{"id":"47c6294f-070f-4dd1-aac7-e9b752e9e7df","arxiv_id":"2505.10040","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"IPAL improves non-exemplar continual graph learning by combining prototype contrastive learning with PageRank-weighted prototypes, instance-prototype affinity distillation, and decision-boundary hard-example mining, outperforming prior methods on four node classification benchmarks.","lead":"The authors propose IPAL, a training method for graph neural networks that learns new tasks without storing old raw examples, by replaying class prototypes instead. It combines prototype contrastive learning, PageRank-weighted Gaussian prototypes, relational distillation, and hard-example mining, and reports higher average accuracy than prior continual graph learning methods on four benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof assumes the conclusion it seeks to establish: the claim that PCL's negative-sample term reduces previous-class mean shift (Eq. 21 area) is asserted, not derived, so the theoretical motivation for IPAL is unsupported.","rationale":"The reader's weakest_assumption identifies exactly the unproved premise in Theorem 1's proof: the assertion that PCL's negative-sample regularization keeps the mean of previous-task feature distributions close to its pre-update value. My reading of Appendix A confirms this is asserted rather than derived. The step from Eq. 17 (gradient w.r.t. current samples) to Eq. 21 (expectation over previous samples) requires an additional mechanism, such as Lipschitz continuity plus a bound on task similarity, that is not stated. The proof also does not connect the PCL gradient to a contraction of the old-class mean; it simply declares that the objective 'essentially regularizes' the mean shift. This is not an internal inconsistency in the sense of a false equation, but it is a genuine gap in the logical chain between the PCL loss and the claimed lower drift. The empirical comparison and ablation are internally consistent, and the paper includes a limitations section that is honest about scope, so a conditional acceptance is appropriate. My stress-test does not introduce a new concern beyond the reader's; it confirms the concern and specifies the precise textual location. Therefore I recommend no change to the reader's verdict: the paper should be accepted only if the theorem is rewritten as a conjecture or properly derived, and if a quantitative drift comparison is added to replace the qualitative Figure 1 claim. The proposed concrete test would settle whether the theorem is actually true, since a single counterexample in a simplified setting would invalidate the proof's premise.","tokens_in":14598,"tokens_out":2759,"duration_ms":27107,"concrete_test":"Construct a controlled two-task experiment with a linear encoder: task 0 has two classes with Gaussian features; task 1 has two new classes with a chosen mean direction relative to old classes. Train the encoder on task 0, then take a single gradient step on task 1 under L_PCL and under L_PR (Eq. 1) with identical initialization and step size. Measure the resulting old-class mean shift ||µ'_m − µ_m||_2 and trace ratio tr(Σ'^{-1} Σ_m). If, for some reasonable configuration (e.g., new-class means aligned with an old-class mean), PCL yields a larger old-class mean shift than PR, Theorem 1 is contradicted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical assertion is Theorem 1 (Appendix A, Eq. 15): PCL incurs less feature drift than PR. The proof hinges on the claim, around Eq. 21, that 'L_PCL essentially regularizes ||µ'_m − µ_m||^2_2, effectively alleviating feature shift.' This is not derived from the PCL objective. The gradient in Eq. 17 is computed for current-task samples x_t, while the mean shift in Eq. 21 is an expectation over previous-task samples (x_m, y_m), which are inaccessible. The proof would need to show how the negative-sample term for x_t—minimizing F_θt(x_t)^T µ_m—induces a contractive update on the previous-task feature means. No such argument is given; the sentence merely restates the desired conclusion. Similarly, the covariance argument that PR 'allows Σ_m to expand uncontrollably' is qualitative and not tied to the PR objective. Consequently, Theorem 1 is unproven: the claimed smaller drift for PCL does not follow from the equations presented. This matters because the paper explicitly motivates the entire IPAL design by this theorem ('To elucidate this phenomenon, we present Theorem 1 in Appendix A, along with a rigorous theoretical proof'). If the premise is false—e.g., PCL's negative-sample regularization can increase drift when current and old tasks are correlated—the theoretical justification collapses, leaving only the qualitative Figure 1 and the empirical Table 1. The empirical gains could stand independently, but the paper's central scientific claim about PCL's inherent drift advantage is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IPAL, a non-exemplar continual graph learning method built on prototype contrastive learning (PCL). The authors argue empirically and via a stated theorem (Theorem 1, Appendix A) that PCL induces less feature drift than conventional prototype replay (PR), and they introduce