Skip to content

Robust Zero-shot Anomaly Detection under Limited Auxiliary Anomaly Priors

Conference: ECCV 2026
arXiv: 2606.29428
Code: https://github.com/ZhouF-ECNU/DIVE
Area: Anomaly Detection
Keywords: Zero-Shot Anomaly Detection, CLIP, Vision-Language Models, Prompt Learning, Decoupled Representations

TL;DR

DIVE is the first to investigate the scenario of zero-shot anomaly detection with limited anomaly priors in auxiliary data. By employing a shallow-to-deep text embedding injection strategy, it enables the vision encoder to abstract cross-domain general anomaly concepts from limited auxiliary anomaly patterns. It also introduces a decoupling mechanism to eliminate the interference of object semantics on vision embeddings. When using the texture dataset DTD as auxiliary data, DIVE outperforms SOTA baselines on 12 target datasets by up to 28.5% (AP) in classification and 47.0% (AUPRO) in segmentation, while compressing the performance degradation caused by insufficient auxiliary data diversity by approximately 46.5%-77.3%.

Background & Motivation

Background: Zero-shot anomaly detection (ZSAD) aims to identify defects in arbitrary unseen target domains without needing target domain data collection. Existing methods are mostly based on large vision-language models like CLIP, aligning visual features with textual descriptions of anomalies through manual prompt engineering or learnable prompt tuning. ZSAD methods have continuously evolved, from hand-crafted prompts in WinCLIP, to object-agnostic learnable prompts in AnomalyCLIP, and to fine-grained prompt designs like TPS and FAPrompt.

Limitations of Prior Work: All existing ZSAD methods build upon the optimistic assumption that auxiliary training data contains rich and diverse anomaly patterns. When auxiliary data lacks anomaly diversity (e.g., training only on the texture anomaly dataset DTD and generalizing to industrial defects and medical lesions), these methods severely overfit to the specific anomaly distribution of the auxiliary data. Consequently, they fail almost completely on complex defects and emerging object classes in the target domain. As shown in Fig. 1 of the paper, when original auxiliary training data MVTec (with various industrial defects like holes, stains, cracks) is replaced with DTD (only texture anomalies), the average AP of five SOTA baselines drops by 9.9% to 23.3%.

Key Challenge: Real-world target domains exhibit unpredictable anomaly variations and continuously emerging new anomaly patterns, whereas auxiliary data can never comprehensively cover all possible anomaly categories. This key challenge of "limited anomaly priors in auxiliary data" universally exists in reality but has never been systematically investigated before.

Another Limitation: The feature space of pre-trained vision encoders is dominated by object semantics. When aligning such visual embeddings with object-agnostic text prompts designed solely for anomaly detection, the dominant object semantics act as interfering noise. The model fails to separate subtle defect variations from salient object identity, severely hindering the learning of discriminative anomaly representations.

Core Idea: Rather than memorizing specific anomaly appearances in the auxiliary data, the model should learn the general concept of "what constitutes an anomaly." This is achieved by injecting general knowledge of anomalies/normality described in text during the visual encoding process (shallow-level injection for cross-class generalization, deep-level injection for abstracting anomaly concepts), while decoupling global visual embeddings into independent state and semantic subspaces. Consequently, the model generalizes robustly to unseen target domains even with limited auxiliary anomaly priors.

Method

Overall Architecture

The core problem DIVE addresses is: how to enable the CLIP model to learn to distinguish anomalies without relying on specific anomaly appearances when the auxiliary training data has limited anomaly patterns. The general approach is to "inject general anomaly knowledge from text during visual feature extraction, and then separate the entangled object semantics and anomaly states." Instead of modifying the CLIP architecture, it attaches learnable injection modules and decoupling branches to a frozen backbone.

