{"id":"34d03f6b-8613-44be-9045-cbc250b37bbb","arxiv_id":"2608.11690","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A layer-wise information-theoretic decomposition bounds replay-based continual learning's generalization gap, predicting memory scaling, an interior stabilization layer, and gradient-alignment signals that track forgetting.","lead":"This paper derives a layer-wise information-theoretic bound that separates the generalization gap of replay-based continual learning into finite-memory drift and optimization dependence. The theory yields diagnostics for which layer to stabilize and when old and new task gradients align, with experiments supporting the predicted memory scaling and alignment-forgetting link.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SGLD diagnostics assume a frozen prefix but benchmarks train jointly, leaving the alignment diagnostic's theoretical cover conditional.","rationale":"The reader's weakest_assumption identifies the frozen-prefix mismatch in Section V-A, and I agree that this is the single most load-bearing concern. The main theorem, Theorem IV.1, is a general decomposition that does not require the prefix to be frozen, and its proof appears internally consistent under the stated subgaussian and absolute-continuity assumptions (with the usual caveat that KL drift may be infinite under support mismatch, which the paper explicitly routes around via the Wasserstein relaxation). What the frozen-prefix assumption actually threatens is the SGLD branch: Corollary V.2 and Theorem V.3 are proved for an algorithm that updates only the suffix Θ while W^{1:l} is held fixed, and the derivation of the sensitivity-aware alignment cos_H depends on that conditional dynamics. The benchmark experiments, however, train all parameters jointly with SGD, so the theoretical justification for applying cos_H as a forgetting diagnostic in those experiments is conditional on a regime the experiments do not implement. This does not invalidate the empirical finding—the correlations could still hold for reasons the theory does not capture—but it does weaken the paper's claim that the diagnostic is derived from the bound in the setting where it is deployed. The Corollary IV.2 no-ties issue is real but secondary: it affects a simplification at the input layer and the comparison to prior bounds, not the main theorem or the empirical diagnostics. I therefore recommend keeping the reader's CONDITIONAL verdict: the central decomposition survives, but the operational SGLD branch should either be explicitly framed as a heuristic extrapolation to joint training or be supported by a derivation that does not require a frozen prefix.","tokens_in":41945,"tokens_out":31820,"duration_ms":337643,"concrete_test":"Re-run the Split-CIFAR-100 / ER benchmark cell with the prefix W^{1:ℓ*} frozen at the empirically identified funnel layer ℓ* during each task's training, so that only the suffix is updated (optionally with SGLD noise added as in Theorem V.3). Compute the task-controlled partial correlation of cos_H with pairwise forgetting ΔF_{i,t} under the same statistical protocol as Table I. Compare with the reported jointly-trained value of −0.948: if the frozen-prefix run does not reproduce a strong negative correlation, the joint-regime result is an unsupported extrapolation of the SGLD theory; if it does, the frozen-prefix assumption is not the source of the empirical signal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-A fixes the prefix W^{1:l} and analyzes only the suffix parameters Θ under SGLD. Theorem V.3 and Corollary V.2 are proved under this conditioning: the update rule Θ_r = Θ_{r-1} + η_r G_r + N_r defines G_r using a fixed W^{1:l}, and Corollary V.2's Markov chain U_T -> M_{1:T-1} -> Θ_{T-1}^R requires the old-task component of U_T to be a deterministic function of the buffer and the fixed prefix. The benchmark experiments in Section VI train ER, DER++, and iCaRL with all parameters updated jointly by SGD, not SGLD, and no layer is frozen. Consequently, the trajectory log-determinant budget, the sensitivity metric H = (I + αV)^{-1}, and the alignment cosine cos_H—presented as bound-derived diagnostics—are derived for a training dynamics that the experiments do not execute. This is the load-bearing gap: the paper's strongest operational claim that the alignment diagnostic tracks forgetting is empirically demonstrated in the jointly-trained regime, but theoretically justified