three components: Topology-Integrated Gaussian Prototypes (TIGP), which PageRank-weight node contributions when computing class prototypes; Instance-Prototype Affinity Distillation (IPAD), which regularizes instance–prototype affinity via Mixup-based synthetic features; and Decision Boundary Perception (DBP), which adds hard-example repulsion to the PCL objective. The method is evaluated on CS-CL, CoraFull-CL, Arxiv-CL, and Reddit-CL against regularization, rehearsal, and non-exemplar baselines, reporting improved average performance in Table 1, with ablations and parameter sensitivity studies in Tables 2 and Figure 6.","tokens_in":14929,"tokens_out":5817,"duration_ms":56058,"significance":"If the claims hold, IPAL is a competitive empirical contribution to non-exemplar continual graph learning, a setting where memory and privacy constraints rule out raw-example replay. The paper's strengths are the standard CGLB-style benchmark setup, the inclusion of several strong non-exemplar baselines (POLO, EFC), the clear ablation isolating TIGP, IPAD, and DBP, and the consistent average-performance gains in Table 1. However, the central scientific claim that PCL provably incurs less feature drift than PR is not supported by the proof in Appendix A; the key step is asserted rather than derived. Because the main text explicitly presents Theorem 1 as a rigorous proof that motivates the entire design, this gap is load-bearing and must be repaired, or the theoretical claim must be removed and reframed as an empirical observation.","major_comments":[{"comment":"Theorem 1 is not established by the proof given. The pivotal step is the sentence after Eq. (21): \"Eq. 2 regularizes the negative sample gradients in Eq. 17 by minimizing F_θt(x_t)⊤·µ_m, encouraging ∆θ to drive F_θt(x_m) closer to F_θt−1(x_m).\" Eq. (17) is a gradient with respect to the current-task feature f_θt = F_θt(x_t), whereas Eq. (21) is an expectation over previous-task samples (x_m, y_m) ∈ T_{0:t−1}, which are not available when training on T_t. No argument shows that optimizing the PCL objective on current-task samples contracts the previous-task prototype means or bounds ||µ′_m − µ_m||². The subsequent covariance claim that PR \"allows Σ_m to expand uncontrollably\" is qualitative and is not derived from the PR objective in Eq. (1). Consequently, the inequality D_KL(P^{t−1}∥P^t)_{PCL} < D_KL(P^{t−1}∥P^t)_{PR} does not follow from the equations presented. Because Section 1 presents this theorem as the rigorous justification for the PCL-based design, this gap needs to be repaired or the theorem must be removed and the motivation reframed as an empirical observation.","section":"Appendix A, Theorem 1 and Eq. (21)"},{"comment":"The empirical premise that \"PCL exhibits less pronounced drift than conventional PR\" is currently supported only by the qualitative visualization in Figure 1. No quantitative drift metric is reported for any dataset or task, such as the average prototype mean shift or an estimated KL divergence between the old and new feature distributions. Since Theorem 1 is the only formal support for this premise and that proof is incomplete, the paper should either report such a metric (e.g., on CS-CL and CoraFull-CL) or explicitly downgrade the claim to an observed trend that motivates the method rather than a demonstrated property of PCL.","section":"Section 5 / Figure 1"}],"minor_comments":[{"comment":"As typeset, the drift compensation and the TIGP generation/allocation appear inside the \"for epoch\" loop, so they would be executed every epoch and the memory buffer would be rewritten multiple times; move them outside the epoch loop to match the post-task descriptions in Sections 4.1 and 4.4.","section":"Algorithm 1, lines 9–10"},{"comment":"The notation \"idx = arg max ... == y_m\" is unclear: y_m is not defined for the prototype index m, and the comparison with the arg max result should be spelled out (e.g., by defining the prototype label and how the boolean mask is applied).","section":"Eq. (7)"},{"comment":"\"PyTorch 3.10 framework\" should read \"PyTorch with Python 3.10\" (or equivalent); PyTorch does not have a version 3.10.","section":"Section 5.1, Implementation Details"},{"comment":"The text says β is \"assigned an exceedingly small value\" in light of Figure 1, but the experiments use β = 0.1; please clarify whether 0.1 is considered small in this context or whether the reported results use a different value.","section":"Section 4.4, Eq. (11)"},{"comment":"The caption describes top/middle/bottom rows, but the figure contains multiple dataset columns; please annotate each subplot with its method and dataset, or arrange the panels by method, to remove ambiguity about which heatmap corresponds to which configuration.