The DIVE framework consists of three parallel text prompt learning pathways and a two-stage visual embedding process. The three text pathways are: a state-aware prompt learner (generating "normal"/"abnormal" text embeddings \(z_n, z_a\)), a semantic-aware prompt learner (generating object class text embeddings \(z_c\)), and description embeddings \(Z_g\) obtained by passing LLM-generated anomaly/normal description dictionaries through a frozen CLIP text encoder. On the visual side, with ViT as the backbone, learnable tokens of semantic prompts are projected as visual prompts and injected into the first ViT layer via a V-L coupling function \(\mathcal{F}\) in the shallow layers. In the deep layers (the last 4 ViT blocks), \(Z_g\) is fused with the patch tokens via cross-attention. After the ViT outputs the global feature \(z_x\) and local patch features \(z_e\), \(z_x\) is decoupled into state embeddings \(z_{st}\) and semantic embeddings \(z_{se}\) through two parallel residual MLP branches (\(\mathcal{G}_{state}\) and \(\mathcal{G}_{semantic}\)). These are respectively aligned with state and semantic text embeddings to obtain the image-level anomaly score \(s\). Meanwhile, \(z_e\) is used to calculate the similarity map with state text embeddings, yielding the pixel-level anomaly map \(\mathbf{M}\) after upsampling and Gaussian smoothing.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input Image x"] --> B["Shallow Text Injection<br/>Semantic prompts projected as visual prompts"]
    B --> C["ViT Vision Encoder<br/>（Frozen Backbone）"]
    D["LLM-generated abnormal/normal<br/>text description embeddings Z_g"] -->|"Deep Cross-Attention<br/>（Last 4 ViT blocks）"| C
    C --> E["Global Feature z_x"]
    C --> F["Local patch feature z_e"]
    E --> G["Decoupling branches<br/>G_state / G_semantic<br/>+ Orthogonal constraint"]
    G --> H["z_st aligned with z_n / z_a<br/>→ Image-level anomaly score s"]
    F --> I["z_e similarity → Upsampling<br/>→ Anomaly map M"]

Key Designs

1. Shallow and Deep Text Embedding Injection: Abstracting General Anomaly Concepts from Limited Anomaly Samples

This design directly addresses Challenge 1—model overfitting when auxiliary data lacks anomaly diversity. DIVE's solution is split into two levels: shallow-level injection handles "unseen target object categories," and deep-level injection handles "unseen novel anomaly patterns."

Shallow-level injection adopts the multimodal prompting mechanism of MaPLe: projecting the first \(h=4\) learnable tokens of the semantic text prompt into the visual space through a V-L coupling function \(\mathcal{F}\) (a learnable linear projection) to serve as additional visual prompt tokens \(\tilde{C}\) concatenated into the ViT input sequence: \([c_1, E_1, \tilde{C}_1] = \text{ViT}_1([c_0, E_0, \mathcal{F}(C_0|_1^h)])\). The intuition here is that the shallow ViT layers are mainly responsible for capturing coarse-grained local structures like edges and textures. Projecting semantic prompts into these layers ensures that the visual encoding process retains cross-modal alignment capability—even if target domains present object categories never seen in the auxiliary data, the model does not go completely "blind."

Deep-level injection is the more critical design: first, GPT-4o generates 100 anomaly descriptions and 35 normal descriptions (e.g., "a photo of an object with a crack", "a photo of an object with a stain", "a photo of an object with a flawless surface"). These descriptions are input into the frozen CLIP text encoder to obtain \(Z_g\), which is then projected into the visual space via a learnable projection matrix \(W_{proj}^\top\). In the last 4 layers of the ViT (layers 22-24, i.e., after \(Q'=21\)), each patch token is used as a query, and the projected text description embeddings act as keys and values to fuse via cross-attention \(\mathcal{CA}\): \(E_{d'} = E_{d'-1} + \gamma \cdot \mathcal{CA}_{d'-1}(E_{d'-1}, Z_g W_{proj}^\top, Z_g W_{proj}^\top)\), where \(\gamma\) is a hyperparameter balancing the original visual features and the injected textual information. Why place this in the deep layers? Because the deep layers of ViT possess strong concept abstraction capabilities—at this stage, the patch tokens have undergone 21 layers of hierarchical abstraction and are well-suited to receive language-level general anomaly concepts. The effect of deep-level injection is equivalent to letting the model map specific anomaly textures seen in the auxiliary data (such as fabric defects in DTD) to general descriptions ("a photo of an object with a stain"). Thus, when encountering target domain anomalies with similar patterns but entirely different surface materials (such as tumor shadows in medical imaging), the same anomaly concept can still be activated.

Ablation studies show that removing deep-level injection (DIVE\(_{\text{-text(d)}}\)) causes a more severe performance drop than removing shallow-level injection (DIVE\(_{\text{-text(s)}}\)) (MVTec AUROC drops by 2.0 vs 0.5). Removing all text injections (DIVE\(_{\text{-text}}\)) leads to the largest drop (AUROC drops by 2.4, AP by 4.4). This confirms that text injection—especially deep-level LLM description injection—is the core pillar of DIVE's generalization capability when auxiliary anomaly priors are limited.

