acceptodds
Under review as a conference paper at ICLR 2027

Value-order Decomposition for Generalist Anomaly Detection

Abstract

Industrial anomaly detection suffers from limited data, making cross-domain generalization particularly challenging. Generalist Anomaly Detection (GAD) aims to train a unified model on a source domain that can effectively detect anomalies in unseen target domains. In the initial semantic feature space, strong entanglement between anomalies and object categories or defect types hinders effective generalization across domains. Recent works address this issue by projecting features into a residual space; however, such methods primarily increase cross-domain overlap for normal features, while anomalous features remain specific to object categories, defect types and data domains, leading to poor alignment and generalization. To address this limitation, we propose Value-order Decomposition (VOD), a simple yet effective technique that bridges three types of generalization gaps across object categories, defect types (including real and synthetic defects), and data domains. VOD disentangles and suppresses object-category-, defect-type-, and domain-specific information, promoting alignment within normal and abnormal samples while preserving their separability, thereby enabling robust generalization across the three gaps. Leveraging the strong alignment between real and synthetic defects within the same object, we perform anomaly detection using only normal and synthetic-abnormal reference, and effectively generalize to unseen real defect types. Experiments on diverse industrial and medical benchmarks demonstrate that our method, using a simple cut-and-paste anomaly simulation strategy, achieves strong generalization across the three gaps.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Generalist anomaly detection, Industrial anomaly detection, Cross-domain generalization, Representation disentanglement

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_f9zCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Value-Order Decomposition (VOD) for generalist anomaly detection, splitting residual features into a gap-invariant value component and a gap-specific order component. Using only the value component, VODA combines few-shot normal references with synthetic anomalies for image- and pixel-level detection.

Critical strengths
  1. The problem formulation is clear and easy to follow.
  2. The proposed method is simple-yet-effective and large-scale experiments were conducted to verify its effectiveness.
  3. The proposed method does not rely on real anomalies, which are hard to collect in real world.
Critical weaknesses

Novelty. Despite the effectiveness of the proposed method VODA, the novelty is relatively limited, and I am afraid that it fails to provide a new insight for GAS. The core technique in VODA is VALUE-ORDER DECOMPOSITION. Particularly, it sorts the residual features in a descending order, and then select the largest value for anomalous score. This is similar to TopK operation. Is it possible to directly use TopK? Besides, after value-order decomposition, the order information is not used but is discarded in VODA. Why? More related works on value-order decomposition can be introduced.

Limited theoretical contributions. Proposition 2 and 3 essentially state "strict monotonic transformation preserves sorting" and "sorting remains invariant to permutation", which are mathematically correct but relatively obvious. They illustrate the properties of value and order, but cannot directly prove the core assumption that "abnormal information is mainly encoded by residual magnitude and gap information is mainly encoded by order". This hypothesis is more supported by experiments, and the theoretical evidence is still weak.

Concerns about Fig. 3c. In Fig. 3(c), the value for 1D and 2D examples are converse. Is that reasonable? Could you explain this?

Questions for authors

See weakness.

5 comments

R3W1

We thank the reviewer for this insightful comment. We emphasize that our contribution does not lie in introducing a new sorting operation. Rather, the novelty of VOD lies in explicitly decomposing residual features into value and order components and investigating their distinct roles in generalist anomaly detection. Our key insight is that these two components exhibit different behaviors under cross-gap variations: the complete value distribution provides a more stable representation, whereas the original channel ordering can capture gap-specific structural patterns that may hinder generalization. To substantiate this contribution, we further clarify the distinction between VOD and Top-k selection, explain why order is excluded from the final anomaly scoring, and discuss conceptually related rank- and ordinal-based representations.
(1) Difference between VOD and Top-k selection. VOD does not simply select the largest residual value after sorting. Specifically, we sort to the entire residual feature vector in descending order, retaining all residual magnitudes while removing their original channel ordering. The resulting full-dimensional value representations of query patches and synthetic abnormal references are then used for similarity-based abnormal reference matching. To further investigate the difference, we compare VOD with Top-k variants (k=5,10k=5,10), as reported in Table F. Top-k can be regarded as a truncated version of the value component, retaining only the largest residual responses while discarding the remaining values. Consequently, Top-k loses the relative magnitude relationships between the retained and discarded values within each patch, potentially overlooking informative distributional patterns. In contrast, VOD preserves the complete sorted residual value distribution. Moreover, Top-k is sensitive to the choice of kk. As shown in Table F, the Top-k variants consistently underperform VOD. These results demonstrate the importance of preserving the complete residual value distribution rather than relying only on a subset of high-magnitude responses.

