{"id":"a312f406-8f5b-4e6d-a1ae-139821c3a30e","arxiv_id":"2411.09558","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"Adaptive Deviation Learning combines a soft-label deviation loss with instance reweighting to improve visual anomaly detection when the training set contains unlabeled anomalies.","lead":"This paper introduces Adaptive Deviation Learning, a method that reweights training samples and uses soft anomaly scores to detect defects when the training data itself is contaminated with anomalies. The reported gains over prior methods on MVTec and VisA are undercut by an evaluation design that pulls training contamination from the test set.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation protocol leaks test anomalies into training as Gaussian-noised copies; the reported SOTA under contamination is not established by this evidence.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing issue: the contamination protocol in Section 4.1 is not a valid measure of robustness because it draws contamination from the test-set anomalies. I agree with that assessment. The paper's central claim requires that ADL handle unlabeled anomalies that corrupt normal training data; for that claim to be tested, the contamination must be independent of the test anomalies. Sampling from the test anomalies and adding Gaussian noise makes the task artificially easy and uneven across methods: discriminatively trained methods such as DRÆM, DestSeg, and ADL can exploit the noisy copies, while memory-based PatchCore may be less able to do so. The reported comparisons therefore do not establish state-of-the-art performance under real data contamination. This is not an internal inconsistency in the method: the deviation-learning formulation and the reweighting scheme in Section 3 are coherent, and the ablation shows plausible incremental gains. But the empirical core is compromised by the evaluation protocol. I would not change the reader's rejection; if anything, the broken code link and missing error bars reinforce it. The disjoint-contamination test I propose would settle whether the method's advantage is real or an artifact of near-duplicate leakage, and if the method passes that test the paper could be resubmitted with a corrected protocol.","tokens_in":13441,"tokens_out":3289,"duration_ms":33413,"concrete_test":"Redo the 10/15/20% contamination experiments with contamination sourced from anomalies disjoint from the test set: for each class, hold out half of the anomaly images for testing and use the other half (plus Gaussian noise) for training contamination, or replace the contamination source with an external outlier set such as DTD textures. Keep the same model, hyperparameters, and evaluation. If ADL's average AUC margin over DestSeg and LOE at 15-20% contamination shrinks or reverses, the SOTA claim is an artifact of near-duplicate leakage; if the margin persists, the robustness claim is supported. Also report per-seed mean and standard deviation over at least three runs and an epsilon=0 clean-data baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 states that artificial contamination is introduced by \"sampling from the test set anomalies and adding zero-mean Gaussian noise with a relatively large variance.\" The test set then contains the original anomalies while the training set contains noisy variants of those same images. This is a train/test leakage: the model is trained on near-duplicates of exactly the images it is later scored on. Any method capable of memorizing these noisy copies will report inflated AUC, and the comparison with PatchCore, DestSeg, and DRÆM measures adaptation to known anomaly appearance rather than robustness to realistic, previously unseen contamination. The central claim of state-of-the-art performance under data contamination therefore rests on an evaluation premise that is false for the intended real-world setting. The algorithm itself is not internally inconsistent; the weakness is in the evidence. Missing error bars, the absence of a clean-data baseline, and the non-functional code link further weaken reproducibility, but the test-derived contamination protocol is the single load-bearing flaw.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adaptive Deviation Learning (ADL), a self-supervised framework for visual anomaly detection on contaminated training data. ADL combines a soft deviation loss with instance reweighting via constrained optimization, using a classification head to produce soft labels and a segmentation loss for stability. The authors claim state-of-the-art performance under 10%–20% training contamination on MVTec AD and VisA, supported by tables and robustness curves. The key evaluation protocol, however, is flawed: Section 4.1 states that artificial contamination is created by sampling from the test-set anomalies and adding Gaussian noise, so the training set contains near-duplicates of the exact test images being scored. This leaks test information into training and invalidates the reported AUC improvements.","tokens_in":13629,"tokens_out":4868,"duration_ms":47571,"significance":"If the method worked as advertised, it would be a useful contribution to anomaly detection under label contamination, since most existing methods assume clean training data. The paper introduces a plausible combination of deviation learning and sample reweighting, and the ablation study in Table 2 provides some evidence that the individual components contribute. However, the empirical evidence for the headline claim is not trustworthy because of the train/test