only in the frozen-prefix regime. The main Theorem IV.1 itself does not depend on the frozen-prefix choice, so this concern is about the operational branch rather than the core decomposition, but it directly affects the claimed status of the SGLD diagnostics as bound-driven rather than heuristic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a layer-wise information-theoretic framework for replay-based continual learning. Theorem IV.1 decomposes the expected generalization gap into a replay-centroid drift term K^(l) and an optimization-dependence term that is further resolved into stability S^(l), plasticity P^(l), interaction R^(l), and residual coupling C^(l), with an effective sample size N_eff governing the variance scale. A Wasserstein relaxation (Theorem IV.4) replaces the KL-based drift by a drift--sensitivity product and predicts an interior \"generalization funnel\" layer. An SGLD instantiation (Theorem V.3) upper-bounds the optimization branch by a cumulative trajectory-level log-determinant budget, which is then expanded into instability and interaction costs and leads to a curvature-aware gradient-alignment statistic cos_H. Controlled Gaussian and MNIST experiments plus Split-CIFAR-100 and Split-TinyImageNet benchmarks test three predictions: the m^{-1/2} variance-floor scaling, the interior funnel location and its depth trend, and the negative correlation between the alignment diagnostic and task-wise forgetting.","tokens_in":42143,"tokens_out":13984,"duration_ms":150096,"significance":"If the main theorem holds, the paper is a genuine step beyond hypothesis-level information-theoretic bounds: it isolates a finite-memory representation drift from reuse-induced optimization dependence at every layer, and the two refinements yield falsifiable structural predictions. The manuscript is unusually honest: it reports the fitted m-exponent as -0.443 with CI [-0.498, -0.391] rather than claiming the idealized -1/2, it explicitly presents the funnel proxy as weak for DER++ and iCaRL, it includes iCaRL as a deliberate scope boundary, and it acknowledges that the certified Lipschitz product is too loose to localize a funnel in deep networks. The Appendix B proofs are detailed, and the core decomposition in Theorem IV.1 does not depend on the SGLD or frozen-prefix assumptions. The main weakness is that the SGLD-derived diagnostics are rigorously justified only under a fixed prefix and injected noise, while the experiments that carry the operational claims use joint training and, on the benchmarks, SGD without injected noise.","major_comments":[{"comment":"The SGLD analysis conditions on a fixed prefix W^{1:l} and updates only the suffix parameters; Corollary V.2's Markov chain U_T -> M_{1:T-1} -> Theta_{T-1}^R requires the old-task component of U_T to be a deterministic function of the buffer and the frozen prefix. The experiments in Section VI-A, however, train all parameters jointly: the benchmark pipelines update every layer with SGD, and even the controlled MNIST SGLD stream does not freeze a prefix. Consequently, the log-determinant budget of Theorem V.3 and the alignment statistic cos_H of Corollary V.7, which are presented as bound-derived diagnostics, are not theoretically justified for the dynamics actually executed. This is load-bearing for the paper's operational claim that the alignment diagnostic is a reading of the bound rather than a heuristic. I would ask the authors either to extend the analysis to joint prefix updates, or to explicitly reclassify the SGLD diagnostics as heuristics motivated by the frozen-prefix analysis and to validate them in a controlled experiment with a genuinely frozen prefix.","section":"Section V-A / Theorem V.3 / Corollary V.2"},{"comment":"The benchmark experiments use SGD with learning rate 0.03 and no injected Gaussian noise, so the SGLD signal-to-noise ratio alpha = eta^2/tau^2 entering the sensitivity metric H = (I + alpha V)^{-1} is undefined unless an ad hoc regularization is introduced. The paper does not state how alpha is set when computing cos_H for the SGD-trained ER, DER++, and iCaRL models in Table I, nor does it specify the diagonal approximation details beyond saying it is used. Since the central empirical claim -- that the alignment diagnostic tracks forgetting -- is made on these SGD benchmarks, the absence of a well-defined alpha and the lack of a theoretical bridge from SGLD to SGD leave the benchmark validation of the