{"id":"5c6f8298-9925-4a3a-bebe-e080a67e5a26","arxiv_id":"2506.13097","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A prototype-based reconstruction method with an expanded set of prototypes and a prototype-based constraint achieves state-of-the-art anomaly detection on MVTec-AD, VisA, and Real-IAD.","lead":"Pro-AD is a new method for multi-class unsupervised anomaly detection that learns a large set of prototypes and adds a 'prototype-based constraint' to prevent anomalies from being reconstructed. It reports state-of-the-art scores on three industrial benchmarks, mainly by improving pixel-level anomaly localization.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 4 updates prototypes P^{l+1} from the test query Q^l, so the Prototype-based Constraint in Eq. 9 is an input-dependent residual, not a normal-only constraint; this undermines the core claim that it blocks Soft Identity Mapping.","rationale":"The reader's weakest assumption identifies the additive residual P^{l+1}_reg as a learned term that could encode anomalous information. My concern is stronger and more specific: Eq. 4-5 make the prototypes themselves functions of the test query Q^l, so P^{l+1}_reg is not merely a learned residual but an input-dependent residual. This directly contradicts the paper's narrative that the Prototype-based Constraint limits the expressive capability of prototypes or prevents Soft Identity Mapping. The central claim of the paper is that Pro-AD achieves state-of-the-art performance by learning comprehensive prototypes and applying this constraint. If the constraint is actually an input-conditioned layer, the proposed mechanism is internally inconsistent as written. The ablation in Table 5 provides some evidence that the component helps, but it does not test the mechanism; a generic capacity increase or the dynamic prototype update could explain the gain. Given the absence of code, the lack of error bars, and the per-dataset tuning of dropout rates, the current version does not adequately support the central claim. I therefore recommend rejection of the paper in its present form, with the concrete frozen-prototype experiment as a way for the authors to demonstrate that the constraint works as claimed.","tokens_in":9821,"tokens_out":6755,"duration_ms":74483,"concrete_test":"Run the trained model on the MVTec-AD test set with two variants: (A) the current dynamic update from Eq. 4-5, and (B) a frozen prototype variant in which P^{l+1} is computed once from normal training features and held constant during inference (e.g., using a normal reference batch or a detached prototype update). If variant B retains image- and pixel-level AUROC within a small tolerance, the input-dependent update is not necessary and a normal-anchored constraint could work as claimed. If variant B degrades substantially, the model is exploiting the input-dependent prototype update, so the Prototype-based Constraint is not preventing anomalous reconstruction. Additionally, measure the norm of the prototype-update term LCA(P^l, Q^l, Q^l) on anomalous versus normal test inputs; a large anomaly-induced update would directly show that anomalies influence the 'prototype' reference.","verdict_should_be":"REJECT","load_bearing_attack":"In the Dynamic Bidirectional Decoder, Eq. 4-5 update the prototypes from the target feature itself: P^{l+1} = P^l + LCA(P^l, Q^l, Q^l) + FFN(...). Thus, at inference, P^{l+1} is a function of the potentially anomalous query Q^l. The Prototype-based Constraint in Eq. 9 is then P^{l+1}_reg = FFN(P^{l+1}), and the final reconstruction is f^l_D = f^l_rec + P^{l+1}_reg, where f^l_rec = Q^l + LCA(Q^l, P^{l+1}, P^{l+1}). Both terms depend on Q^l, so the 'constraint' is not a fixed normal reference: it is an input-conditioned residual that can carry anomalous information. Section 3.4 argues that f^l_rec is 'implicitly tasked' with reconstructing f^l_en - P^{l+1}_reg, but this decomposition is arbitrary when both terms are learned functions of the same input. Nothing in the loss or architecture forces P^{l+1}_reg to be normal-only; the model is free to split reconstruction between f^l_rec and P^{l+1}_reg in any way that matches the supervised normal outputs. Consequently, the claimed mechanism for preventing Soft Identity Mapping is not established by the presented equations. The only empirical support is the one-dataset ablation in Table 