{"id":"5c768742-3ba4-4f79-936e-66a56c3690e2","arxiv_id":"2411.13852","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Synthetic image contamination degrades online continual learning, and ESRM's entropy-based buffer selection plus contrastive feature alignment significantly mitigates the degradation.","lead":"Synthetic AI-generated images harm online continual learning models, and the paper's ESRM method reduces the damage by keeping high-entropy (real) samples in the buffer and aligning real and synthetic features. It matters because web-scraped training data is increasingly contaminated by synthetic images, and this study offers a practical defense for streaming-learning systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ESRM's core entropy discriminator rests on a single moderate AUC (0.71); if the synthetic-vs-real entropy gap shifts or reverses for other generators or prompt distributions, both ES and RM lose their grounding and the claimed robustness is unsupported.","rationale":"I read the paper as an empirical demonstration plus a method. The primary evidence is Table 2 and the ablations; the method's two components are both tied to the entropy observation. The reader's weakest assumption is correct and is the most load-bearing: a single AUC of 0.71 is weak quantitative support, and no cross-generator measurement is provided. However, the paper is explicit about its limitations and the reported improvements are consistent and large in the tested settings. The low clean accuracy of ESRM on CIFAR-10 (67.35 vs OnPro 74.87) is an additional caveat but does not undermine the 'alleviation' claim. An external generalization test would resolve the concern; until then the verdict stands, with the limitation clearly noted.","tokens_in":29484,"tokens_out":10537,"duration_ms":108509,"concrete_test":"Generate a CIFAR-100 synthetic twin with a non-evaluated diffusion model (e.g., DALL-E 3 or a fine-tuned SDXL) using LLM-generated diverse prompts, contaminate at P=70%, then: (i) train vanilla ER and measure the entropy ROC AUC on the training set; (ii) train ESRM and ER, comparing final accuracy. If the AUC drops below ~0.6 or ESRM no longer beats ER, the entropy gap assumption is violated and the central robustness claim fails to generalize.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ESRM markedly reduces contamination-driven degradation, with performance attributed to ES and RM. Both components in Sec. 5.1 and 5.2 depend on Obs. 3: synthetic images yield lower prediction entropy than real images under the online CL model. The only quantitative support is Fig. 7: an ROC AUC of 0.7098 for an ESRM-trained model on In-100/SDXL. This is a weak, single-point estimate: at a median threshold roughly a third of samples are misclassified, and it is measured after ESRM training, so it may partly reflect the method itself rather than a stable property. The paper's own limitations (Appendix A) acknowledge only five diffusion models and simple prompts; no evidence is given that the gap persists for newer generators, diverse prompts, or other contamination distributions. If entropy no longer separates real from synthetic, ES will not preferentially retain real samples (Fig. 4, Alg. 1), and RM's group split (Sec. 5.2) becomes arbitrary; the method could provide no benefit or even reinforce contamination-induced bias. The claim as stated therefore holds only conditional on an unquantified empirical regularity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the effect of AI-generated (synthetic) images on online continual learning (CL). It simulates contaminated training streams by replacing a fraction P of four benchmark datasets (CIFAR-10, CIFAR-100, TinyImageNet, ImageNet-100) with synthetic twins produced by Stable Diffusion XL, Stable Diffusion v1.4/v2.1, VQDM, and GLIDE, using simple class-name prompts. It reports that six existing online CL methods degrade as P increases, and it proposes ESRM, a replay-based method with two components: Entropy Selection (ES), which preferentially stores high-entropy (more likely real) samples in the buffer, and Real-synthetic similarity Maximization (RM), a contrastive loss that pulls features of high- and low-entropy groups together. Experiments over four datasets, two contamination settings (SDXL-only and a five-model mix), contamination ratios from 50% to 95%, DIL results, ablations, and buffer-size variants show that ESRM generally suffers smaller contamination-induced degradation and often achieves higher absolute accuracy than the baselines. Source code and detailed