is load-bearing. Section IV-A gives the network architecture and optimizer but never says whether the forward pass is inductive or transductive. If the full adjacency matrix is used during support-only fine-tuning, then query nodes' features and topology are aggregated into support node representations, and gradients flow through those computations. That means the model is not trained 'exclusively on the novel support sets' as claimed in Section III-A; it has transductive access to query inputs, even if labels are withheld. The comparison to HAG-Meta and GEOMETER is then not about removing query sets, just removing query labels. This is concrete and fixable—mask query nodes or sample support-only subgraphs—but until it is stated, the central contribution is ambiguous.\n\nOther issues are proportionate. The abstract claims 2.60-7.43% improvement on the first novel task, but DBLP's first task shows +1.83%. Reddit results have no error bars, and no significance testing is provided; some margins (e.g., Amazon Task 9) are within noise. The adapted GFSCIL baselines are admitted to be degraded, so the SOTA claim rests partly on baselines that had their anti-forgetting mechanisms removed. MCTF is close in spirit to prior meta-continual learning (Meta-CL, OML), which is not cited.\n\nThis paper is worth a serious referee. A major revision that specifies the graph inductive/transductive setting, releases code, adds error bars on Reddit, and either properly tunes adapted baselines or reframes them as ablations would make it a solid contribution to the graph continual learning subfield. As written, it is a plausible, internally consistent method with an unresolved leakage ambiguity and overstated headline numbers.\n\nRecommendation: revise-majorly rather than accept; do not desk-reject.","headline":"A genuinely stricter GFSCIL setting and a modular meta-training stage, but the no-query-leakage claim is compromised by unspecified transductive message passing, and the SOTA margins are less clean than the abstract suggests.","tokens_in":22980,"tokens_out":4036,"would_cite":false,"duration_ms":37765,"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":"Graph few-shot class-incremental learning models can avoid catastrophic forgetting if their meta-training stage simulates the incremental stage through an incremental second-order gradient and aligned loss functions, as the proposed MEGA…","keywords":["graph few-shot class-incremental learning","catastrophic forgetting","meta-learning","second-order gradient","graph continual learning","knowledge distillation","experience replay","node classification"],"falsifier":"Evaluate the incremental stage with message passing restricted to the support-induced subgraph only, so the model cannot see any unlabeled or test nodes before evaluation; if MEGA's large accuracy advantage over the baselines vanishes, the claim that the meta-learned prior alone drives forgetting mitigation is falsified.","tokens_in":21934,"feed_emoji":"🧠","tokens_out":7931,"duration_ms":63349,"temperature":0.7,"pith_summary":"Graph few-shot class-incremental learning (GFSCIL) asks a model to keep recognizing old classes while learning new classes from only a handful of labeled nodes. This paper claims that existing GFSCIL methods leak information by fine-tuning on the new classes' query sets, and that this masks catastrophic forgetting. It proposes a stricter setting in which incremental training uses only the few support nodes, and a meta-learning framework called MEGA that computes an incremental second-order gradient over a sequence of pseudo-tasks drawn from the base classes. The claim is that this produces an initialization whose behavior matches the incremental stage, so the model adapts quickly to new classes without forgetting old ones. On four graph datasets, MEGA is reported to outperform adapted baselines and to improve existing graph continual-learning methods when they are wrapped in its meta-training stage.","feed_headline":"Second-order meta-gradients stop graph models from forgetting","feed_subtitle":"Meta-training on pseudo-task streams makes new few-shot classes stick without erasing old ones.","key_machinery":"The central mechanism is the incremental second-order gradient, a variant of model-agnostic meta-learning's outer-loop update. Where standard meta-learning optimizes a single task's query loss, MEGA samples an ordered sequence of pseudo-tasks from the base label space, performs inner-loop adaptation on each, and computes the gradient of the cumulative sum of query losses across all tasks with respect to the initial parameters (Eq. 2). This injects second-order information about the whole task stream into the initialization. The