5, which shows a performance improvement but cannot distinguish the stated mechanism from a generic increase in model capacity or from the input-dependent prototype update. This is an internal correctness concern about the central contribution, not a dispute about external consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents Pro-AD, a multi-class unsupervised anomaly detection method that learns an expanded set of learnable prototypes and reconstructs target features with a dynamic bidirectional decoder. The main novel component is a Prototype-based Constraint that adds a feed-forward transformation of the updated prototypes to the naive reconstruction, with the stated goal of preventing anomalies from being reconstructed through attention, a failure mode the authors call 'Soft Identity Mapping.' The paper reports state-of-the-art or near-state-of-the-art image- and pixel-level metrics on MVTec-AD, VisA, and Real-IAD, together with an ablation study on MVTec-AD.","tokens_in":10318,"tokens_out":6014,"duration_ms":56368,"significance":"The paper addresses a relevant problem in multi-class unsupervised anomaly detection and reports strong empirical results. If the proposed mechanism were established, the simple recipe of expanded prototypes plus a residual constraint would be practically useful. The main weakness is that the central mechanistic claim is not supported by the equations: the prototype update and the Prototype-based Constraint both depend on the potentially anomalous query feature. The empirical evidence is also limited by the absence of error bars, the lack of code release, per-dataset hyperparameter choices, and a one-dataset ablation. The paper does not provide machine-checked proofs or reproducible code, so the SOTA claims rest entirely on the reported tables.","major_comments":[{"comment":"The Prototype-based Constraint is not a normal-only reference. In Eq. (4), the updated prototype P^{l+1} is computed by cross-attending to the target feature Q^l, so at inference P^{l+1} is a function of the potentially anomalous query. Consequently, both terms in Eq. (9), f_D^l = f_rec^l + P_reg^{l+1}, are functions of Q^l, and the statement that f_rec^l is 'implicitly tasked with reconstructing f_en^l - P_reg^{l+1}' is an arbitrary decomposition when both summands are learned from the same input. Nothing in the loss or architecture restricts P_reg^{l+1} to encode only normal information. The claimed disruption of Soft Identity Mapping is therefore not established by the presented equations; the ablation in Table 5 cannot distinguish the stated mechanism from a generic increase in model capacity or from an input-dependent residual.","section":"Section 3.4, Eqs. (4)-(9)"},{"comment":"The paper's core premise—that increasing the number of prototypes causes Soft Identity Mapping and that the Prototype-based Constraint removes it—is supported only by a qualitative argument and a single ablation on MVTec-AD. No controlled experiment varies the number of prototypes N, and no direct measurement of reconstruction fidelity on anomalous regions with and without the constraint is reported. Without such evidence, the phenomenon and the proposed remedy are not demonstrated as claimed.","section":"Section 3.3 and Table 5"},{"comment":"The SOTA claims are not supported with uncertainty estimates. The reported gains over INP-Former are small in several cases (e.g., image-level AUROC 99.8 vs. 99.7 on MVTec-AD and 91.5 vs. 90.5 on Real-IAD), and the per-dataset dropout rates (0.2, 0.3, 0.4) are given without a validation procedure. No code is released, so the results cannot be independently checked. Error bars or an explicit multi-seed evaluation protocol are needed to establish that the improvements are statistically meaningful.","section":"Section 4.1 and Tables 2-4"}],"minor_comments":[{"comment":"The notation 'Dropout prob' is ambiguous; please specify whether 'prob' is the dropout rate or the keep probability.","section":"Section 3.2, Eqs. (1)-(2)"},{"comment":"The loss formula uses T(·) while the text defines F(·) as the flatten operation, and the terms G_D^l(h,w) and g_d(·,·) are not defined in the text. Please make the notation consistent and define all quantities.","section":"Section 3.5, Eq. (11)"},{"comment":"The prototype parameter count (605,952) is not justified; please provide the formula used (e.g., N x C) and state the values of N and C.","section":"Table 1"},{"comment":"The paragraph states the method is effective only for certain logical anomalies and acknowledges weaker performance on MVTec-LOCO; this admission tempers the broad robustness claim and should be reflected in the conclusion.","section":"Limitations and Future Works"},{"comment":"Figure 3 is not referenced in the text; please add an in-text citation where the anomaly localization visualizations are discussed.