hyperparameter settings are provided.","tokens_in":29746,"tokens_out":4881,"duration_ms":52885,"significance":"If the results hold, the paper identifies a realistic and understudied threat to online CL—synthetic data on the Internet—and offers a simple, replay-compatible mitigation. The empirical protocol is extensive: four datasets, six baselines, multiple contamination ratios, five generators, ablations, and error bars over 5–10 runs. The paper also contributes four qualitative observations about synthetic data in online CL, and releases code, which supports reproducibility. The main limitation is that the mechanism of both ESRM components rests on a single moderate quantitative estimate of the entropy gap (ROC AUC 0.7098 on one dataset/model), and the paper's own limitation section acknowledges the narrow generator and prompt coverage. The central claim that contamination degrades online CL and that ESRM alleviates the degradation is therefore empirically supported on the tested settings, but the generality of the entropy-based mechanism is not yet established.","major_comments":[{"comment":"Both ES and RM are explicitly grounded in Observation 3, that synthetic images have lower predictive entropy than real images. The only quantitative support is the ROC curve in Fig. 7, which reports AUC = 0.7098 on In-100/SDXL (P = 50%) for an ESRM-trained model. This is a single, moderate discriminability estimate: at an operating point near the ROC curve, roughly one third of samples would be misclassified. Moreover, the AUC is measured after ESRM training, so it may partly reflect the method's own effect rather than a pre-existing property that ES and RM exploit. Since ES's filtering and RM's group splitting both become arbitrary if the entropy ordering weakens or reverses, the robustness claim is conditional on an empirical regularity that is not yet quantified across the test datasets, training stages, or unseen generators. I request additional evidence: entropy-gap statistics or AUC values for all datasets and baselines, before and after training, and for at least one held-out generator or more diverse prompt distribution; alternatively, the authors should explicitly restrict the claim to the tested generator set.","section":"4.3, 5.1, 5.2, Fig. 7"},{"comment":"On CIFAR-10, the headline claim of 'alleviating performance deterioration' is less clear in absolute terms. At P = 70%, ESRM achieves 67.47 ± 1.43 while OnPro achieves 68.79 ± 1.17, and at P = 80% the two are close (66.81 vs. 66.07). ESRM's smaller degradation is measured relative to its own clean accuracy of 67.35, which is substantially below OnPro's clean accuracy of 74.87. A method with lower clean accuracy can show smaller absolute drops even if its contaminated accuracy is not the best. The paper should discuss this trade-off explicitly and, ideally, report a paired analysis of contamination-induced drops or normalized degradation for all methods, not only the raw accuracy tables.","section":"Table 2 (C10/SDXL rows)"},{"comment":"The paper uses the word 'significant' repeatedly (e.g., in the abstract, Section 4.1, and Section 6.2) but reports only means and standard deviations. Some of the key comparisons that support the method—ES vs. random in Table 5 (45.92 ± 0.42 vs. 44.84 ± 0.80 at P = 70%) and the loss-term ablation in Table 6 (Baseline 42.61 ± 0.84 vs. ESRM 45.92 ± 0.42)—are close enough that a formal paired test or confidence interval would materially strengthen the claim. I recommend adding paired significance tests or bootstrap confidence intervals for the main comparisons, especially the ES/Random comparison and the ESRM-vs-best-baseline comparisons at severe contamination levels.","section":"6.2, Tables 5 and 6"}],"minor_comments":[{"comment":"In the sentence introducing the total loss, 'balancing hyperparamteres' is a typo for 'balancing hyperparameters.'","section":"5.3"},{"comment":"The phrase 'the fixed copy of ϕ(f(aug(X)))' is ambiguous; it should be clarified that this is the same network with gradients stopped, not a separate or fixed network copy.","section":"Eq. (4)"},{"comment":"The ROC curve would be more informative if the paper also reported the entropy distributions of real and synthetic samples for the same model, and ideally for a baseline model before ESRM training, so readers can see how the separation evolves.","section":"Fig. 7"},{"comment":"For ESRM, Relative Forgetting in Table 4 is lower at high contamination (e.g., 35.40 at P = 95% vs. 49.79 at P = 0%), which seems counterintuitive; the text says ESRM 'solves the problem of stability