optimization branch incomplete. The paper should either provide the SGD analogues of Theorem V.3 or clearly report the heuristic choice of alpha and argue its insensitivity.","section":"Section VI-A / Table I / Figure 6"}],"minor_comments":[{"comment":"The citation [47] is given for stochastic gradient Langevin dynamics, but the listed reference is Chen et al., \"On the stability-plasticity dilemma in continual meta-learning,\" which is not the standard SGLD source; please cite Welling and Teh (2011) or an equivalent canonical SGLD reference.","section":"Section V / References"},{"comment":"The caption contains the fragment \"MNIST-100×y-axis,\" which appears to be a formatting artifact or missing explanation; please clarify what the y-axis scaling refers to.","section":"Figure 1 caption"},{"comment":"The proof states that the training sequence is block-wise i.i.d. and therefore C^(0)=0; this is correct, but the argument would be clearer if it explicitly noted that the random subset index set is independent of the i.i.d. task data, so without-replacement sampling does not create dependence at the input layer.","section":"Appendix B, proof of Corollary IV.2"},{"comment":"The funnel study uses the on-distribution Jacobian norm as a lower bound for the suffix Lipschitz constant, while Theorem IV.4 and Corollary IV.5 concern the certified upper bound product; the paper acknowledges this in Section VII, but a sentence in Section VI-C stating that the empirical funnel is a surrogate-location result, not a certified one, would improve precision.","section":"Section VI-C / Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The core decomposition in Theorem IV.1 is careful and the experimental reporting is unusually honest; the main obstacle is the mismatch between the frozen-prefix SGLD theory and the jointly trained SGD/SGLD experiments for the alignment diagnostics. This is fixable either by adding a frozen-prefix controlled experiment or by weakening the claim from 'bound-derived' to 'bound-motivated heuristic.' The reader's concern about Corollary IV.2 does not survive scrutiny in my reading, so I would not require changes there beyond a clarifying sentence. The paper fits the journal's scope and, after the operational branch is reconciled, would be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main decomposition is real and deserves a serious referee. Theorem IV.1 separates the replay generalization gap into a centroid drift term and an optimization-dependence term, then splits the latter into stability, plasticity, interaction, and residual coupling. That is a genuine advance over the hypothesis-level bounds cited, including the authors' own earlier work. The layer-wise treatment is not cosmetic: it exposes a depth trade-off that a black-box bound cannot, and the Wasserstein relaxation gives a principled criterion for which layer to stabilize. The appendix proofs for the main theorem and the SGLD instantiation are detailed enough to check line by line, and the controlled experiments are honestly analyzed — the m^{-1/2} scaling fit, the funnel location in closed-form populations, and the task-controlled partial correlations are all sensible. Credit also where due: the paper states its own scope boundary with iCaRL and reports the funnel's weak benchmark transfer explicitly, rather than burying it.\n\nTwo soft spots, one minor and one load-bearing. First, Corollary IV.2 asserts that C^(0)=0 because the input-layer training sequence is \"block-wise i.i.d.\" That is false under the paper's without-replacement buffer sampling: the stored examples within a task are exchangeable but not independent, so the KL to the product reference is not zero. The corollary can be repaired by assuming with-replacement sampling or by bounding the residual, but as written the proof overclaims. This affects only the corollary, not Theorem IV.1.\n\nSecond, and more important: the SGLD branch in Section V fixes the prefix W^{1:l} and analyzes only the suffix, and Corollary V.2's Markov structure relies on that conditioning. The benchmark experiments train all parameters jointly with SGD, not SGLD, so the log-det budget, the H-metric, and the alignment cosine are theoretically justified in a regime the experiments do not execute. The main theorem does not depend on the frozen-prefix choice, but the paper's strongest operational claim — that alignment cos_H is a bound-derived forgetting diagnostic — is only rigorously connected to the frozen-prefix SGLD setting. As written, the benchmark alignment results are heuristic. This is fixable by either extending the analysis to joint training or explicitly reframing the benchmark diagnostics as motivated by, rather than derived from, the SGLD bound.