Table F. Image-level/pixel-level AUROC under the 4-shot setting.

DatasetOursTop-5Top-10
VisA93.3/98.590.4/97.191.0/97.3
MVTecLOCO70.5/76.967.9/75.668.1/75.7
BraTS90.5/97.988.2/97.188.4/97.1

(2) Why discard order information. Our motivation for retaining value while discarding the original channel ordering is to obtain a more generalizable representation for cross-gap AD. As analyzed in Section 3.1, value preserves the complete distribution of residual magnitudes and is invariant to channel permutations (gap-invariant). In contrast, order captures the relative rankings of feature channels, which reflects gap-specific structure patterns. Although such information can be useful for distinguishing anomaly-types, it may also introduce sensitivity to structural variations across different gaps. Therefore, excluding order reduces the dependence on potentially gap-specific channel-ranking patterns, allowing the model to focus on the more stable residual magnitude distribution. Please refer to R3W2 for further discussion of the semantic correspondence between value/order and gap-invariant /gap-specific abnormal information.
(3) Related work. To the best of our knowledge, explicit value-order decomposition of feature representations has received little attention in computer vision. We therefore discuss several conceptually related studies on rank- and ordinal-based representations. For example, SIFT-Rank [1] represents local descriptors using the relative ranks of feature values, while Yagnik et al. [2] construct ordinal embeddings based on partial-order statistics. Although these methods exploit relative ordering information, they do not explicitly disentangle feature magnitudes from their original ordering. In contrast, VOD explicitly decomposes residual features into the sorted magnitude distribution (value) and the corresponding channel ordering (order), allowing us to investigate their distinct roles in generalist anomaly detection.
[1] Toews M, Wells W. Sift-rank: Ordinal description for invariant feature correspondence. CVPR 2009.
[2] Yagnik J, Strelow D, Ross D A, et al. The power of comparative reasoning. ICCV 2011.

R3W2
We thank the reviewer for this insightful comment. Specifically, we clarify the core hypothesis underlying VOD and explain how the mathematical properties established in Propositions 2 and 3 provide a theoretical basis for interpreting the value and order components in terms of gap-invariant and gap-specific abnormality-related information, respectively. First, we would like to clarify the semantic correspondence between the two components and the information they capture. Our hypothesis is not simply that abnormal information is encoded in value while gap information is encoded in order. Rather, we associate value with gap-invariant abnormality-related information and order with gap-specific abnormality-related information. Both components may contain cues for distinguishing normal and abnormal patterns; the distinction lies in whether the captured abnormality-related information remains stable across different gaps.

The mathematical properties established in Propositions 2 and 3 provide a theoretical basis for understanding the respective roles of the Value and Order components in capturing gap-invariant and gap-specific abnormality-related information, as follows:

(1) Order is associated with gap-specific abnormality-related information because it captures structural characteristics that vary across different gaps. Specifically, different object categories exhibit different shapes and textures, different defect types exhibit different geometric patterns, and different domains (e.g., industrial and medical domains) exhibit distinct structural and semantic characteristics. Such variations can affect the relative responses of feature channels and thus their rankings. This interpretation is supported by previous studies on ordinal representations and feature distribution matching, particularly EFDM, which suggest that feature ordering can preserve structural and content information. This observation is also illustrated in Fig. 3(c), motivating us to associate order with gap-specific structural and abnormality-related information.

