{"id":"cf3afa8d-c13f-48b7-a391-6082adb309ae","arxiv_id":"2507.23534","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Experience Blending with Support Boundary Data improves continual learning accuracy by 6-14% over strong replay baselines on four image benchmarks.","lead":"A continual learning method adds synthetic, boundary-like data points to replay memory and blends two models, improving accuracy on four image benchmarks. The result suggests a cheap way to make replay-based continual learning more robust against catastrophic forgetting.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Frozen-encoder SBD generation degrades over longer task horizons; Table III shows EB(IN) falls below CLIB at 25 tasks, so the claimed consistency is not supported for the default protocol.","rationale":"The reader's weakest assumption is exactly the one I find most load-bearing: the first-task-frozen encoder must keep producing useful boundary-adjacent SBD for every later task. The paper's Table III provides direct evidence that this assumption degrades: with the default ImageNet encoder, EB(IN) loses 20.6% accuracy between 5 and 25 tasks and falls below CLIB at 25 tasks. That is an internal contradiction with the caption's claim that EB outperforms compared methods, and it undercuts the headline claim of consistent improvements. The memory-budget issue on ImageNet-1k and the abstract/body discrepancy over the ImageNet gain are real but secondary; they affect the magnitude of one result, whereas the frozen-encoder issue affects the mechanism across all longer-horizon settings. The concern is not that the method is fraudulent or that the 5-task results are fabricated; the concern is that the strength of the conclusion depends on a fixed-encoder assumption that the authors themselves show is fragile. Because the paper already reports the 25-task degradation, the conditional verdict is appropriate: the central claim should be conditioned on task horizon and encoder quality, and the authors should provide the margin analysis or an updated-encoder variant to show whether SBD remains boundary-adjacent in later tasks.","tokens_in":11327,"tokens_out":6100,"duration_ms":59694,"concrete_test":"On the 25-task CIFAR-100 i-Blurry-50-10 setup, log the margin (as defined in Section III-D) of the SBD samples generated by the frozen PR+SA and of held-out raw samples at the end of each task, using the final classifier. If the median absolute margin of SBD for tasks beyond 5 increases substantially above the value observed for tasks 1-5 (i.e., SBD are no longer near the boundary), the central mechanism fails for later tasks. A complementary rerun with PR/SA refreshed per task, or with the EB(CLIP) variant already in Table III, would show whether the 20.6% drop is explained by encoder staleness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SBD+EB consistently improves replay-based CL. The load-bearing assumption is that the first-task-frozen encoder PR+SA continues to generate boundary-adjacent SBD for all later tasks. Section III-C states: \"For all subsequent tasks, both PR and the SA layer are kept fixed to ensure stable and consistent feature encoding.\" The paper's own Table III shows this assumption is not safe: EB(IN) drops from 59.06 at 5 tasks to 50.27 at 10 tasks and 38.46 at 25 tasks, a 20.6% decline, and at 25 tasks EB(IN) is below CLIB (43.64), directly contradicting the caption \"Our method outperforms the compared methods.\" The paper attributes the decline to \"limited and biased data coverage in the first task,\" but that is precisely the same assumption: the no-update encoder never sees later-task classes, so its SBD may no longer lie near the current model's decision boundaries. Because SBD only regularizes if it is boundary-adjacent, the method's benefit is contingent on encoder stability; the reported 5-task gains cannot be extrapolated to longer task horizons. This is not an outside-consensus disagreement; it is an internal weakening visible in the paper's own numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Experience Blending (EB), a replay-based continual learning method that augments stored exemplars with Support Boundary Data (SBD): latent representations generated by injecting Laplace noise, motivated by the differential-privacy mechanism, into features produced by a first-task-fine-tuned encoder and self-attention layer. Training uses two models, one trained on replay plus SBD and one on SBD alone, aggregated by FedAvg-style weight averaging (DMA). The authors report large average-accuracy gains over prior replay methods on CIFAR-10, CIFAR-100, Tiny ImageNet, and ImageNet-1k under the i-Blurry-50-10 setting, and provide ablations, margin analyses, and memory-budget experiments supporting the contribution of each component.","tokens_in":11665,"tokens_out":5308,"duration_ms":56994,"significance":"If the reported gains hold under controlled comparisons, EB would be a simple, training-only addition to replay-based continual learning that improves decision-boundary stability. The paper's strengths