2. Visual Embedding Decoupling Mechanism: Eliminating Object Semantics Interference on Anomaly Discrimination

This design addresses Challenge 2—the entanglement of object semantics and anomaly states in visual embeddings, making direct alignment with object-agnostic text prompts suboptimal. DIVE's solution is to explicitly project the global visual feature \(z_x\) into two orthogonal subspaces through two learnable residual MLPs.

Specifically: \(z_{st} = z_x + \mathcal{G}_{state}(z_x)\) and \(z_{se} = z_x + \mathcal{G}_{semantic}(z_x)\), where \(\mathcal{G}_{state}\) and \(\mathcal{G}_{semantic}\) are two independent MLP networks that learn the residual offsets \(\Delta z_{st}\) and \(\Delta z_{se}\). Residual learning is adopted rather than absolute transformation for two benefits: alleviating catastrophic forgetting of pre-trained visual concepts and stabilizing the optimization process. To ensure the two subspaces are truly independent, an orthogonal constraint is applied—the decoupling loss is defined as the squared cosine similarity between the two residual offset vectors: \(\mathcal{L}_{dis} = (\frac{\Delta z_{st}^\top \Delta z_{se}}{\|\Delta z_{st}\|_2 \|\Delta z_{se}\|_2})^2\).

After decoupling, the state branch \(z_{st}\) is aligned only with "normal/abnormal" text embeddings (via cross-entropy loss \(\mathcal{L}_{state}\)), and the semantic branch \(z_{se}\) is optimized via contrastive learning only with object class text embeddings (\(\mathcal{L}_{sem}\) to pull it closer to the ground-truth class embedding and push it away from other classes). Only the state branch is used to compute anomaly scores during inference, since the category of the object and whether it has a defect are two independent matters.

In the ablation study, removing the decoupling mechanism (DIVE\(_{\text{-dis}}\)) primarily leads to a decrease in classification metrics (MVTec AUROC drops by 0.6, AP drops by 1.8). This is because decoupling acts on the global visual embedding, from which the classification score is derived by aligning with the state text embedding. In contrast, the drop in segmentation metrics is smaller, as segmentation uses the un-decoupled local patch features \(z_e\).

3. RCPRO Metric: Correcting AUPRO's Leniency on False Positives

This is not a part of the DIVE model itself, but a new segmentation evaluation metric proposed in the paper. The standard AUPRO metric focuses solely on whether the prediction covers ground-truth anomaly regions (per-region overlap), completely ignoring false positive predictions in normal areas and over-segmentation of anomaly regions. As shown in the right image of Fig. 1, AnomalyCLIP predicts three anomaly regions on a colonoscopy image, two of which are obvious false positives (including the one with the highest score), but AUPRO still yields an inflated 96.5% because those three regions indeed cover the ground truth. RCPRO adds a region-calibrated penalty to the PRO curve: for each predicted connected anomaly region \(P_m\), if its maximum overlap with any ground-truth region is less than the threshold \(\tau = 0.05 \times |T_{n^*}|\), its quality score is set directly to 0 (punishing false positives in normal areas); if it passes the filtering but is excessively expanded (the area ratio exceeding the ground-truth boundary \(E(P_m) > \theta_{exc} = 2.0\)), its quality score is inversely scaled by \(\theta_{exc} / E(P_m)\) (punishing over-segmentation). The final RCPRO is the Area Under the Curve formed by the PRO scores (X-axis) and region calibration scores (Y-axis) at different thresholds. Under RCPRO, the score for the aforementioned colonoscopy case is adjusted to a reasonable 28.7%.

Loss & Training

The total loss function of DIVE consists of five components: \(\mathcal{L} = \mathcal{L}_{state} + \lambda \mathcal{L}_{sem} + \lambda' \mathcal{L}_{dis} + \mathcal{L}_{seg} + \lambda'' \mathcal{L}_{reg}\).