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are credible and the ablation is informative, so this could become a solid empirical paper. The main risk is the theoretical appendix: the proof of Theorem 1 is currently circular in the sense that it assumes the contractive effect it claims to establish. If the authors cannot provide a real derivation, they should remove the theorem and present the reduced-drift observation as an empirical motivation; I would not require a theorem for acceptance, but as written the unsupported theorem is a central claim and needs to be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real contribution here is empirical. IPAL combines three sensible ideas — PageRank-weighted Gaussian prototypes, instance-prototype affinity distillation with mixup, and entropy-based hard-example mining — inside a prototype contrastive loss, and it consistently beats prior non-exemplar methods on four standard graph benchmarks. The ablations isolate each component cleanly, and the gains over EFC and POLO look real. That alone makes this paper useful to anyone working on continual graph learning.\n\nThe soft spot is the theory, and the stress-test note is right. Theorem 1 claims PCL incurs less feature drift than PR, but the proof never delivers. The key step around Eq. 21 simply asserts that the PCL negative-sample term \"essentially regularizes\" the previous-task mean shift. It doesn't show how a gradient computed on current-task samples contracts the expectation over previous-task features. The covariance argument is equally hand-wavy. This is not a proof; it's a restatement of the desired conclusion. The paper explicitly motivates the whole design with this theorem, so this is a load-bearing flaw — though not a fatal one for the empirical claims.\n\nTwo smaller issues: the \"empirical finding\" that PCL drifts less than PR rests entirely on the qualitative Figure 1, with no quantitative drift metric. And no code is released, which undercuts reproducibility for a results-driven paper. Both are fixable.\n\nIn proportion: the empirical story stands on its own. The components are not wildly novel individually, but the combination is new relative to the cited literature, and the benchmark numbers are the message. If the authors rewrite the appendix as an empirical observation or a formal conjecture, add a measured drift comparison, and release code, this becomes a solid paper. Even as-is, it deserves referee time rather than a desk reject.\n\nI'd bring it to a reading group as a useful baseline paper, and I'd cite it if I were working on non-exemplar continual graph learning. The theory needs to be relabeled, not the whole paper thrown out.","headline":"Solid empirical paper with a flawed theory appendix: IPAL works on the benchmarks, but Theorem 1 should be a conjecture, not a theorem.","tokens_in":15449,"tokens_out":1456,"would_cite":true,"duration_ms":16877,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"IPAL claims that prototype contrastive learning drifts less than prototype replay, and builds a non-exemplar continual graph learner on that finding.","keywords":["non-exemplar continual graph learning","class-incremental learning","prototype replay","feature drift","prototype contrastive learning","graph neural networks","knowledge distillation","PageRank prototypes"],"falsifier":"Compute the mean feature of each old class before and after training on a new task under PCL and under standard prototype replay; if PCL's mean shift is not consistently smaller across task transitions, the theorem on feature drift is not supported by the data.","tokens_in":14401,"feed_emoji":"🕸️","tokens_out":6009,"duration_ms":62669,"temperature":0.7,"pith_summary":"The paper proposes a new method, IPAL, for continual node classification on graphs without storing or replaying raw examples from earlier tasks. Its central claim is that Prototype Contrastive Learning (PCL) suffers less feature drift than the standard Prototype Replay used by prior non-exemplar approaches, and the authors give a theorem to that effect. On top of PCL, IPAL adds PageRank-weighted Gaussian prototypes, relational distillation between instances and prototypes, and a decision-boundary module that uses hard examples as extra negatives. Across four node-classification benchmarks, the paper reports that IPAL outperforms existing methods and achieves a better balance between plasticity and stability.","feed_headline":"Exemplar-free graph learner tops four continual benchmarks","feed_subtitle":"Stores class prototypes, not raw data; balances absorbing new classes with remembering old ones.","key_machinery":"The carrying object is the PCL objective, a prototypical contrastive loss that treats each node embedding as a query, its own class prototype as the positive, and all other class prototypes as negatives. IPAL makes three modifications to it: TIGP replaces mean prototypes with PageRank-weighted Gaussian distributions, biasing class geometry toward influential nodes; IPAD distills the inner products between current-task instances and old-class prototypes, preserving relational structure rather than rigid feature distances; and DBP selects the top-entropy instances near class boundaries and adds them as extra negative pairs. Together these pieces carry the argument that old-class geometry can be protected without raw exemplars.","core_discovery":"The paper's central discovery is that replacing the cross-entropy objective of prototype replay with a contrastive objective aligns current-task instances with their class prototypes and actively repels instances from other classes, so old feature distributions move less when the