leakage, and the paper also lacks statistical error bars, a clean-data baseline, and a working code link. The contribution is therefore not established; its significance remains conditional on a valid evaluation.","major_comments":[{"comment":"The contamination protocol leaks test anomalies into training. The text states: 'we introduced artificial anomalies by sampling from the test set anomalies and adding zero-mean Gaussian noise with a relatively large variance to the samples.' Since the test set contains the original anomaly images and the training set contains Gaussian-noised copies of those same images, the test set is no longer independent of the training set. A model can memorize the noisy copies and receive inflated AUC on the original test anomalies, so the reported comparison against PatchCore, DestSeg, DRÆM, and LOE does not measure robustness to previously unseen contamination. This directly undermines contribution (v) in Section 1. To fix this, the contamination should be drawn from a held-out subset of anomalies that is disjoint from the evaluation set, or from an external anomalous source that does not overlap with the test images. The authors should also verify that the noise transformation itself does not create near-duplicates by, for example, reporting results with and without noise on a disjoint split.","section":"Section 4.1"},{"comment":"The reported results contradict the text. Section 4.2 claims that 'our method outperforms the top-performing DestSeg model at contamination levels of 15% and 20%, and surpasses other competitive baselines across all contamination levels.' However, Table 1 shows that on MVTec at 10% contamination, DestSeg has an average AUC-ROC of 0.927 while ADL has 0.924, so ADL does not surpass DestSeg at that level. The claim is therefore inaccurate. Moreover, the table reports only mean values without standard deviations, making it impossible to judge whether the small differences (e.g., 0.908 vs 0.894 at 15%) are statistically meaningful. The authors should provide error bars or at least standard deviations across runs, and correct the overstatement.","section":"Section 4.2, Table 1"},{"comment":"The soft-deviation loss in Eq. (5) depends on p(x_i), which is produced by a classification head trained with labels that alternate between the (corrupted) original labels y_i and the k-means assignments \\tilde{y}_i computed from the anomaly score \\psi_K(x_i) of the same scoring network. This creates a circular dependency: the soft labels used to train the scoring network are derived from the scoring network's own output. The paper does not analyze the fixed-point behavior of this loop or provide an ablation that trains the classification head using only fixed synthetic labels. Without such an analysis, it is unclear whether Eq. (5) actually reduces contamination or merely propagates the network's own bias. The authors should add an experiment that isolates the effect of the k-means-derived labels from the soft-deviation objective.","section":"Equations (5) and (6)"},{"comment":"The robustness analysis lacks a clean-data baseline. All experiments in Figure 2 start at 5% contamination, and there is no comparison to training on uncontaminated data. As a result, the claim that the method shows 'stability and robustness' is not anchored: a method that is robust but has a large clean-data performance drop would be unattractive. The authors should include a 0% contamination condition for all methods and report the absolute performance level, not just the rate of decline with contamination.","section":"Section 4.2, Figure 2"}],"minor_comments":[{"comment":"The code link in the footnote is malformed: 'https://github.com/anindyasdas/ADL4V AD/' contains a space and does not resolve to a working repository. Please provide a valid anonymous link for review.","section":"Abstract and Section 1"},{"comment":"The cable row for PatchCore contains '0.7370.713' with a missing separator; it should read '0.737/0.713' or similar. Please check all table entries for formatting.","section":"Table 1"},{"comment":"In the last sentence of the first paragraph, 'approached' should be 'approach'.","section":"Section 2"},{"comment":"The comparison with LOE is not fully clear: LOE applies contamination in a learned latent space, while the authors add Gaussian noise in image space. The paper should explain how the noise variance is chosen and whether it is intended to mimic a specific real-world distortion, or whether it is arbitrary.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The core idea is not without interest, but the experimental protocol is fundamentally invalid for the stated claim. The leakage of test anomalies into training is not a minor issue; it invalidates every empirical comparison and the central SOTA claim. A revision could potentially fix this by re-running all experiments with a proper disjoint contamination split, but that would be a substantial new evaluation rather than a targeted correction. Given the current state of the evidence, I cannot recommend publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The method has a genuinely interesting combination: soft deviation loss that replaces hard labels with classifier probabilities, adaptive instance reweighting from constrained optimization, and Perlin-mask pseudo-anomalies. The ablation study is honest and shows each component helps a little, and the sensitivity analysis on λ and α is useful. The writing is clear and the related work is properly cited.