second half of the machinery is behavioral alignment: the inner loop, outer loop, and incremental stage share the same loss components (cross-entropy, knowledge distillation, single-instance replay), so the update dynamics during meta-training match the dynamics during incremental fine-tuning. That alignment is what, according to the paper, activates the learned priors at test time.","core_discovery":"MEGA's core proposal is that catastrophic forgetting in GFSCIL can be mitigated by making the model's meta-training stage mimic the incremental-learning stage at the level of both the optimization trajectory and the loss functions. In the meta-training stage, the base class set $\\mathbb{Y}_0$ is randomly split into an ordered sequence of pseudo-tasks; for each task the model performs $m$ inner-loop steps on the support set, then the outer loop computes the gradient of the sum of query losses over all tasks seen so far with respect to the original parameters (Eq. 2). This 'incremental second-order gradient' accumulates curvature information across the task stream, unlike single-task meta-learning. The same loss terms — cross-entropy, knowledge distillation, and a single-instance replay buffer — are used in the inner loop, outer loop, and incremental fine-tuning, so the model's behavior in all three phases is aligned. The paper claims this alignment is what turns the meta-learned initialization into a usable prior for sequential novel classes, and it presents experiments on Amazon-Clothing, DBLP, Cora-Full, and Reddit showing accuracy gains of 2.60% to 7.43% on the first novel task, along with improvements when existing continual-learning losses (FTF-ER, MAS, TWP) are embedded into the same framework.","pith_inferences":["Editorial inference: If behavioral alignment is truly the active ingredient, the same incremental second-order gradient recipe should transfer to non-graph modalities, such as image or text few-shot class-incremental learning, where base classes can be reorganized into pseudo-task sequences; this is a direct test the paper does not run.","Editorial inference: The paper does not state whether incremental training propagates messages through the full graph (transductive) or only through the support-induced subgraph. If full-graph propagation is used, part of the reported advantage may come from seeing unlabeled nodes during training; an inductive evaluation would separate graph-structure effects from the meta-learned prior.","Editorial inference: The stricter support-only protocol invalidates direct comparison with prior GFSCIL numbers; adopting it as a community standard would require re-benchmarking existing methods, and some previously reported improvements may shrink once query-set fine-tuning is removed.","Editorial inference: The mechanism is described for node classification; a natural extension is to test the same MCTF wrapper on link prediction or graph classification tasks, where the definition of a 'task' over a label space differs and the transfer assumption may or may not hold."],"forward_implications":["Under the stricter support-only incremental protocol, MEGA is reported to keep higher accuracy on both base and novel classes than adapted GFSCIL baselines, GFSL methods, and GCL methods on four graph datasets.","The reported 2.60–7.43 percent improvement on the first novel task indicates that the benefit of the meta-learned prior appears immediately after the first incremental step, not only late in the stream.","Existing graph continual-learning methods (FTF-ER, MAS, TWP) show consistent accuracy gains when their loss functions are used inside MCTF, supporting the model-agnostic claim rather than a KDSIR-specific effect.","Ablations show that the meta-training framework alone raises the performance ceiling, while the knowledge-distillation and single-instance-replay terms add further forgetting resistance; removing any component degrades the final accuracy.","Performance stays roughly flat as the support set size $K$ varies from 1 to 7, suggesting the method does not depend on having several labeled nodes per new class."],"supporting_citations":[{"why":"Supplies the model-agnostic meta-learning algorithm whose outer-loop gradient MEGA extends into an incremental second-order accumulation over a task sequence.","marker":"[35]"},{"why":"One of the two existing GFSCIL methods that MEGA critiques for fine-tuning on novel query sets; serves as the primary baseline adapted to the stricter setting.","marker":"[16]"},{"why":"The other existing GFSCIL method, also adapted as a baseline; defines the prototype-representation approach that MEGA argues limits architectural flexibility.","marker":"[17]"},{"why":"First graph meta-learning