degradation' but does not explain why forgetting would decrease as contamination increases. A brief explanation would help.","section":"Tables 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study with a clear protocol and useful reproducible assets. The main risk is not the experimental evidence for the degradation, which is strong, but the load-bearing role of the entropy gap in explaining why the method works: only a single moderate AUC supports it, and the method's two components would lose their grounding if the gap shifts. I recommend major revision rather than acceptance because this mechanism claim needs direct supporting experiments. The authors' own limitation section is transparent, and the required additions are feasible within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-executed empirical paper and the main claims hold within the tested scope. It is the first systematic study of AI-generated image contamination in online continual learning, and the proposed ESRM method consistently reduces contamination-induced degradation across four datasets, six baselines, two contamination settings (single generator and five-generator mix), and ratios from 50% to 95%. The ablations show both components contribute, the gains are large at high contamination ratios, and the clean-setting performance is competitive. The paper ships code, reports error bars over 5-10 runs, and includes an honest limitations section. These are real strengths.\n\nWhat is new: the four observations (contamination hurts, buffer content matters, synthetic entropy is lower, embeddings are misaligned) are specific to online CL, and prior work only looked at standard training. The ESRM combination of entropy-based buffer selection and contrastive alignment is a simple, reasonable response to those observations.\n\nThe soft spots are mostly about external validity, not execution. The biggest one is the entropy discriminator. Both ES and RM rest on Observation 3: synthetic images have lower prediction entropy than real ones. The only explicit quantitative support is one ROC AUC of 0.71 (Fig. 7, In-100/SDXL), measured on an ESRM-trained model at the end of training. That is a moderate signal, and it could partly reflect the method rather than a stable property of synthetic data. The qualitative histograms across baselines (Fig. 8) help, and the fact that ESRM works across many settings is indirect evidence the entropy signal holds within the tested distribution. But there is exactly one AUC number, and the paper confirms it used only five diffusion models and simple prompts. If future generators produce higher-entropy or more diverse synthetic images, the method could lose its grounding. That is a stated limitation, not a hidden one.\n\nSecond, the paper never compares against a straightforward alternative: use a synthetic-image detector to filter the stream or buffer. The idealized \"Real Only\" result in Table 1 shows that near-clean performance is achievable if you know the status; a detector baseline would tell us whether ESRM's soft approach is actually needed. That is an easy experiment and its absence is a real gap.\n\nMinor: no significance tests, just error bars; the differences are large enough that this probably does not change any conclusion. The one self-citation (Ref. [42]) supports the self-distillation component and is fine.\n\nBottom line: the core claim is supported for these generators and prompts, and the paper says what is not covered. Whoever works on continual learning with web-scraped data, or on training with AI-generated content, gets value from this. It deserves a serious referee; my recommendation would be accept with requests for a detector comparison and a second entropy-discrimination measurement, ideally on the Mix setting or a held-out generator.","headline":"Solid empirical paper on synthetic contamination in online continual learning; ESRM works on the tested distributions, but the entropy discriminator's generalization is the main open question.","tokens_in":30255,"tokens_out":3509,"would_cite":true,"duration_ms":34820,"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":"Entropy-based replay buffer selection plus a contrastive alignment loss keeps online continual learners accurate even when 95% of training images are synthetic.","keywords":["online continual learning","synthetic data contamination","diffusion models","replay buffer","entropy-based sample selection","contrastive learning","class-incremental learning","catastrophic forgetting"],"falsifier":"Construct a contamination stream with