  • \(\mathcal{L}_{state}\): Classification loss of the state branch, computed as cross-entropy over \(p_{st} = \text{softmax}([z_{st}^\top z_n, z_{st}^\top z_a] / \tau)\) to enable the model to distinguish between normal and abnormal images.
  • \(\mathcal{L}_{sem}\): Contrastive loss of the semantic branch, pulling \(z_{se}\) closer to the text embedding of the actual category and pushing it away from others, ensuring the semantic branch captures object identity rather than anomaly state.
  • \(\mathcal{L}_{dis}\): Decoupling orthogonal loss, which constrains \(\Delta z_{st}\) and \(\Delta z_{se}\) to be orthogonal, forcing independence of the two subspaces.
  • \(\mathcal{L}_{seg}\): Segmentation loss = Focal Loss + Dice Loss (anomaly channel against ground-truth mask + normal channel against inverse mask) to optimize pixel-level anomaly localization.
  • \(\mathcal{L}_{reg}\): Regularization term for visual prompts \(\frac{1}{|\tilde{C}|}\|\tilde{C}\|_2^2\), preventing the magnitude of injected visual prompt tokens from becoming too large and destabilizing training.

During training, the CLIP ViT and text encoder backbones are frozen, and only three sets of learnable parameters are trained: prompt learners (state + semantic, introducing \(h=4\) learnable tokens per layer with prompt depth \(P=9\)), the V-L coupling function \(\mathcal{F}\) with the cross-attention layer \(\mathcal{CA}\), and the two decoupling MLPs (\(\mathcal{G}_{state}\), \(\mathcal{G}_{semantic}\)). During inference, frozen model weights are directly used: the image-level anomaly score is calculated as \(s = \frac{\exp(z_{st}^\top z_a / \tau)}{\exp(z_{st}^\top z_n / \tau) + \exp(z_{st}^\top z_a / \tau)}\), and the pixel-level anomaly map is \(\mathbf{M} = G_\sigma(Up(\mathbf{S}_a))\), where \(\mathbf{S}_a\) is the anomaly channel of the similarity map and \(G_\sigma\) is a Gaussian filter with standard deviation \(\sigma\).

Key Experimental Results

Main Results

Table 1: Performance comparison between DIVE and various baselines when DTD is used as auxiliary data (average results across 12 datasets). DTD only contains texture anomalies, simulating the real-world scenario of "limited auxiliary anomaly priors."

Task Metric AnomalyCLIP AdaCLIP AF-CLIP AA-CLIP TPS DIVE Gain (vs best)
Classification (Avg of 8) AUROC 79.3 80.1 75.8 71.1 81.6 87.3 +5.7
Classification (Avg of 8) AP 51.7 49.6 43.9 40.8 60.5 69.3 +8.8
Segmentation (Avg of 8) AUROC 72.9 84.1 82.3 84.4 63.9 87.3 +2.9
Segmentation (Avg of 8) AUPRO 32.5 24.1 65.5 64.9 24.6 71.1 +5.6
Segmentation (Avg of 8) RCPRO 16.3 28.1 29.4 34.3 17.2 40.4 +6.1

Table 2: Average performance comparison between DIVE and baselines across 8 classification and 8 segmentation datasets when MVTec is used as auxiliary data (sufficient anomaly diversity).

Task Metric AnomalyCLIP AdaCLIP AF-CLIP AA-CLIP TPS DIVE
Classification (Avg of 8) AUROC 88.6 89.1 87.4 78.6 88.7 90.1
Classification (Avg of 8) AP 68.4 70.0 67.2 52.2 70.3 74.6
Segmentation (Avg of 8) AUROC 90.0 88.3 91.1 83.3 73.1 91.1
Segmentation (Avg of 8) AUPRO 71.7 31.0 71.5 65.5 40.9 75.0
Segmentation (Avg of 8) RCPRO 41.4 34.6 42.7 34.3 19.6 43.7

Two key comparisons: (1) When switching the auxiliary data from MVTec to DTD, the average AP of the baselines drops by 9.9%-23.3%, while DIVE drops by only 5.3%, compressing the performance degradation rate by about 46.5%-77.3%. (2) Even when auxiliary data is sufficiently diverse (MVTec), DIVE remains the best or tied for the best across all metrics, indicating its design does not compromise performance in scenarios with abundant data.

Ablation Study

Table 3: Ablation of DIVE components (DTD as auxiliary data, evaluated on MVTec and Visa).