(2) The theoretical properties, together with the above interpretation of order, provide a basis for associating value with gap-invariant abnormality-related information. Specifically, Property 1 establishes that value preserves the empirical distribution of residual magnitudes, which retains magnitude-based information relevant to anomaly identification, since abnormal patches tend to exhibit larger residual responses. Moreover, the subsequent propositions characterize the complementary behaviors of value and order: Value is invariant to permutations of the original channel ordering, whereas order preserves the corresponding relative channel rankings. More intuitively, when structural variations across different gaps induce different permutations of the same residual-value multiset, value remains unchanged, whereas order captures the corresponding structural differences. Thus, value retains magnitude-based abnormality cues that are invariant to such structural variations, while excluding the gap-specific information reflected in channel ordering.

Thus, the mathematical properties provide a theoretical basis for interpreting the invariance characteristics of the two components in terms of gap-invariant and gap-specific information. This interpretation is further supported by the t-SNE visualizations in Figs. 3(a) and 6, as well as the probing experiments in Table 5. We will revise the theoretical discussion to make the connection between these mathematical properties and their implications for gap-invariant and gap-specific representations more explicit.

R3W3
We thank the reviewer for pointing this out. The inconsistency in Fig. 3(c) was caused by a coding error in generating the figure. We have corrected the error and updated the figure to ensure that, for both the 1D and 2D examples, the value-based and order-based representations are correctly obtained through VOD with descending sorting. We apologize for this oversight.

Thank you for your response and the additional experiments. The response has addressed some of my concerns about Fig. 3(c) and the theoretical assumption.

However, I still have some questions. 1) As I mentioned before, the VOD method is similar to top-K operation. While the authors have provided explanations of the differences between the two and provided related experiments to support the effectiveness of the proposed method, I do not think that comparative experiments is convincing. In the proposed method, the authors use top 1% highest values, which generates over 2000 values from S (448 $\times $ 448). However, the authors only adopt top-5 and top-10 to compare with VOD. 2) Although the theoretical assumption is relatively convincing, the Proposition is still simple and obvious, which weakens the theoretical contribution.

Reviewer_rxftOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Author proposes VOD, which sorts the channel values of each patch-level absolute residual relative to its nearest normal reference.

Critical strengths
  1. VOD is interesting to reduce variation in patch-level anomaly residuals
  2. The hypothesis which author proposed can be evaluated through controlled experiments
Critical weaknesses
  1. Compare to recent work like AnomalyDINO, SubspaceAD, the model does not show advantages using more training steps.

  2. This work combines several existing approaches : nearest-normal residuals and reference-based scoring are closely related to prior residual and DINO nearest-neighbor methods; synthetic defect generation follows DRAEM; and the separation of feature values and order is explicitly inspired by EFDM. The real difference lies in applying channel-wise sorting to patch residuals and using the sorted residuals to score synthetic abnormal references. Could authors provide more analysis or ablation towards this mechanisms?

Questions for authors
  1. Inconsistent results: the VODA image AUROC for two-shot MVTec3D differs between Tables 1 and 12 (84.2 versus 82.4), and the eight-shot MVTecLOCO result differs between Tables 1 and 14 (71.7 versus 71.1).

  2. Lack of clarity. Table 4 full-model results appear to correspond largely to the four-shot results in Table 1, yet the MVTec AD image AUROC is 97.8 in Table 4 and 96.8 in Tables 1, 3, 10, and 13.

Please specify the setting, seeds, and aggregation procedure for each table, and identify which values should be used for the main comparison.

2 comments

R2W1
We note that the reviewer's original comparison involves different backbone architectures and input resolutions: SubspaceAD uses DINOv2-G, whereas our method uses DINOv2-S; AnomalyDINO uses an input resolution of 672, whereas ours uses 448. These differences may affect the comparison. To provide a more controlled evaluation, we additionally compare these methods under matched backbone and resolution settings (DINOv2-S and an input resolution of 448). As shown in Tables E1–E3, our method outperforms AnomalyDINO on most datasets and SubspaceAD on all datasets under the 2-, 4-, and 8-shot settings, demonstrating its effectiveness under comparable experimental conditions..