method using the meta-learning algorithm; provides the support/query split configuration and the Meta-GNN baseline for few-shot node classification.","marker":"[8]"},{"why":"Prototypical networks are the architecture behind prior GFSCIL methods; MEGA's critique of PN-based loss coupling relies on this.","marker":"[18]"},{"why":"FTF-ER is one of the GCL methods embedded into MCTF in the versatility experiments; its improvement under MCTF supports the model-agnostic claim.","marker":"[45]"},{"why":"MAS is another GCL baseline and MCTF-embedded method; its performance collapse without MEGA and recovery with MCTF-MAS is a key ablation result.","marker":"[46]"},{"why":"TWP is the third GCL method used to demonstrate that MCTF improves existing continual-learning losses in the few-shot setting.","marker":"[47]"}],"fun_headline_variants":["Second-order gradients align meta and incremental learning to stop forgetting","Meta-learned gradients align tasks so new classes stick without forgetting","Aligning second-order gradients across phases stops catastrophic forgetting in GFSCIL","Meta-training on pseudo-tasks with second-order gradients prevents forgetting","Second-order gradient alignment makes few-shot classes stick in graph models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The base classes, randomly split into an ordered sequence of pseudo-tasks, must be a faithful stand-in for the future novel-task stream, because the meta-learned initialization is trusted to transfer from base-class gradients to genuinely new classes with different labels, features, and graph neighborhoods.","fun_headline_variants_meta":{"raw":{"variants":["Second-order gradients align meta and incremental learning to stop forgetting","Meta-learned gradients align tasks so new classes stick without forgetting","Aligning second-order gradients across phases stops catastrophic forgetting in GFSCIL","Meta-training on pseudo-tasks with second-order gradients prevents forgetting","Second-order gradient alignment makes few-shot classes stick in graph models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001053,"raw_usage":{"total_tokens":4473,"prompt_tokens":1051,"completion_tokens":3422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":3335}},"tokens_in":667,"tokens_out":3422,"duration_ms":20061,"temperature":1.0,"reasoning_tokens":3335,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:02:14.544201+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the incremental stage with message passing restricted to the support-induced subgraph only, so the model cannot see any unlabeled or test nodes before evaluation; if MEGA's large accuracy advantage over the baselines vanishes, the claim that the meta-learned prior alone drives forgetting mitigation is falsified.","supporting_citations":[{"cited_title":"Graph few-shot class-incremental learning,","cited_arxiv_id":null,"evidence_quote":"One of the two existing GFSCIL methods that MEGA critiques for fine-tuning on novel query sets; serves as the primary baseline adapted to the stricter setting."},{"cited_title":"Geometer: Graph few-shot class-incremental learning via prototype representation,","cited_arxiv_id":null,"evidence_quote":"The other existing GFSCIL method, also adapted as a baseline; defines the prototype-representation approach that MEGA argues limits architectural flexibility."},{"cited_title":"Meta- gnn: On few-shot node classification in graph meta-learning,","cited_arxiv_id":null,"evidence_quote":"First graph meta-learning method using the meta-learning algorithm; provides the support/query split configuration and the Meta-GNN baseline for few-shot node classification."},{"cited_title":"Prototypical networks for few-shot learning,","cited_arxiv_id":null,"evidence_quote":"Prototypical networks are the architecture behind prior GFSCIL methods; MEGA's critique of PN-based loss coupling relies on this."},{"cited_title":"Ftf-er: Feature-topology fusion-based experience replay method for continual graph learning,","cited_arxiv_id":null,"evidence_quote":"FTF-ER is one of the GCL methods embedded into MCTF in the versatility experiments; its improvement under MCTF supports the model-agnostic claim."},{"cited_title":"Memory aware synapses: Learning what (not) to forget,","cited_arxiv_id":null,"evidence_quote":"MAS is another GCL baseline and MCTF-embedded method; its performance collapse without MEGA and recovery with MCTF-MAS is a key ablation result."},{"cited_title":"Overcoming catastrophic forgetting in graph neural networks,","cited_arxiv_id":null,"evidence_quote":"TWP is the third GCL method used to demonstrate that MCTF improves existing continual-learning losses in the few-shot setting."}],"review_version":1}