\n\nWho this is for: researchers working on theory of continual learning or on principled replay diagnostics will get real value. The paper should go to peer review — it is substantive, mostly careful, and the flaws are repairable rather than fatal. I would recommend the editor send it out, with referees asked to verify the corollary fix and to press the authors on the training-dynamics mismatch.","headline":"A genuinely new layer-wise decomposition of replay generalization with real controlled validation, but the SGLD diagnostic branch is derived for a frozen-prefix regime the benchmarks do not run, and Corollary IV.2 overclaims independence at the input layer.","tokens_in":42663,"tokens_out":1902,"would_cite":true,"duration_ms":25934,"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":"Replay-based continual learning splits its generalization gap at every layer into a finite-memory drift term and an optimization-dependence term, and derives concrete diagnostics from the split.","keywords":["continual learning","experience replay","catastrophic forgetting","information-theoretic generalization bounds","Wasserstein distance","stochastic gradient Langevin dynamics","layer-wise analysis","gradient alignment"],"falsifier":"In the paper's controlled Gaussian–linear setting, train two agents with identical current-task data but independently resampled equal-size buffers while keeping the feature prefix fixed; if the difference in their generalization gaps is not explained by the change in the estimated drift term $K^{(l)}$ (with the estimated dependence term $S^{(l)}+P^{(l)}-R^{(l)}+C^{(l)}$ held fixed), the layer-wise separation claimed in Theorem IV.1 would be falsified.","tokens_in":41724,"feed_emoji":"🧠","tokens_out":13787,"duration_ms":135227,"temperature":0.7,"pith_summary":"Replay-based continual learning stores a small buffer of past examples and mixes them into current training; the paper argues that the resulting generalization gap is governed by two coupled but separable effects that existing theory folds together. Finite memory replaces each past task distribution with an empirical proxy, producing a representation drift that is reshaped by every layer of the network, while repeated reuse couples the buffer, the current data, and the final hypothesis through a shared optimization trajectory. The paper's main theorem (Theorem IV.1) bounds the expected gap by a drift term plus an optimization-dependence term, and shows the latter decomposes into stability, plasticity, interaction, and residual coupling at every layer. If this decomposition is right, it would let practitioners identify which interior layer to stabilize and monitor forgetting online through gradient alignment, and the experiments report the predicted memory scaling, an interior funnel layer, and an alignment signal that tracks forgetting on standard replay benchmarks.","feed_headline":"Replay generalization splits into drift and dependence per layer","feed_subtitle":"A layer-wise bound isolates finite-memory bias from reuse coupling, and picks the layer to stabilize.","key_machinery":"The load-bearing machinery is a layer split of the network at depth $l\\in\\{0,\\dots,L\\}$. The paper defines three layer-$l$ laws: the population law $P_{A_l,Y|i,W^{1:l}}$, the empirical proxy $\\hat P_{A_l,Y|S_i,W^{1:l}}$ on the stored buffer, and the replay centroid $Q_{A_l,Y|i,W^{1:l}} = \\mathbb{E}_{S_i}[\\hat P_{A_l,Y|S_i,W^{1:l}}\\mid W^{1:l}]$, which averages over buffer construction while conditioning on the learned features. Comparing the centroid to the population isolates finite-memory drift; comparing the actual training sequence to a decoupled reference $\\tilde Q$ (independent coordinates with the same marginals) isolates reuse coupling. The optimization dependence is then rewritten through the interaction-information identity as $S^{(l)}+P^{(l)}-R^{(l)} = I(U_{\\mathrm{old}};U_{\\mathrm{new}};W^{l+1:L}\\mid W^{1:l})$, and the geometric relaxation replaces KL by Wasserstein-1 distance with a suffix Lipschitz factor $\\bar\\rho_l(W)=\\rho_0(1\\vee\\prod_{h>l}\\rho_h\\|W_h\\|_{\\mathrm{op}})$, while the SGLD refinement bounds the incremental dependence by $\\frac{1}{2}\\log\\det(I+(\\eta^2/\\tau^2)M)$ per step and reads it through the sensitivity metric $H=(I+\\alpha V)^{-1}$.","core_discovery":"The central discovery is Theorem IV.1: for a $\\sigma$-subgaussian loss and any split layer $l$, the expected generalization gap is bounded by $(T-1)\\sqrt{2\\sigma^2 K^{(l)}} + \\sqrt{2\\sigma^2/N_{\\mathrm{eff}}\\,(S^{(l)}+P^{(l)}-R^{(l)}+C^{(l)})}$, where $N_{\\mathrm{eff}} = 1/((T-1)/m + 1/n)$. Here $K^{(l)}$ is the average KL divergence between the layer-$l$ replay centroid and the population distribution of each old task, and $S^{(l)},P^{(l)},R^{(l)},C^{(l)}$ are layer-wise mutual-information terms measuring, respectively, how much the suffix parameters retain about replayed old representations, how much they encode the current task, the signed synergy/interference between old and new representations, and residual coupling of the training sequence. The drift term does not vanish as current-task size grows: it is the cost of compressing memory to $m$ exemplars per past task, and it survives even when the buffer contents are i.i.d. because the learned feature map is itself fitted to the stored samples. A Wasserstein relaxation of the drift gives a depth-dependent drift–sensitivity product whose minimizer the paper calls a generalization funnel layer, and an SGLD instantiation of the optimization term turns the bound into a trajectory-level log-determinant budget that separates gradient-covariance instability from a curvature-aware gradient-alignment interaction cost.","pith_inferences":["An implication the paper leaves implicit: the same layer-wise decomposition offers a testable criterion for buffer design—buffer selection policies should be judged by how much they reduce $K^{(l)}$ and $C^{(l)}$ at the funnel layer, not just by end-task accuracy.","Because the theorems condition on a frozen prefix $W^{1:l}$ while the benchmark experiments train all parameters jointly, the operational force of the log-determinant budget in standard training is an extrapolation; comparing the alignment diagnostic under frozen versus jointly trained prefixes would delimit when the bound's optimization branch applies.","The signed interaction term $R^{(l)}$ suggests a curriculum prediction not tested in the paper: if old and new tasks are ordered so that their sensitivity-metric gradient alignment is positive, the accumulated dependence budget should be lower, so task orderings with high $\\cos_H$ among adjacent tasks should forget less.","The funnel-layer and alignment quantities could in principle be combined into an online intervention—freeze or lightly regularize the layer with the smallest drift–sensitivity product and down-weight replay gradients that anti-align in the $H$ metric—but the paper only provides the diagnostics, not the combined controller."],"forward_implications":["The variance term obeys an effective-sample-size scaling: as the number of current examples $n\\to\\infty$, the bound does not shrink to zero but settles at a finite-memory floor of order $\\sqrt{(T-1)/m}$; controlled experiments confirm a fitted exponent near $-1/2$ for the gap versus buffer size $m$.","At the input layer the bound reduces to the standard mutual-information bound with effective sample size $N_{\\mathrm{eff}}$, while at the output layer only drift and residual coupling remain; the dominant source of error therefore shifts from parameter-information complexity at shallow layers to representation mismatch at deep layers.","Minimizing the drift–sensitivity product selects an interior generalization funnel layer; the paper argues that feature distillation or partial freezing should target that interior basin rather than either end of the network, and reports that the interior minimizer appears in every tested depth cell and moves deeper as networks deepen.","For gradient-replay methods (experience replay and DER++), the curvature-aware alignment $\\cos_H$ of old and new mean gradients negatively tracks pairwise forgetting, with benchmark partial correlations between $-0.65$ and $-0.95$, several times stronger than task-order baselines.","The same SGLD analysis yields a local replay/current mixing coefficient that minimizes the sensitivity-weighted mean-update energy; replay-heavy values are warnings of interference rather than a globally optimal replay policy."],"supporting_citations":[{"why":"Supplies the mutual-information generalization bound that is recovered as the