Table E1. Image-level/pixel-level AUROC under the 2-shot setting.

VisAMVTecADBTADMVTec3DMVTecLOCOBraTS
our92.4/98.495.7/97.495.4/96.084.2/98.066.7/75.588.5/97.6
SubspaceAD89.1/96.494.9/96.992.0/96.676.8/97.255.5/70.078.9/95.3
AnomalyDINO88.3/97.896.7/96.592.5/95.676.3/96.662.7/75.383.6/96.9

Table E2. Image-level/pixel-level AUROC under the 4-shot setting.

VisAMVTecADBTADMVTec3DMVTecLOCOBraTS
our93.3/98.596.8/97.595.0/96.885.0/98.770.5/76.990.5/97.9
SubspaceAD90.1/96.496.0/96.991.3/96.679.8/97.558.5/65.582.5/96.6
AnomalyDINO91.3/98.097.6/96.792.6/96.679.1/97.270.5/76.587.3/97.2

Table E3. Image-level/pixel-level AUROC under the 8-shot setting.

VisAMVTecADBTADMVTec3DMVTecLOCOBraTS
our93.0/98.697.8/97.694.6/97.089.0/98.671.7/76.790.1/98.1
SubspaceAD90.7/96.495.8/96.991.5/96.879.5/97.459.4/65.081.9/96.7
AnomalyDINO92.6/98.298.0/97.093.9/96.779.8/97.573.6/77.389.3/97.6

R2W2
We do not simply combine these existing approaches. Instead, our method provides an in-depth investigation of residual features, with two key contributions: (1) VOD, which explicitly decomposes residual features into complementary Value and Order representations and investigates their distinct properties under cross-gap variations; and (2) the A-Value projection module, which leverages the Value representation for similarity-based matching with synthetic abnormal references.
In the original manuscript, we conducted component-wise ablations to evaluate the roles of adaptation, the value representation in VOD, the normal reference branch SnS^n, and the synthetic abnormal reference branch SaS^a. Specifically, comparing the full model with its variant using the original residual representations instead of the value representations evaluates the benefit of VOD in synthetic abnormal reference matching. Removing SaS^a further assesses the contribution of synthetic abnormal reference matching itself. Please refer to Table B in R1W2 for a more straightforward interpretation.

R2Qs
We thank the reviewer for pointing out these inconsistencies. After carefully rechecking the experimental results, we identified these discrepancies as reporting errors rather than differences in experimental settings.Specifically, we have corrected the following values:
Table 12: The image AUROC for 2-shot MVTec3D is corrected from 82.4 to 84.2.
Table 14: The image AUROC for 8-shot MVTecLOCO is corrected from 71.1 to 71.7.
Table 4: The image AUROC for the full model on MVTec AD is corrected from 97.8 to 96.8.
We have also systematically cross-checked all tables against the original experimental records. These corrections do not affect the main conclusions of our work.
The experimental protocol is described in Appendix C. We provide the following additional details to clarify the settings, random seeds, and result aggregation procedure.

  1. Experimental Protocol and Reproducibility. All experiments are conducted using three random seeds: [0,1,2][0, 1, 2]. These seeds control the randomness in few-shot reference sampling, episodic sampling, the initialization of trainable modules, and synthetic anomaly generation. For each KK-shot setting, KK normal reference samples are randomly selected and the sampling is performed independently for each run.
  2. Model Selection and Evaluation. Following NAGL, we construct a validation set from VisA and select the best checkpoint based on the sum of image-level and pixel-level AUROC on this validation set. For cross-dataset evaluation, the model is trained on MVTec AD and evaluated on the other datasets. For evaluation on MVTec AD, the model is instead trained on VisA. The target test sets are used exclusively for final evaluation and are not involved in model training or checkpoint selection. Each target dataset is evaluated under three few-shot settings K∈2,4,8K \in \\{2,4,8\\}.
  3. Result Aggregation. All reported results are the means over three independent runs with different random seeds. Specifically, each evaluation metric is calculated separately for each run, and the resulting values are then averaged across the three runs. The same aggregation procedure is used for both the main comparisons and ablation studies.
  4. Table-specific Experimental Settings. The common experimental protocol described above applies to all tables unless otherwise specified. Any deviations from the common protocol are explicitly stated in the corresponding captions.
Reviewer_nGPiOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Value-Order Decomposition (VOD) for few-shot generalist anomaly detection. It decomposes residual features into a value component -- residual magnitude distribution, and an order component--ranking information. The core idea is to build a permutation-invariant feature space so that cut-paste synthetic anomalies can serve as effective proxies for real anomalies. The work is well-motivated with extensive benchmarks, but it has notable conceptual and experimental limitations that need careful revision.