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible empirical contribution, but its central mechanism is not established from the equations. I would ask for a formal or controlled demonstration that P_reg^{l+1} behaves as a normal-only reference, plus error bars and code release. If the mechanism cannot be defended, the paper could still be published as an engineering improvement, provided the claims are repositioned accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2506.13097.\n\nThe paper reports a genuine improvement on three industrial anomaly detection benchmarks, and it names a real problem: with many prototypes, attention can reconstruct anomalies too well (\"Soft Identity Mapping\"). The expanded prototype set and the bidirectional decoder with shared weights are reasonable engineering, and the parameter count stays below INP-Former. The ablation on MVTec shows each component adds something, especially at pixel level. They also honestly note the method struggles on MVTec-LOCO logical anomalies. That part is solid and I'd credit it.\n\nThe soft spot is the central mechanism. Section 3.4 claims the Prototype-based Constraint stops the network from reconstructing anomalies because the reconstruction is 'implicitly tasked' with producing f_en − P_reg. But Eq. 4–5 update the prototypes P^{l+1} from the query Q^l itself. So P_reg = FFN(P^{l+1}) is an input-conditioned residual, not a fixed normal reference. Nothing in the loss or architecture forces it to be normal-only; the network is free to split reconstruction between f_rec and P_reg in any way that matches the training data. The claimed mechanism for blocking Soft Identity Mapping is not established by the equations, and the Table 5 ablation can't separate the mechanism from a generic capacity increase. That is a real flaw in the paper's central theoretical claim.\n\nOther issues are more minor. No code, no error bars, and dropout rates are tuned per dataset. The 'SOTA' claim is slightly overstated—on VisA and Real-IAD they are second on a couple of pixel-level metrics. The related work and citations look appropriate; they build directly on INP-Former and Dinomaly and say so.\n\nOverall, the empirical work is plausible and the architecture is well specified, but the explanation of why it works is shaky. This is exactly the kind of paper that a serious referee can fix: require code, variance reporting, and either a formal argument or a careful control experiment showing that P_reg actually acts as a norm rather than a residual. I would send it to review, not desk reject, and push for major revision on the mechanism section.\n\nFor a reading group, it's a maybe—the flaw is instructive. I wouldn't cite it in my own work in the next year.","headline":"Real empirical gains on three benchmarks, but the paper's core mechanism—the Prototype-based Constraint—is an input-dependent residual, not a normal-only constraint, so the central theoretical claim doesn't hold as stated.","tokens_in":10732,"tokens_out":4322,"would_cite":false,"duration_ms":38983,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Pro-AD claims a large prototype set plus a prototype-derived constraint solves the Soft Identity Mapping problem and reaches state-of-the-art multi-class anomaly detection scores on three industrial benchmarks.","keywords":["unsupervised anomaly detection","multi-class anomaly detection","prototype learning","reconstruction-based detection","soft identity mapping","dynamic bidirectional decoder","linear cross-attention","industrial anomaly detection"],"falsifier":"Train Pro-AD on MVTec-AD, then remove the Prototype-based Constraint (set $\\mathbf{P}^{l+1}_{\\rm reg}=0$) and measure image-level AUROC on anomaly classes with large spatial extent, such as logical anomalies. If the gain from the constraint disappears or reverses on a broader set of classes, or if a test shows an anomaly being reconstructed through $\\mathbf{P}^{l+1}_{\\rm reg}$ alone, the mechanism is not doing what the paper claims.","tokens_in":9610,"feed_emoji":"🔍","tokens_out":4329,"duration_ms":37136,"temperature":0.7,"pith_summary":"Pro-AD is a reconstruction-based method for multi-class unsupervised anomaly detection: one model trained only on normal images must flag defects across many object classes. The paper argues that prototype-based reconstruction faces a trade-off: too few learnable prototypes cannot represent normal variation, while too many let the attention mechanism reconstruct anomalies as well as normal regions, a failure it calls the \"Soft Identity Mapping\" problem. Pro-AD's proposal is to use a large set of learnable prototypes plus a Prototype-based Constraint that forces the final reconstruction to be composed of a weighted prototype combination and a regularized prototype term, so anomalies cannot be rebuilt by a simple weighted mix. On MVTec-AD, VisA, and Real-IAD the method reports state-of-the-art or near-top image- and pixel-level scores, with the largest gains at the pixel level.","feed_headline":"New prototypes block anomaly reconstruction in defect detection","feed_subtitle":"Pro-AD pairs an expanded prototype set with a constraint that stops attention from rebuilding defects, lifting pixel-level scores.","key_machinery":"The load-bearing object is the Prototype-based Constraint, defined as $\\mathbf{P}^{l+1}_{\\rm reg} = \\mathrm{FFN}(\\mathbf{P}^{l+1})$ added to the attention reconstruction $\\mathbf{f}^l_{\\rm rec}$, giving $\\mathbf{f}^l_D = \\mathbf{f}^l_{\\rm rec} + \\mathbf{P}^{l+1}_{\\rm reg}$. This term recasts the attention reconstruction as learning a residual complementary to a prototype-derived constant, so no single anomaly can be rebuilt by weighted combination alone. It works inside a Dynamic Bidirectional Decoder whose prototype-update and target-reconstruction branches share weights and use linear cross-attention, with an Adaptive Noisy Bottleneck keeping prototypes and query features in the same feature space.","core_discovery":"The central claim is that the \"Soft Identity Mapping\" problem, where an enlarged prototype set gives the attention mechanism enough expressive power to reconstruct anomalous features, can be broken by adding a Prototype-based Constraint inside the decoder. The reconstruction of each target feature is written as the sum of a linear-cross-attention reconstruction from prototypes and a per-position regularized term derived from the same prototypes through a feed-forward network; this changes the learning objective so that the attention part must reconstruct the encoder feature minus the regularized term, preventing the prototypes from acting as an identity map. With this constraint, an expanded set of prototypes (as many as there are image patches) can safely supply comprehensive normal information, and the paper reports the method outperforms prior multi-class anomaly detectors on three industrial benchmarks, particularly in pixel-level anomaly localization.","pith_inferences":["An implication the paper leaves implicit is that the constraint deserves testing on logical-anomaly datasets like MVTec-LOCO, where the paper itself notes limited performance; such a test would show whether the mechanism mainly helps structural defects or generalizes to context-dependent anomalies.","Because the constraint is an additive learned residual, a natural extension is to bound or normalize $\\mathbf{P}^{l+1}_{\\rm reg}$ so it cannot exceed a fixed budget, turning the soft constraint into a true capacity limit.","The shared weights between prototype learning and reconstruction suggest the same decoder could be reused for few-shot or zero-shot anomaly detection if the initial prototypes are replaced by a small set of normal exemplars.","If the residual interpretation is correct, similar constraints could improve other attention-based reconstruction models, not just prototype-based ones, by forcing the attention output to be complementary to a fixed reference."],"forward_implications":["If Pro-AD's mechanism is correct, pixel-level anomaly localization improves enough to matter for industrial inspection, with pixel-level AP and F1 gains of several points over the prior best on MVTec-AD.","A single multi-class model can cover diverse industrial categories without a per-class memory bank, since the prototypes are shared and