\n\nThe load-bearing problem is the evaluation protocol in Sec. 4.1. They contaminate the training set by sampling from test-set anomalies and adding zero-mean Gaussian noise with large variance. The test set contains the original anomalies, so the model is trained on near-duplicates of exactly the images it is later scored on. The AUC gains over PatchCore, DestSeg, and DRÆM likely reflect memorization of those noisy copies, not robustness to unseen contamination. The stress-test note is correct: this is train/test leakage, and the central claim of state-of-the-art performance under contamination is not established.\n\nThere are secondary issues: Table 1 reports only mean AUC-ROC without standard deviations, Figure 2 has no error bars, and there is no clean-data baseline to show the actual cost of contamination. The code link is dead, preventing reproduction. The k-means alternation for the BCE labels introduces a mild circularity, but that is secondary and not fatal.\n\nWho is this for? Researchers working on anomaly detection with noisy training data, and anyone designing contamination benchmarks. The paper deserves a serious referee, but only to demand a corrected protocol: contaminate with anomalies from a held-out source, not from the test set. If the results survive that, the method is worth publishing. As it stands, I would not accept it; I would send it back for major revision with the evaluation redone.","headline":"The soft-deviation + reweighting idea is worth a second look, but the contamination protocol leaks test anomalies into training, so the headline SOTA claim is not supported.","tokens_in":14131,"tokens_out":1687,"would_cite":false,"duration_ms":18069,"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":"The paper claims a new self-supervised method, adaptive deviation learning, that keeps visual anomaly detection accurate even when up to 20% of training images are anomalies.","keywords":["anomaly detection","data contamination","deviation learning","sample reweighting","self-supervised learning","soft labels","industrial defect detection"],"falsifier":"Re-run the comparison with training contamination drawn from anomaly classes that are completely absent from the test set; if ADL's advantage over DestSeg and LOE shrinks or vanishes, the claimed state-of-the-art robustness is an artifact of the near-duplicate contamination protocol.","tokens_in":13240,"feed_emoji":"🔍","tokens_out":6274,"duration_ms":51235,"temperature":0.7,"pith_summary":"The paper proposes Adaptive Deviation Learning (ADL), a self-supervised method for visual anomaly detection that is explicitly designed for training sets containing unlabeled anomalies. ADL learns an anomaly scoring function in which normal images score near a reference mean drawn from a standard Gaussian, while anomalous images must deviate from that mean by a margin. To stay robust when some training images are themselves anomalies, the method replaces hard labels with soft likelihood estimates from a classification head and adaptively reweights each instance's loss through a constrained optimization solved per mini-batch. The authors claim that on the MVTec and VisA benchmarks, at contamination levels of 10%, 15%, and 20%, ADL outperforms previous methods including PatchCore, DestSeg, DRÆM, and LOE, and remains relatively stable as contamination rises.","feed_headline":"Adaptive deviation learning tops anomaly detection on contaminated data","feed_subtitle":"Soft-deviation loss and sample reweighting stay reliable when up to 20% of training images are anomalies.","key_machinery":"The central object is the adaptive deviation learning objective, a combination of a soft deviation loss and an instance-reweighting scheme. Deviation learning scores each image by the mean of its top-K patch-level anomaly scores, pushes normal scores toward the mean of a prior Gaussian reference, and forces anomaly scores at least gamma standard deviations away; the soft version swaps hard binary labels for the probabilities produced by a k-means-refined classification head. Instance reweighting solves a constrained optimization that keeps the weight distribution close to uniform under KL, reverse-KL, or alpha-divergence, yielding closed-form per-sample weights computed per mini-batch. A segmentation head trained with focal loss on Perlin-noise masks stabilizes the features.","core_discovery":"On the authors' terms, the paper's central discovery is that deviation learning, which was previously used with a small set of clean labeled anomalies, can be made to work in a fully self-supervised setting with contaminated unlabeled normal data. The key to this is a soft-deviation objective, where the binary anomaly label is replaced by the probability output of an auxiliary classification head, together with an adaptive sample-importance mechanism that assigns closed-form weights to individual instances based on their loss values, derived from a constrained divergence-minimization problem. With these components, the top-K patch-based anomaly score trained on Perlin-noise pseudo-anomalies and contaminated originals produces image-level AUC-ROC scores that the paper reports as state-of-the-art on MVTec (for example, 92.4/90.8/87.3 at 10/15/20% contamination) and VisA (84.4/82.7/80.7).","pith_inferences":["Going beyond the paper, the closed-form reweighting formulas are not specific to deviation learning and could be dropped into other self-supervised anomaly objectives, such as reconstruction-based or contrastive losses, whenever contamination is suspected.","The paper's contamination protocol uses Gaussian-noised copies of test-set