Critical strengths
  1. The paper identifies a key pain point of residual-based GAD methods: residual signals are entangled with “three generalization gaps”: object categories, defect types, and domains. The proposed value-order decomposition explicitly separates magnitude distribution from channel ranking information, and leverages the permutation invariance of the sorted value vector.
  2. Well-structured framework and favorable practical properties.The VOD module is lightweight and plug-and-play, and the two-score fusion design is reasonable and brings clear complementary effects.
  3. The paper is clear and easy to follow.
Critical weaknesses
  1. The order component is defined and highlighted in the decomposition formulation, yet it is never used in training, supervision, or inference scoring and merely serves as a mathematical term to characterize the discarded permutation-sensitive information. This raises a critical concern: whether “value-order decomposition” is an appropriate name for the proposed method, or whether the pipeline is essentially residual sorting followed by magnitude-distribution matching. More importantly, since the method discards all order-related channel ranking information, a natural question arises: could discarding order information introduce negative effects for certain anomaly types? For example, anomalies in MVTec LOCO are logical/structural anomalies where the abnormal signature may reside in structural arrangement and relational patterns rather than residual magnitudes. The authors are expected to analyze and justify whether discarding order impairs the model’s capacity for such cases.
  2. The ablation shows that S^n brings limited gain on MVTec AD/VisA but nearly 4-point drop when removed on BTAD. The paper does not provide targeted analysis linking this phenomenon to dataset characteristics. Without such explanation, readers cannot understand the boundary of the two-score fusion strategy. For example, is that the performance improvement is built upon adding synthetic anomaly references branch (S^a)?. While the ablation “w/o VOD” is provided, further evidence is required to verify: whether simply feeding raw residuals together with the same synthetic anomaly reference can achieve comparable performance. The current experiments cannot fully rule out the alternative hypothesis that the main gain comes from introducing anomaly references, while VOD only plays a secondary supporting role.
  3. Sorting appears useful, but the domain-invariance explanation is less convincing. Appendix A proves invariance properties that do not establish that domain differences are mainly carried by channel order. The probes and gap experiments provide some support. However, Equation (6) is the query residual norm multiplied by its maximum cosine similarity to a synthetic direction. A norm-only baseline and a unit-normalized-query variant, with matched training and score scaling, would help show how much the synthetic directions add beyond residual magnitude.
  4. The BraTS ablation deserves more attention: removing adaptation raises image AUROC from 90.5 to 97.2. The authors already attribute this to industrial-domain bias. Given the size of the reversal, the discussion should be clearer about when source adaptation helps and when the frozen representation may be preferable.
Questions for authors

Since the method discards all order-related channel ranking information, a natural question arises: could discarding order information introduce negative effects for certain anomaly types? Relevant analysis is needed to address this concern. With synthetic anomaly samples incorporated into the pipeline, there is a risk that the model could exhibit a bias toward such anomalies. In-depth analysis backed by quantitative experiments is highly desirable to address this concern.

5 comments

R1W1

We clarify that VOD is not merely residual sorting followed by magnitude-distribution matching, and that its name reflects the explicit decomposition and analysis of residual information.