a newer generator and diverse per-class prompts, train ESRM and a random-replay baseline, then measure the entropy separation (ROC AUC) between synthetic and real samples at the end of training; if AUC approaches 0.5 or if ESRM no longer outperforms random reservoir sampling, the method's grounding is refuted.","tokens_in":29301,"feed_emoji":"🖼️","tokens_out":6108,"duration_ms":54962,"temperature":0.7,"pith_summary":"This paper claims that synthetic images from diffusion models, when mixed into the training stream of an online continual learner, cause measurable accuracy loss that grows with the contamination ratio, and that this loss can be largely undone without knowing which images are synthetic. The proposed method, Entropy Selection with Real-synthetic similarity Maximization (ESRM), keeps a replay buffer stocked with high-entropy (real) samples and uses a contrastive loss to pull real and synthetic feature embeddings closer together. On four benchmarks, ESRM's accuracy drop stays small even at 70-80% contamination, while six existing replay-based methods lose up to roughly twenty points. The reason it works is an observed regularity: across the five tested diffusion generators, synthetic images receive lower prediction entropy from the continual learner and cluster more tightly in embedding space.","feed_headline":"Entropy-based replay buffer survives 95% AI-image contamination","feed_subtitle":"By storing high-entropy samples and aligning real/synthetic features, ESRM keeps accuracy nearly flat even at 95% contamination.","key_machinery":"Two mechanisms carry the method. Entropy Selection (ES) is a buffer-management rule: each incoming batch drops its lower-entropy half, applies reservoir sampling to the rest, and when replacement is needed selects the buffer sample of the same class with the lowest prediction entropy; all buffer entropies are refreshed at task ends. Real-synthetic similarity Maximization (RM) is a contrastive loss that maximizes cosine similarity between the high-entropy half of the stream and the low-entropy half, and between stream and memory, pulling embedded synthetic features toward real ones. A self-distillation term from a fixed copy of the augmented classifier prevents overconfidence. Together they exploit Observation 3 and Observation 4: entropy separates synthetic from real, and aligning embeddings removes the observed feature misalignment.","core_discovery":"The paper establishes that AI-generated images are a distinct corruption source for online class-incremental learning and that a replay method can be made robust to them using only learner-internal signals. Four observations support the claim: contamination degrades all tested baselines; keeping real images in the buffer nearly restores clean accuracy; synthetic images have lower prediction entropy; and synthetic features are better clustered, misaligning with real features. ESRM combines Entropy Selection, which discards low-entropy incoming samples and replaces same-class lowest-entropy buffer entries, with a Real-synthetic similarity Maximization loss that aligns high- and low-entropy groups and stream with memory. In experiments, ESRM is the least affected method under contamination, often improving on its own clean baseline, and remains usable at P=95% where baselines collapse.","pith_inferences":["If the entropy gap persists across future generators, ES could double as an unsupervised data-quality filter for any replay-based learner, not just the specific architecture here.","The paper's limitation that prompts were single-template 'an image of a class' suggests a natural stress test: with diverse user-style prompts, synthetic entropy may rise, and the method's advantage should shrink; measuring that would bound the method's validity.","A detector-free buffer policy that increases the real-to-synthetic ratio may also reduce bias amplification and diversity loss reported in non-continual settings, though the paper does not evaluate those outcomes.","The near-flat accuracy of ESRM across contamination ratios suggests online continual learning could tolerate streams that are mostly synthetic if a small anchor set of real data is stored, an implicit argument for curating small clean seeds rather than entire clean datasets."],"forward_implications":["At contamination ratios above 70%, existing replay methods lose accuracy by 5 to more than 20 points; ESRM keeps the drop under about 7 points on most benchmarks.","A learner does not need a separate synthetic-image detector or ground-truth synthetic labels; entropy and embedding