Configuration MVTec Class AUROC MVTec Seg AUPRO Visa Class AUROC Visa Seg AUPRO
DIVE (Full) 89.3 80.9 77.8 80.2
w/o Decoupling (DIVE\(_{\text{-dis}}\)) 88.7 (-0.6) 80.5 (-0.4) 76.2 (-1.6) 79.5 (-0.7)
w/o Shallow Inj. (DIVE\(_{\text{-text(s)}}\)) 88.8 (-0.5) 80.5 (-0.4) 76.9 (-0.9) 79.4 (-0.8)
w/o Deep Inj. (DIVE\(_{\text{-text(d)}}\)) 87.3 (-2.0) 77.3 (-3.6) 75.3 (-2.5) 77.9 (-2.3)
w/o All Text Inj. (DIVE\(_{\text{-text}}\)) 86.9 (-2.4) 76.6 (-4.3) 74.4 (-3.4) 74.4 (-5.8)

Key Findings

  • Deep text injection is the most critical component: The performance drop when removing deep-level injection (AUROC -2.0, AUPRO -3.6) is far more severe than removing shallow-level injection (AUROC -0.5, AUPRO -0.4). This illustrates that injecting LLM-generated general anomaly/normal descriptions via cross-attention with patch tokens in the deep layers of ViT is the core source of DIVE's generalization capability. Completely removing text injection causes performance to collapse (AP drops by 4.4-6.1), confirming that pure visual features cannot generalize when auxiliary anomaly priors are limited.
  • Decoupling mechanism primarily affects classification rather than segmentation: Removing decoupling leads to a clear drop in classification metrics (MVTec AP -1.8, Visa AUROC -1.6), but segmentation metrics remain almost unaffected (AUPRO -0.4/-0.7). This is because decoupling only operates on the global visual embedding \(z_x\), whereas segmentation uses the un-decoupled local patch features \(z_e\). This also suggests different feature space requirements for classification and segmentation—classification requires eliminating object semantic interference, while segmentation relies more on local details.
  • An observable side effect of DIVE: The cross-attention module occasionally exhibits a slight tendency to focus on object boundaries (e.g., the contours of hazelnuts, the edges of cables), resulting in sporadic false positives along object edges in the anomaly map. The paper provides a visual analysis of this in Fig. 3 and Fig. 4—high attention scores are assigned to both defective areas and contours, which is a boundary effect of the text description's "object-agnostic anomaly concept."
  • Inappropriate inflation of AUPRO confirmed by RCPRO: AA-CLIP achieves high AUPRO on multiple datasets (since its predictions cover ground-truth regions) but low RCPRO (since its predictions over-expand, introducing many false positives). Conversely, AdaCLIP achieves moderate AUPRO but higher RCPRO (locating core anomaly areas accurately, despite incomplete boundaries). This aligns perfectly with the motivation for RCPRO—evaluation should not solely measure ground-truth coverage, but also penalize false positives.

Highlights & Insights

  • "Teaching vision models to identify anomalies via language descriptions" is clever and transferable: Instead of exposing the model to more anomaly images, it uses 135 general anomaly/normality descriptions from an LLM, injected into the deep ViT layers via cross-attention. This essentially uses textual knowledge to compensate for visual data scarcity. This paradigm of "vision-centric, language-assisted, and multimodal-complemented data shortcomings" can be transferred to other visual data-scarce scenarios (e.g., few-shot medical image classification, rare event detection) as long as a language anchor describing "what constitutes an anomaly" can be defined.
  • Decoupling with orthogonal constraint on residuals is simple and highly effective: Two lightweight MLPs learn residual offsets (instead of absolute mapping), and a squared cosine similarity orthogonal loss separates the anomaly state and object semantics in the visual embedding. This three-formula approach avoids complex adversarial training or decoupling VAEs, making it suitable as a general decoupling module for other multi-task vision models.
  • RCPRO highlights a widely overlooked evaluation loophole: AUPRO is the standard segmentation metric in the ZSAD domain, but its deficiency of evaluating only "recall" without "precision" has been largely unaddressed. This metric design possesses independent value and invites future ZSAD papers to report both AUPRO and RCPRO.
  • Rigorous and fair experimental design: Contrasting DTD and MVTec auxiliary setups cleanly separates method behaviors under "sufficient" versus "insufficient" auxiliary data scenarios. 12 datasets cover both industrial and medical categories, and the ablation study carefully separates the contributions of shallow and deep injections instead of just removing the entire block.