input-layer limit.","marker":"[15]"},{"why":"Provides the conditional-mutual-information and on-average stability tools used at the buffer-equals-full-sample limit.","marker":"[16]"},{"why":"Gives the hierarchical layer-wise generalization bound that this paper adapts to replay.","marker":"[18]"},{"why":"Is the prior hypothesis-level information-theoretic bound for replay that this work refines layer by layer.","marker":"[36]"},{"why":"Supplies the trajectory-level SGLD generalization analysis that the optimization branch generalizes.","marker":"[37]"},{"why":"Represents the gradient-projection family against which the sensitivity-aware alignment diagnostic is contrasted.","marker":"[31]"},{"why":"Is the feature-distillation stabilization technique that the funnel-layer result directs to an interior layer.","marker":"[44]"},{"why":"Is the replay baseline used in the benchmark validation of the diagnostics.","marker":"[48]"}],"fun_headline_variants":["Drift vs dependence: how replay forgets, layer by layer","Layer-wise bounds decompose replay forgetting into drift and coupling","Interior funnel layer emerges from replay drift bounds","New bounds pick optimal layer to stabilize in continual learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The optimization-branch analysis fixes the bottom $l$ layers and analyzes only the suffix under SGLD, while the benchmark experiments update all parameters jointly; the claim that the log-determinant budget and alignment diagnostic remain valid under joint training is therefore an assumption, not a proven consequence.","fun_headline_variants_meta":{"raw":{"variants":["Drift vs dependence: how replay forgets, layer by layer","Layer-wise bounds decompose replay forgetting into drift and coupling","Interior funnel layer emerges from replay drift bounds","New bounds pick optimal layer to stabilize in continual learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00069,"raw_usage":{"total_tokens":3199,"prompt_tokens":1095,"completion_tokens":2104,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":711,"completion_tokens_details":{"reasoning_tokens":2039}},"tokens_in":711,"tokens_out":2104,"duration_ms":14090,"temperature":1.0,"reasoning_tokens":2039,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:31:55.397482+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the paper's controlled Gaussian–linear setting, train two agents with identical current-task data but independently resampled equal-size buffers while keeping the feature prefix fixed; if the difference in their generalization gaps is not explained by the change in the estimated drift term $K^{(l)}$ (with the estimated dependence term $S^{(l)}+P^{(l)}-R^{(l)}+C^{(l)}$ held fixed), the layer-wise separation claimed in Theorem IV.1 would be falsified.","supporting_citations":[{"cited_title":"Information-theoretic analysis of generalization capability of learning algorithms,","cited_arxiv_id":null,"evidence_quote":"Supplies the mutual-information generalization bound that is recovered as the input-layer limit."},{"cited_title":"Reasoning about generalization via conditional mutual information,","cited_arxiv_id":null,"evidence_quote":"Provides the conditional-mutual-information and on-average stability tools used at the buffer-equals-full-sample limit."},{"cited_title":"Hierarchical generalization bounds for deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Gives the hierarchical layer-wise generalization bound that this paper adapts to replay."},{"cited_title":"Information-theoretic generalization bounds of replay-based continual learning,","cited_arxiv_id":null,"evidence_quote":"Is the prior hypothesis-level information-theoretic bound for replay that this work refines layer by layer."},{"cited_title":"Information-theoretic generalization bounds for sgld via data-dependent estimates,","cited_arxiv_id":null,"evidence_quote":"Supplies the trajectory-level SGLD generalization analysis that the optimization branch generalizes."},{"cited_title":"Podnet: Pooled outputs distillation for small-tasks incremental learning,","cited_arxiv_id":null,"evidence_quote":"Is the feature-distillation stabilization technique that the funnel-layer result directs to an interior layer."},{"cited_title":"Dark experience for general continual learning: a strong, simple baseline,","cited_arxiv_id":null,"evidence_quote":"Is the replay baseline used in the benchmark validation of the diagnostics."}],"review_version":1}