encoder is updated. Theorem 1 formalizes this: under Gaussian feature distributions, the KL divergence between previous-task feature distributions before and after an update is smaller for PCL than for Prototype Replay. IPAL operationalizes the result with Topology-Integrated Gaussian Prototypes (TIGP), which weight class means and covariances by PageRank; Instance-Prototype Affinity Distillation (IPAD), which regularizes instance-prototype relationship patterns instead of exact features; and Decision Boundary Perception (DBP), which adds high-entropy boundary instances as negative pairs. In the reported experiments, IPAL has the highest average accuracy on all four benchmarks, such as 83.07 percent on CS-CL versus 81.19 percent for the strongest rehearsal-based comparison.","pith_inferences":[],"forward_implications":["If PCL drifts less than Prototype Replay, non-exemplar continual graph learners can reduce their reliance on post-hoc drift-compensation machinery.","IPAL's benchmark results suggest that relational distillation plus contrastive prototypes can substitute for feature distillation, which the paper argues over-constrains plasticity.","Topology-weighted prototypes imply that node importance affects prototype quality, so uniform averaging of node features is suboptimal for graph continual learning.","Boundary-aware negative pairs imply that explicitly mining hard instances improves inter-class separation during incremental graph learning.","A better plasticity-stability trade-off means graph classifiers can be trained on sequentially arriving classes while keeping no raw data from past tasks.","This is an editorial inference: if the drift-reduction claim holds, the same contrastive-prototype mechanism may transfer to other exemplar-free continual learning settings, such as images or text, where stored prototypes are also affected by feature drift.","This is an editorial inference: a quantitative drift metric, measuring the mean shift of old-class features before and after each task, would make the paper's central premise directly testable; the paper currently supports it with a qualitative visualization.","This is an editorial inference: the PageRank weighting suggests a testable family of variants that replace PageRank with other centrality measures and compare plasticity, or that recompute prototype weights as the graph evolves."],"supporting_citations":[{"why":"Supplies the Prototype Contrastive Learning objective that IPAL builds on and whose lower drift the paper claims.","marker":"[14]"},{"why":"Provides the non-exemplar baseline and the argument that feature distillation over-constrains plasticity, motivating IPAD.","marker":"[19]"},{"why":"Supplies the PageRank algorithm used to weight nodes in TIGP.","marker":"[23]"},{"why":"Provides the Lipschitz-continuity assumption used in the feature-drift proof.","marker":"[26]"},{"why":"Supplies prototypical networks, the root of the PCL formulation.","marker":"[30]"},{"why":"Provides the first-order Taylor expansion used to relate parameter updates to feature drift in the appendix.","marker":"[38]"},{"why":"Provides the benchmark task construction and evaluation protocol used for the four datasets.","marker":"[42]"}],"fun_headline_variants":["Contrastive prototypes curb drift in continual graph learning","No raw data: prototype affinity learning tops graph benchmarks","IPAL: remembering graphs without storing examples","Prototype affinity beats replay without exemplars","Graph continual learning: contrastive prototypes win without memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the contrastive objective stabilizes old-class feature means, an effect the paper asserts in its proof rather than derives from the loss.","fun_headline_variants_meta":{"raw":{"variants":["Contrastive prototypes curb drift in continual graph learning","No raw data: prototype affinity learning tops graph benchmarks","IPAL: remembering graphs without storing examples","Prototype affinity beats replay without exemplars","Graph continual learning: contrastive prototypes win without memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1490,"prompt_tokens":979,"completion_tokens":511,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":439}},"tokens_in":595,"tokens_out":511,"duration_ms":4979,"temperature":1.0,"reasoning_tokens":439,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:17:33.071650+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the mean feature of each old class before and after training on a new task under PCL and under standard prototype replay; if PCL's mean shift is not consistently smaller across task transitions, the theorem on feature drift is not supported by the data.","supporting_citations":[{"cited_title":"Inductive graph few-shot class incremental learning","cited_arxiv_id":null,"evidence_quote":"Supplies the Prototype Contrastive Learning objective that IPAL builds on and whose lower drift the paper claims."},{"cited_title":"Graph neural networks: Architectures, stability, and transferability.Proceedings of the IEEE, 109(5):660–682, 2021","cited_arxiv_id":null,"evidence_quote":"Provides the Lipschitz-continuity assumption used in the feature-drift proof."}],"review_version":1}