Limitations & Future Work

  • False positives at object boundaries: DIVE's cross-attention module occasionally produces false positive predictions on object contours (e.g., the cables and colons in Fig. 3/Fig. 4), which the authors acknowledge as a limitation. A possible cause is that the attention is activated by sharp texture changes between the object and background (similar to anomaly features), without further differentiating "boundary transitions" from "defect transitions." A potential remedy is to include negative descriptions in the LLM prompts (e.g., "a photo of an object with a clean edge") or introduce boundary negative samples during cross-attention training.
  • Peculiarity of medical segmentation datasets without normal samples: Three medical segmentation datasets (ISIC, ColonDB, TN3K) contain only abnormal images. In the ablation study, this led to AA-CLIP's RCPRO being abnormally higher than DIVE's on ISIC (89.3 vs 74.2), as the lesion regions in ISIC occupy most of the image, and AA-CLIP's over-expansion turned into an "advantage." This exposes the limit of RCPRO in scenarios with extremely high anomaly ratios, where the difference between over-segmentation and accurate segmentation is narrowed.
  • Dependency on LLM description quality: The effectiveness of deep-level injection relies on whether the LLM-generated anomaly/normal descriptions cover sufficiently diverse anomaly concepts. Although the paper uses 100+35 descriptions generated by GPT-4o, if target-domain anomaly categories completely fall outside the description vocabulary (e.g., a completely new anomaly structure), the injection might fail. However, this limitation stems from the nature of zero-shot learning—which always has boundaries between the known and unknown—though future work could explore dynamically expanding the descriptions or making descriptions learnable.
  • Single auxiliary domain setting: The paper validates DIVE under a single auxiliary domain (DTD or MVTec) and does not explore multi-auxiliary-domain joint training or incremental auxiliary domain expansion, which might align better with the realistic challenges of continuous deployment.
  • vs AnomalyCLIP (ICLR 2024): Both execute object-agnostic prompt learning, but AnomalyCLIP optimizes learnable prompts purely on the text side while directly aligning frozen visual features. This performs well with adequate auxiliary anomaly priors but severely overfits when priors are scarce. DIVE's innovation lies in injecting text knowledge directly into the visual encoding process (rather than aligning only at the output layer), making visual features inherently more sensitive to anomalies. This suggests that prompt learning should not be limited to inputs or outputs; integrating it into intermediate representation layers can be far more effective.
  • vs MaPLe (CVPR 2023): DIVE's shallow V-L coupling function directly adopts the multimodal prompting mutual projection concept from MaPLe, but successfully adapts it from general vision recognition to anomaly detection, while appending the critical extension of deep-level LLM description injection. This demonstrates that multimodal prompting methods can be repurposed across tasks, provided that injection hierarchy and content are tailored to downstream tasks.
  • vs AA-CLIP (CVPR 2025): AA-CLIP also performs decoupling (dual-perspective prompts to decouple object semantics and defect appearance) but executes it on the text side (designing different prompts), whereas DIVE decouples on the visual side (explicitly separating global feature subspaces). Both papers independently identified the significance of decoupling for ZSAD but took different routes—text-side decoupling is more lightweight but potentially less thorough, whereas visual decoupling is more complete but adds learnable parameters.
  • Evaluation philosophy of RCPRO: Analogous to the transition from [email protected] to [email protected]:0.95 in object detection—moving from purely evaluating coverage to incorporating precision defines the maturity of an evaluation standard. The dual mechanisms of "false-positive penalization + over-segmentation decay" introduced by RCPRO can inspire evaluation metric designs for other dense prediction tasks, such as change detection or forgery localization.

Rating

  • Novelty: 4/5 —— First to define and systematically study the realistic ZSAD scenario of "limited auxiliary anomaly priors." Although individual components borrow from existing work (e.g., MaPLe's mutual projection, basic decoupling concepts), their combination and adaptation to this new scenario are novel. The identified limitations in AUPRO and the proposed RCPRO also hold independent value.
  • Experimental Thoroughness: 5/5 —— Evaluated across 12 datasets under two auxiliary data settings with 5 SOTA baselines. Includes thorough ablation studies, hyperparameter sensitivity analyses, and visual comparisons of anomaly maps and attention maps. The experimental design is meticulous.
  • Writing Quality: 4/5 —— Clearly structured and driven by two main challenges. Formulations and algorithm descriptions are complete and reproducible. Fig. 1's combination of quantitative curves and qualitative examples is highly persuasive. A minor flaw is some formula truncation in cached texts.
  • Value: 5/5 —— Defines a new subproblem in ZSAD. The proposed shallow-to-deep injection paradigm and decoupling module can be directly reused by future ZSAD works. RCPRO has the potential to become a standard metric in the field. The 16%-47% performance gains under limited auxiliary data settings offer significant practical value for deployment.