alignment suffice.","Keeping the buffer mostly real is the dominant protective mechanism: with oracle real-only memory, ER matches clean accuracy even at 80% contamination, and ES approaches this without labels.","ESRM also transfers to domain-incremental learning, holding degradation below baselines on a 20-superclass split.","Because ESRM's plasticity is higher, it does not trade away new-task learning to gain stability."],"supporting_citations":[{"why":"Establishes Experience Replay, the replay baseline whose contamination response underpins Observations 1-2.","marker":"[33]"},{"why":"Provides the DER++ baseline that stores logits and is one of the six compared methods.","marker":"[5]"},{"why":"Provides the ER-ACE asymmetric-cross-entropy baseline used in contamination comparisons.","marker":"[6]"},{"why":"Supplies the OCM mutual-information replay baseline.","marker":"[16]"},{"why":"Supplies the GSA gradient self-adaptation baseline.","marker":"[17]"},{"why":"Supplies the OnPro prototype-learning baseline.","marker":"[45]"},{"why":"Motivates the self-distillation loss and the plasticity/stability metrics used in ESRM.","marker":"[42]"},{"why":"Is the supervised contrastive loss that RM adapts into inter-group alignment.","marker":"[22]"},{"why":"Supplies the diffusion generator used for the single-generator SDXL synthetic twin datasets.","marker":"[34]"},{"why":"Provides the reservoir-sampling procedure used inside Entropy Selection.","marker":"[40]"}],"fun_headline_variants":["Entropy-based replay withstands 95% AI-image contamination","ESRM keeps continual learning stable despite synthetic data","How to train online models when the web is full of AI images","Entropy selection rescues online learning from AI-generated data","Robust replay for online learning under synthetic contamination"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the empirical regularity that synthetic images receive lower prediction entropy than real images in the online continual learner, with the gap large enough that entropy-based buffer choice beats random choice.","fun_headline_variants_meta":{"raw":{"variants":["Entropy-based replay withstands 95% AI-image contamination","ESRM keeps continual learning stable despite synthetic data","How to train online models when the web is full of AI images","Entropy selection rescues online learning from AI-generated data","Robust replay for online learning under synthetic contamination"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000294,"raw_usage":{"total_tokens":1696,"prompt_tokens":915,"completion_tokens":781,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":701}},"tokens_in":531,"tokens_out":781,"duration_ms":7492,"temperature":1.0,"reasoning_tokens":701,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:47:42.191578+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a contamination stream with a newer generator and diverse per-class prompts, train ESRM and a random-replay baseline, then measure the entropy separation (ROC AUC) between synthetic and real samples at the end of training; if AUC approaches 0.5 or if ESRM no longer outperforms random reservoir sampling, the method's grounding is refuted.","supporting_citations":[{"cited_title":"Online prototype learning for online continual learning","cited_arxiv_id":null,"evidence_quote":"Supplies the OnPro prototype-learning baseline."},{"cited_title":"Improving Plasticity in Online Continual Learning via Collaborative Learning","cited_arxiv_id":"2312.00600","evidence_quote":"Motivates the self-distillation loss and the plasticity/stability metrics used in ESRM."},{"cited_title":"Supervised contrastive learning","cited_arxiv_id":null,"evidence_quote":"Is the supervised contrastive loss that RM adapts into inter-group alignment."},{"cited_title":"Experi- ence replay for continual learning","cited_arxiv_id":null,"evidence_quote":"Establishes Experience Replay, the replay baseline whose contamination response underpins Observations 1-2."},{"cited_title":"Dark experience for general continual learning: a strong, simple baseline","cited_arxiv_id":null,"evidence_quote":"Provides the DER++ baseline that stores logits and is one of the six compared methods."},{"cited_title":"Online continual learning through mutual information maximization","cited_arxiv_id":null,"evidence_quote":"Supplies the OCM mutual-information replay baseline."},{"cited_title":"Dealing with cross-task class discrimination in online continual learning","cited_arxiv_id":null,"evidence_quote":"Supplies the GSA gradient self-adaptation baseline."}],"review_version":1}