(1) Role of the Order component. Order is not considered meaningless. As illustrated in Figs. 3(a) and 6, it can distinguish normal from abnormal patterns, while also capturing channel-wise ranking patterns associated with object categories, defect types, and domain characteristics. However, its sensitivity to structural variations across the three generalization gaps may introduce gap-specific biases and hinder cross-gap generalization. In contrast, Value preserves the empirical distribution of residual magnitudes and is permutation invariant, making it a more stable basis for cross-gap anomaly detection.

(2) VOD explicitly decomposes residual representations into Value and Order to investigate their respective properties and reliability under different generalization gaps. Although only Value is retained for anomaly scoring, analyzing Order is essential for identifying permutation-sensitive, gap-specific information and understanding its impact on cross-gap generalization. Thus, VOD refers to the decomposition and analysis of residual information, rather than requiring both components to contribute to the final scoring. Furthermore, the gap-specific characteristics revealed by Order may provide useful cues for the semantic interpretation of defect types, which we leave for future work.

(3) Impact on logical anomaly detection. To investigate this concern, we conduct additional experiments on the two types of anomalies in MVTecLOCO: visual anomalies (termed structural anomalies in the dataset) and logical anomalies (Table A). We train separate models using Value, Order, or the original residual features, with breakfast_box, juice_bottle, and pushpins for training and the remaining categories for testing.

For visual anomalies, Value consistently outperforms Order and the original residual features across all metrics. For logical anomalies, Order achieves slightly better image-level AUROC, AP and F1, whereas Value performs better in pixel-level PRO and AUROC, suggesting that Order provides some complementary image-level cues but no consistent gains. We attribute this limitation primarily to the patch-wise residual formulation. Logical anomalies often arise from spatial relationships among otherwise normal-looking components. Individual abnormal patches may still find visually similar normal references and consequently produce small residuals. Since Order only captures channel-wise rankings within each patch residual, it cannot explicitly model spatial relationships across patches. Therefore, using Order does not fundamentally resolve this limitation.

Table A. Results on the two types of anomalies in MVTecLOCO under the 4-shot setting.

Visual anomalieslogical anomalies
imgpiximgpix
AUROCAPF1PROAUROCF1AUROCAPF1PROAUROCF1
value76.673.368.887.896.538.057.157.568.443.073.623.2
order60.162.667.934.070.531.558.764.269.430.272.823.5
residual(value+order)75.972.167.783.194.036.658.958.068.342.372.122.8

R1W2

Regarding the role of SnS^n. We attribute the more significant contribution of SnS^n on BTAD to the greater diversity of normal appearances and structures. Under few-shot settings, the limited normal samples may not adequately cover these variations, making some unseen but normal patterns difficult to distinguish from anomalies using SaS^a alone. In this case, SnS^n provides complementary comparisons in the original feature space, helping improve the separation between normal and abnormal patterns. This suggests that SnS^n is more beneficial when normal patterns are diverse and difficult to cover with few-shot samples, while its contribution is relatively limited when normal patterns are more consistent.

Regarding the contribution of VOD. To further clarify the role of VOD, we reorganize the existing ablation results in Table 4 of the main manuscript into Table B, presenting the configurations from a progressive module-integration perspective. Note that Table B does not introduce additional experiments; its configurations correspond to relevant rows in Table 4, with only their descriptions adjusted to make the comparisons requested by the reviewer more explicit. Specifically, we consider three configurations: (a) raw residuals with SnS^n as the baseline; (b) raw residuals with both normal and synthetic anomaly references; and (c) our VODA. The results show that introducing synthetic anomaly references alone (b) actually degrades performance compared with the baseline (a). In contrast, incorporating VOD (c) leads to substantial improvements, indicating that VOD is crucial for effectively exploiting synthetic anomaly references and aligning abnormal representations across gaps.