are the clear component-level ablation in Table V, the margin evidence in Table I, the memory-budget analysis in Figure 6, and the breadth of datasets, including a large-scale ImageNet-1k result. The framing is not circular: the PCA and margin analyses are post hoc interpretations, not quantities that reduce to the fitted SBD by construction. However, two load-bearing concerns must be resolved: the default frozen-encoder version degrades sharply on longer task horizons and even falls below CLIB at 25 tasks, contradicting the paper's consistency claim; and the ImageNet-1k comparison does not control the replay-memory budget between EB and its baselines. The theoretical convergence claim for DMA is also overstated. These issues are fixable with additional experiments and corrected claims, but they currently affect the central message of 'consistent improvements.'","major_comments":[{"comment":"","section":"Section III-C and Table III"},{"comment":"","section":"Section IV and Table II"},{"comment":"","section":"Section III-B"}],"minor_comments":[{"comment":"","section":"Abstract"},{"comment":"","section":"Section III-A"},{"comment":"","section":"Table I"},{"comment":"","section":"Section III-A"},{"comment":"","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is promising, but the two main issues are internal: Table III contradicts the consistency claim for the default encoder, and the ImageNet-1k comparison is not memory-fair. These are addressable with additional controls and corrected claims. I also recommend asking for code or detailed reproducibility instructions, as several baselines (e.g., DPCL) are unavailable and some ImageNet-1k entries are blank. No circularity concern arose from reading the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a serious replay-based CL paper with a genuinely new mechanism and unusually large reported gains. The core idea—DP-style Laplace noise in latent space to create boundary-adjacent support data, blended via FedAvg-style aggregation of two internal models—is not in the cited ER or regularization literature. The ablations in Table V support each component, and the margin analysis in Section III-D is a nice diagnostic. On CIFAR-100 and Tiny ImageNet the gains are large and consistent, and the writing is clear.\n\nNow the soft spots, in order of severity. First, the long-horizon degradation. Table III shows EB(IN) dropping 20.6% from 5 to 25 tasks, and at 25 tasks it is 38.46 vs CLIB's 43.64—yet the caption says \"our method outperforms the compared methods,\" which is false for that row. The explanation about limited first-task coverage is plausible, but it means the benefit hinges on an encoder that doesn't drift. The CLIP variant does much better (57.44 at 25 tasks), so the problem is specifically with the ImageNet-initialized encoder, which is the default for most experiments. The blanket consistency claim overstates what the data show.\n\nSecond, memory fairness on ImageNet-1k: baselines get 20k replay images, EB gets 10k replay plus 10k SBD. That is not a controlled comparison unless memory is measured in bytes and SBD latents are shown to be cheaper. The paper asserts fairness but doesn't show the byte math. Third, the convergence claim: citing FedAvg's non-IID convergence for a two-model internal aggregation overstates what the theorem covers. Fourth, minor but real: the abstract says 2% for ImageNet-1k while the full text says 9.15%, and the code is not released.\n\nBottom line: the core idea is sound and the ablation supports it, but the paper needs revision before the headline claims are credible. A serious referee should focus on memory accounting and long-horizon behavior. I'd send it to review rather than desk-reject.","headline":"Original SBD+DMA idea with strong short-horizon gains, but the ImageNet memory accounting and the 25-task results undercut the headline consistency claim.","tokens_in":12141,"tokens_out":2632,"would_cite":false,"duration_ms":27301,"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":"Latent-space noise injected into replay exemplars generates support boundary data, and blending it into training via dual-model aggregation cuts catastrophic forgetting and improves average accuracy by 6-14 points on four…","keywords":["continual learning","catastrophic forgetting","experience replay","support boundary data","decision boundary regularization","differential privacy","dual-model aggregation","blurred boundary continual learning"],"falsifier":"Train EB on a stream where the first task covers only a small, unrepresentative subset of classes, then check empirically whether the SBD generated for later tasks still falls near the decision boundary (by measuring margin distributions). If the margin gains vanish or accuracy reverts to the ER baseline, the frozen-encoder assumption fails. A sharper test: replace the Laplace noise with Gaussian or uniform noise of the same variance; if the gains disappear, the specific DP-inspired heavy-tailed mechanism, rather than mere perturbation, is what does the