learned rather than stored per class.","The Prototype-based Constraint gives a generic recipe for prototype-based reconstruction models: constrain the reconstruction output by a prototype-derived term, not just by the number of prototypes.","The method's robustness on Real-IAD, the largest benchmark, suggests the approach scales to realistic multi-view industrial settings."],"supporting_citations":[{"why":"The prototype-transformer baseline that Pro-AD extends; supplies the limited-prototype setup and the 'Soft Identity Mapping' failure mode.","marker":"[17]"},{"why":"Dinomaly; supplies the noisy bottleneck and less-is-more reconstruction philosophy that Pro-AD adapts.","marker":"[9]"},{"why":"PatchCore; supplies the pre-stored normal prototype memory-bank baseline the paper argues is memory-heavy and misaligned.","marker":"[21]"},{"why":"UniAD; defines the multi-class unsupervised anomaly detection task and provides a strong reconstruction baseline.","marker":"[24]"},{"why":"MVTec-AD; the primary dataset used for the main experiments and the only dataset used in the ablation.","marker":"[3]"},{"why":"VisA; a second benchmark dataset used to test robustness.","marker":"[28]"},{"why":"Real-IAD; the largest and most challenging benchmark used to test scalability.","marker":"[23]"},{"why":"The pre-trained DINOv2 backbone that Pro-AD adopts as its feature extractor.","marker":"[4]"}],"fun_headline_variants":["Breaking soft identity mapping for sharper anomaly detection","Prototype constraint stops attention from rebuilding defects","Expanded prototypes with constraint: better anomaly detection","Pro-AD: constraints block anomaly reconstruction via attention","Multi-class anomaly detection improved by prototype constraint"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the learned term added to the reconstruction truly limits how well anomalies can be rebuilt, rather than simply being another learned channel that can encode anomaly information; the paper gives no proof of this and tests it on one dataset only.","fun_headline_variants_meta":{"raw":{"variants":["Breaking soft identity mapping for sharper anomaly detection","Prototype constraint stops attention from rebuilding defects","Expanded prototypes with constraint: better anomaly detection","Pro-AD: constraints block anomaly reconstruction via attention","Multi-class anomaly detection improved by prototype constraint"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000556,"raw_usage":{"total_tokens":2648,"prompt_tokens":946,"completion_tokens":1702,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1643}},"tokens_in":562,"tokens_out":1702,"duration_ms":13378,"temperature":1.0,"reasoning_tokens":1643,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:05:15.400559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Pro-AD on MVTec-AD, then remove the Prototype-based Constraint (set $\\mathbf{P}^{l+1}_{\\rm reg}=0$) and measure image-level AUROC on anomaly classes with large spatial extent, such as logical anomalies. If the gain from the constraint disappears or reverses on a broader set of classes, or if a test shows an anomaly being reconstructed through $\\mathbf{P}^{l+1}_{\\rm reg}$ alone, the mechanism is not doing what the paper claims.","supporting_citations":[{"cited_title":"Towards total recall in industrial anomaly detection","cited_arxiv_id":null,"evidence_quote":"PatchCore; supplies the pre-stored normal prototype memory-bank baseline the paper argues is memory-heavy and misaligned."},{"cited_title":"A unified model for multi-class anomaly detection","cited_arxiv_id":null,"evidence_quote":"UniAD; defines the multi-class unsupervised anomaly detection task and provides a strong reconstruction baseline."},{"cited_title":"MVTec AD–A comprehensive real-world dataset for unsupervised anomaly detection","cited_arxiv_id":null,"evidence_quote":"MVTec-AD; the primary dataset used for the main experiments and the only dataset used in the ablation."},{"cited_title":"Spot-the-difference self-supervised pre-training for anomaly detection and segmentation","cited_arxiv_id":null,"evidence_quote":"VisA; a second benchmark dataset used to test robustness."},{"cited_title":"Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection","cited_arxiv_id":null,"evidence_quote":"Real-IAD; the largest and most challenging benchmark used to test scalability."}],"review_version":1}