Table B. Image-level/pixel-level AUROC under the 4-shot setting.

adaptationSnS^nSaS^aVODVisAMVTecADBTADMVTec3DMVTecLOCOBraTS
a✓\checkmark✓\checkmark×\times×\times90.5/97.696.7/97.494.6/96.683.5/97.869.6/76.885.8/97.7
b✓\checkmark✓\checkmark✓\checkmark×\times90.1/97.294.7/96.893.4/93.481.2/97.669.4/77.181.9/96.2
c✓\checkmark✓\checkmark✓\checkmark✓\checkmark93.3/98.596.8/97.595.0/96.885.0/98.770.5/76.990.5/97.9

R1W3

We would like to clarify that the cross-dataset domain variations considered in our work are not limited to conventional low-level appearance shifts, such as changes in illumination or visual style. Rather, they involve substantial differences in visual content and semantic characteristics, including object categories, materials, and even distinct domains such as industrial inspection and medical imaging. Such content-level variations can lead to changes in the channel-wise response patterns of residual features, potentially including their relative rankings.

Our motivation is to investigate whether retaining the distribution of residual magnitudes while excluding permutation-sensitive channel-order information can yield more stable representations for cross-dataset anomaly detection. We have revised the manuscript to provide a clearer rationale linking the mathematical properties of the Value–Order decomposition to our interpretation of the respective roles of these two components under cross-dataset variations. In particular, we clarify how domain variations alter channel order. Please refer to our response to R3W2 for details.

To examine the respective contributions of residual magnitude and the directional information provided by synthetic anomaly references, we conduct additional experiments with three scoring variants: (1) norm-only scoring, (2) direction-only scoring using L2-normalized query residuals, and (3) our method with the original joint magnitude-direction scoring. All variants use the same trained model and reference settings, with only the scoring function changed. As shown in Table C, our joint scoring strategy consistently outperforms both norm-only and direction-only scoring. Meanwhile, norm-only scoring performs better than direction-only scoring, indicating that residual magnitude provides a stronger standalone discriminative signal. Nevertheless, the improvement of our joint scoring over norm-only scoring demonstrates that the similarity between query residuals and synthetic anomaly directions provides additional complementary information beyond residual magnitude alone.

Table C. Image-level/pixel-level AUROC under the 4-shot setting.

VisAMVTecADBTADMVTec3DMVTecLOCOBraTS
Norm-only92.6/97.796.0/96.194.6/96.784.3/97.668.9/75.588.6/96.1
direction-only89.4/96.995.7/95.992.8/94.082.0/96.267.1/74.284.7/94.4
our93.3/98.596.8/97.595.0/96.885.0/98.770.5/76.990.5/97.9

Comments on the paper

2 comments

We sincerely thank all reviewers for their constructive comments and valuable time. Below, we provide point-by-point responses to all comments. Throughout this rebuttal, we use a consistent notation to identify individual comments. For example, "R1-W1/Q1" refers to the first weakness/question raised by Reviewer 1.

Summary of Changes in the Revised PDF

1. At the end of Section 2.2 (Lines 176–193), clarify the distinction between our method and EFDM, and provide additional discussion of related rank-based methods.

2. Introduce a new section (Section 3.2) to explain the semantic correspondence between the Value/Order components and gap-invariant/gap-specific abnormality-related information.

3. In Section 4.4, reorganize the ablation results in Table 4 and revise the accompanying analysis to facilitate a clearer interpretation.

4. In Appendix Section C, provide additional experimental details (Lines 956–971).

5. Add a new section (Section E.8) to discuss the comparison with Top-k statistics.

6. Add a new section (Section E.9) to discuss the contributions of residual magnitude and the directional information provided by synthetic anomaly references.

7. Add a new section (Section E.10) to investigate the impact of Order on logical anomaly detection.

8. Update Figure 3(c) and clarify the experimental settings in the table captions.

9. Revise the propositions to strengthen the theoretical contributions of the paper.

Area_Chair_AycLOpenReview ↗

Please engage in the discussion

Dear Reviewers,

Thank you for your reviews and contributions so far. The authors have now posted their response, and I would like to kindly remind you to review it and engage in the discussion where appropriate.

Please also consider whether the authors' clarifications address any of your concerns and update your assessment if warranted.

Thank you for your time and effort.

Best regards, Your AC.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.