work.","tokens_in":11126,"feed_emoji":"🎯","tokens_out":7518,"duration_ms":68099,"temperature":0.7,"pith_summary":"The paper claims that replay-based continual learning can be substantially improved by adding Support Boundary Data (SBD): latent feature vectors perturbed with differential-privacy-style Laplace noise so that they land near class decision boundaries. Because these synthetic samples are stored alongside ordinary exemplars and used only during training, the model sees boundary-adjacent points that implicitly regularize the classifier and reduce catastrophic forgetting. The authors back this with a framework, Experience Blending (EB), that trains two models—one on replay plus SBD and one on SBD alone—and averages their weights, and report average-accuracy gains of 10.45% on CIFAR-10, 6.26% on CIFAR-100, 13.75% on Tiny ImageNet, and 9.15% on ImageNet-1k over existing continual-learning baselines. If these results hold, boundary-sample generation is a simple, inference-free add-on to experience replay that works across class-incremental, domain-incremental, and blurred-boundary settings.","feed_headline":"Boundary-noise replay lifts continual learning by up to 14 points","feed_subtitle":"Injected latent noise creates boundary samples that stabilize replay memory with zero inference cost.","key_machinery":"The load-bearing object is Support Boundary Data (SBD). For each task, an image encoder followed by a self-attention layer produces features, and the batch is perturbed as $E_b = f'_b + \\mathrm{Lap}\\left(0, \\frac{\\max(f'_b)-\\min(f'_b)}{\\epsilon |S^t_b|}\\right)$, with $\\epsilon=0.005$ by default. The noise is differential-privacy-inspired but used as regularization: it deliberately lowers the model's confidence on these points by placing them in ambiguous regions. The second piece is dual-model aggregation (DMA): two identical models are trained, one on replay memory plus SBD and one on SBD only, and their weights are averaged with $\\alpha=0.5$ (FedAvg-style), which the authors find is necessary because training on SBD alone causes the model to memorize the synthetic samples rather than generalize. The self-attention layer emphasizes class-specific features before noise is added, and first-task fine-tuning aligns the encoder to the downstream domain; after the first task the encoder is frozen.","core_discovery":"The central discovery is that injecting multivariate Laplace noise into the latent representations of stored data—a mechanism borrowed from differential privacy but used for regularization, not privacy—produces synthetic samples that populate the ambiguous region around class decision boundaries. Augmenting experience replay with these Support Boundary Data (SBD) samples, and training with a dual-model aggregation (DMA) step that averages the weights of a replay+SBD model and an SBD-only model, consistently raises average task accuracy. On the i-Blurry-50-10 setting the gains over the best prior baselines are 10.45 points on CIFAR-10, 6.26 points on CIFAR-100, 13.75 points on Tiny ImageNet, and 9.15 points on ImageNet-1k (with a CLIP encoder), and the method also raises average margins and reduces margin variance, which the authors interpret as more stable decision boundaries.","pith_inferences":["The paper does not compare SBD against other boundary-regularization tricks such as mixup in latent space; a natural extension is to test whether the gains come from the Laplace distribution's heavy tails or simply from adding any controlled noise near the boundary.","Because after the first task the encoder is frozen, the method's success depends on the first task being representative; an adaptive version that periodically refreshes the encoder (or the SBD) with new-task data could reduce the 20.6% drop seen at 25 tasks with the ImageNet encoder.","The margin statistics used for the analysis (average margin up, variance down) could serve as a cheap online diagnostic to decide when to regenerate SBD, which the authors leave as an open direction.","The DP-noise mechanism suggests a possible link to certified robustness or to privacy-utility trade-offs, but the paper deliberately separates SBD from privacy, so any privacy guarantee would require a separate analysis."],"forward_implications":["Experience replay methods can be upgraded by adding a latent-noise SBD memory with no change to the inference-time model.","The frozen-encoder design means SBD generation is cheap during training (about +0.13s per epoch for 100 SBD samples) and free at test time.","Under a fixed memory budget, allocating part of the budget to SBD rather than raw images yields roughly 4% higher accuracy than CLIB and FOSTER at the same memory footprint (Figure 6).","The method transfers across encoder choices (ImageNet-pretrained ResNet-18 and CLIP ViT-B/32) and across task counts, with CLIP initialization nearly eliminating the 5-to-25-task degradation.","Blurred-boundary and domain-incremental settings, which are harder for typical class-incremental methods, show the largest relative gains, suggesting SBD helps most when task boundaries