anomalies in the training set, which creates near-duplicates across the train/test split; a disjoint-anomaly evaluation would be a stronger test of the robustness claim.","An immediate testable extension is to vary the noise variance of the injected contamination, since very large Gaussian noise may make the contamination trivially detectable and inflate the apparent robustness.","The failure cases listed (cable, transistor, screw, macaroni2) suggest that texture-like or fine-grained anomalies remain difficult under contamination; a targeted analysis of those categories could reveal where the soft-label head is misled."],"forward_implications":["Anomaly detectors trained with ADL can be applied to real-world data collection pipelines where clean normal data is not guaranteed, without a separate data-purification stage.","The per-mini-batch closed-form reweighting adds negligible computational overhead compared to standard deviation learning.","The soft-deviation objective reduces sensitivity to the true contamination ratio, which the framework does not need to know in advance.","The reported results suggest that ADL degrades more gracefully than PatchCore and DRÆM as contamination rises from 10% to 20%.","Ablations show that the combination of soft-deviation and reweighting, rather than either alone, drives most of the gain on both datasets."],"supporting_citations":[{"why":"It supplies the original deviation loss and network formulation that ADL adapts to the contaminated setting.","marker":"[25]"},{"why":"It supplies the top-K patch scoring and deviation-based few-shot objective that ADL extends with soft labels.","marker":"[24]"},{"why":"It provides the closed-form instance-reweighting formulas under KL, reverse-KL, and alpha-divergence that ADL solves per mini-batch.","marker":"[17]"},{"why":"It supplies the synthetic anomaly generation using Perlin-noise masks that ADL uses for self-supervision.","marker":"[43]"},{"why":"It is the main contamination-robust baseline whose contamination setup ADL compares against in the experiments.","marker":"[29]"},{"why":"It is a leading non-contamination baseline that the paper modifies to include contamination for a fair comparison.","marker":"[31]"},{"why":"It is the strongest competing baseline in the reported tables, which ADL claims to surpass at higher contamination levels.","marker":"[45]"},{"why":"It provides the MVTec AD dataset, one of the two benchmark suites used for the empirical evaluation.","marker":"[3]"},{"why":"It provides the VisA dataset, the second benchmark suite used for the empirical evaluation.","marker":"[46]"}],"fun_headline_variants":["Soft-deviation loss and reweighting tame contaminated anomaly detection","Adaptive deviation learning stays robust when 20% of data is anomalous","Anomaly detection that shrugs off contaminated training data","Deviation learning with sample weights tops MVTec even with noise","Reweighted deviation scores beat noisy anomaly detection baselines"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results assume that contaminating the training set with Gaussian-noised copies of the test-set anomaly images is a valid measure of robustness, even though those same test images then also appear, in nearly identical form, in the test set.","fun_headline_variants_meta":{"raw":{"variants":["Soft-deviation loss and reweighting tame contaminated anomaly detection","Adaptive deviation learning stays robust when 20% of data is anomalous","Anomaly detection that shrugs off contaminated training data","Deviation learning with sample weights tops MVTec even with noise","Reweighted deviation scores beat noisy anomaly detection baselines"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1448,"prompt_tokens":923,"completion_tokens":525,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":438}},"tokens_in":539,"tokens_out":525,"duration_ms":5091,"temperature":1.0,"reasoning_tokens":438,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:31:46.717743+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison with training contamination drawn from anomaly classes that are completely absent from the test set; if ADL's advantage over DestSeg and LOE shrinks or vanishes, the claimed state-of-the-art robustness is an artifact of the near-duplicate contamination protocol.","supporting_citations":[{"cited_title":"Draem- a discriminatively trained reconstruction embedding for sur- face anomaly detection","cited_arxiv_id":null,"evidence_quote":"It supplies the synthetic anomaly generation using Perlin-noise masks that ADL uses for self-supervision."},{"cited_title":"Latent outlier exposure for anomaly detec- tion with contaminated data","cited_arxiv_id":null,"evidence_quote":"It is the main contamination-robust baseline whose contamination setup ADL compares against in the experiments."},{"cited_title":"Towards to- tal recall in industrial anomaly detection","cited_arxiv_id":null,"evidence_quote":"It is a leading non-contamination baseline that the paper modifies to include contamination for a fair comparison."},{"cited_title":"Destseg: Segmentation guided denoising student-teacher for anomaly detection","cited_arxiv_id":null,"evidence_quote":"It is the strongest competing baseline in the reported tables, which ADL claims to surpass at higher contamination levels."},{"cited_title":"Spot-the-difference self-supervised pre- training for anomaly detection and segmentation","cited_arxiv_id":null,"evidence_quote":"It provides the VisA dataset, the second benchmark suite used for the empirical evaluation."}],"review_version":1}