are noisy."],"supporting_citations":[{"why":"Supplies the i-Blurry-50-10 continual-learning setup and the sample-wise importance selection used to build the replay memory.","marker":"[12]"},{"why":"Contributes the differential-privacy Laplace noise-adding mechanism that the paper repurposes to generate support boundary data.","marker":"[18]"},{"why":"Provides the self-attention layer design (from Self-Attention GAN) used to emphasize class-specific features before noise injection.","marker":"[19]"},{"why":"FedAvg is the aggregation algorithm used for dual-model aggregation across the two internal models.","marker":"[20]"},{"why":"Establishes convergence of FedAvg on non-i.i.d. data, which the paper cites to justify DMA's convergence guarantee.","marker":"[21]"},{"why":"First Session Adaptation motivates the first-task fine-tuning (FTF) strategy that aligns the pretrained encoder to the downstream domain.","marker":"[23]"},{"why":"CLIP's vision encoder is used as an alternative initialization for PR, notably for the ImageNet-1k experiments.","marker":"[24]"},{"why":"FOSTER is the strongest prior baseline in the ablations and overall tables, and the paper's gains over it are central to the claim.","marker":"[14]"}],"fun_headline_variants":["Boundary noise boosts replay by up to 14 points","Latent noise creates support samples for stable continual learning","Noise-injected replay samples sharpen decision boundaries","Dual-model blending with boundary support data lifts CL accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"After the first task, the image encoder and self-attention layer are frozen, and the paper assumes this single encoder keeps producing SBD that lies near decision boundaries for every later task, even as the data distribution shifts; the observed accuracy drop of 20.6% when moving from 5 to 25 tasks shows how much rests on that encoder.","fun_headline_variants_meta":{"raw":{"variants":["Boundary noise boosts replay by up to 14 points","Latent noise creates support samples for stable continual learning","Noise-injected replay samples sharpen decision boundaries","Dual-model blending with boundary support data lifts CL accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000912,"raw_usage":{"total_tokens":3901,"prompt_tokens":914,"completion_tokens":2987,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":2923}},"tokens_in":530,"tokens_out":2987,"duration_ms":18769,"temperature":1.0,"reasoning_tokens":2923,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:37:55.608739+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train EB on a stream where the first task covers only a small, unrepresentative subset of classes, then check empirically whether the SBD generated for later tasks still falls near the decision boundary (by measuring margin distributions). If the margin gains vanish or accuracy reverts to the ER baseline, the frozen-encoder assumption fails. A sharper test: replace the Laplace noise with Gaussian or uniform noise of the same variance; if the gains disappear, the specific DP-inspired heavy-tailed mechanism, rather than mere perturbation, is what does the work.","supporting_citations":[{"cited_title":"Online continual learning on class incremental blurry task configuration with anytime inference,","cited_arxiv_id":null,"evidence_quote":"Supplies the i-Blurry-50-10 continual-learning setup and the sample-wise importance selection used to build the replay memory."},{"cited_title":"Differential privacy,","cited_arxiv_id":null,"evidence_quote":"Contributes the differential-privacy Laplace noise-adding mechanism that the paper repurposes to generate support boundary data."},{"cited_title":"Self- attention generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Provides the self-attention layer design (from Self-Attention GAN) used to emphasize class-specific features before noise injection."},{"cited_title":"Communication-Efficient Learning of Deep Networks from Decentralized Data,","cited_arxiv_id":null,"evidence_quote":"FedAvg is the aggregation algorithm used for dual-model aggregation across the two internal models."},{"cited_title":"On the convergence of fedavg on non-iid data,","cited_arxiv_id":null,"evidence_quote":"Establishes convergence of FedAvg on non-i.i.d. data, which the paper cites to justify DMA's convergence guarantee."},{"cited_title":"First session adaptation: A strong replay-free baseline for class-incremental learning,","cited_arxiv_id":null,"evidence_quote":"First Session Adaptation motivates the first-task fine-tuning (FTF) strategy that aligns the pretrained encoder to the downstream domain."},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"CLIP's vision encoder is used as an alternative initialization for PR, notably for the ImageNet-1k experiments."},{"cited_title":"FOSTER: Feature boosting and compression for class-incremental learn- ing,","cited_arxiv_id":null,"evidence_quote":"FOSTER is the strongest prior baseline in the ablations and overall tables, and the paper's gains over it are central to the claim."}],"review_version":1}