{"id":"83e29353-47ab-4e0a-a3ae-23c859c4f330","arxiv_id":"2608.11933","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A dual-anchor CLIP/DINOv3 framework with hierarchical group tokens improves zero-shot anomaly detection by roughly one to three AUROC/AP points across 14 benchmarks.","lead":"This paper proposes a dual-anchor zero-shot anomaly detection method that builds hierarchical image-group tokens and fuses them with text prompts. It reports modest but consistent gains over prior CLIP-based methods across 14 industrial and medical benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All cross-modal similarity computations are dimensionally inconsistent: DINOv3 features are 1024-d, CLIP text embeddings 768-d, and no projection is specified (Eqs. 12, 13, 15).","rationale":"The reader's verdict identified the same region of the paper (Eq. 12) and marked the paper CONDITIONAL. My pass confirms that suspicion but shows it is not limited to the dynamic prompt: the same 1024/768 mismatch appears in the two equations that actually produce the reported anomaly maps and scores (Eqs. 13 and 15). This makes the concern more central because it is not a detail of one module but a property of the entire cross-modal architecture. The claim of 'jointly reinforcing visual and textual semantics' requires a defined way to compare DINOv3 visual tokens with CLIP text embeddings; the manuscript provides none. This is an internal inconsistency, not a disagreement with field consensus. The most likely benign resolution is that the authors silently use a linear projection or an align-then-contrast design; if so, a one-line equation would fix the paper. But until that is stated and the projection is included in released code, the paper's headline results cannot be produced from the text. I therefore move the verdict from CONDITIONAL to UNVERDICTED: the missing projection is a necessary condition for the central claim, and its absence prevents verification. This is not a rejection of the underlying idea; it is a request for the authors to specify the missing component, and for reviewers to withhold judgment until then.","tokens_in":12011,"tokens_out":7675,"duration_ms":76367,"concrete_test":"Obtain the code/checkpoint from the authors, or ask them to specify the projection layer. Concretely, implement the forward pass as written: load DINOv3 ViT-L/16 and CLIP ViT-L/14@336, build Eq. 12 with 1024-d G_N/G_A, Eq. 13 with 1024-d patch features and 768-d text embeddings, and run inference. The shape mismatch will raise an error at the first matmul if no projection exists. Then add one learnable Linear(1024, 768) before all cross-modal interactions and retrain with the same losses; if Tables 1-2 are reproduced, the omitted projection was the issue. If they are not, the reported gains depend on an unspecified component and the central claim is unsupported.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"Sec. 3.2 and the Implementation Details fix the visual backbone as DINOv3 ViT-L/16, whose patch/group tokens are 1024-dimensional, and the text encoder as CLIP ViT-L/14@336, whose token embeddings are 768-dimensional. The framework then routes DINOv3 tokens into the CLIP text branch and multiplies them against CLIP text embeddings in three places. Eq. 12 inserts G_N and G_A directly into the prompt sequence consumed by the CLIP text encoder; the text encoder's embedding space is 768-dimensional, so the 1024-d group tokens cannot be embedded without a projection that is never defined. Eq. 13 computes patch-level anomaly maps as Softmax(Up([...] T^T)) with F^(l), F_assign in R^{M x 1024} and T in R^{2 x 768}; the matrix product is dimensionally invalid. Eq. 15 computes the image-level score as cosine similarity between t_refined_cls and T; cosine similarity is undefined for vectors of different dimensions. No projection, adapter, or shared embedding space is mentioned in Sec. 3.4 or Sec. 3.5, and the loss functions in Eqs. 16-19 do not include any term for training such a layer. This is the single most load-bearing gap: every quantitative result in Tables 1-4 is produced by this forward pass, so if the projection is absent the method as written cannot run; if it is present but omitted, the paper's central claim is not reproducible. The reader singled out Eq. 12, but the same dimensional gap also invalidates Eqs. 13 and 15, so the issue is broader than prompt injection.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a zero-shot anomaly detection (ZSAD) framework that augments the usual text-side anchors with image-side anchors. It extracts 1024-dimensional patch features from DINOv3, progressively merges them into normal/abnormal group tokens via hierarchical merge blocks, uses these tokens to refine the [CLS] token through a Group-Gated Token Refiner, and injects them into CLIP text prompts to form dynamic state prompts. Training on an auxiliary dataset uses classification, segmentation, orthogonality, and group-consistency losses. The method is evaluated on 8 industrial and 6 medical benchmark datasets, with average improvements over six prior ZSAD baselines, supported by component ablations and qualitative visualizations.","tokens_in":12467,"tokens_out":5560,"duration_ms":56806,"significance":"If the method can be made fully specified and reproducible, the dual-anchor idea is a plausible step toward reducing the prompt sensitivity that limits CLIP-based ZSAD. The paper's strengths include broad benchmark coverage, careful module-level ablations, and an explicit attempt to ground text semantics in hierarchical visual structure. However, the current manuscript has a load-bearing dimensional inconsistency in its cross-modal computations, and the shared use of image-derived tokens on both sides of the alignment creates a potential shortcut that is not controlled. These issues must be resolved before the empirical claims can be assessed.","major_comments":[{"comment":"The forward pass is dimensionally inconsistent with the stated backbones. Implementation Details explicitly says DINOv3 ViT-L/16 yields 1024-dimensional patch features and CLIP ViT-L/14@336 yields 768-dimensional embeddings. Eq. 12 inserts the 1024-d group tokens G_N and G_A directly into the prompt consumed by the 768-d CLIP text encoder; Eq. 13 multiplies the 1024-d concatenated visual feature block by T^T, where T is 2 x 768; Eq. 15 computes cosine similarity between the 1024-d refined [CLS] token and the 768-d text embeddings. No projection, adapter, or shared embedding space is described, and none of the losses in Eqs. 16-19 train such a layer. As written, the method cannot execute, so the results in Tables 1-4 are not reproducible. Please specify the missing projection/adaptation, retrain if necessary, and re-report, or clearly remove the unsupported dynamic-state-prompt results.","section":"§3.4–§3.5, Eq. 12, Eq. 13, Eq. 15; Implementation Details"},{"comment":"The design shares the same image-derived group tokens across both branches: G_N and G_A refine the [CLS] token (Eqs. 10-11), and the same tokens are injected into the text prompts whose embeddings are then compared back to t_refined_cls (Eq. 15). This creates a possible shortcut where the normal/abnormal text embeddings are partly a function of the test image itself, so high similarity may reflect the injected visual tokens rather than genuine image-text semantic alignment. The claim that the framework 'stabilizes image-text alignment' needs a control experiment—for example, ablating with group tokens taken from a different image, or removing the prompt injection while keeping the GGTR—to show that the discriminative signal does not come primarily from the shared anchors.","section":"§3.4 and §3.5, Eq. 12 with Eq. 15"},{"comment":"The sentence in §4.2 that 'our model consistently outperforms all existing baselines across both industrial and medical benchmarks' is contradicted by the tables. Bayes-PFL reports higher pixel AUROC than the proposed method on MPDD (97.1 vs 95.9), RSDD (99.6 vs 99.2), KSDD2 (99.6 vs 99.5), and DAGM (99.3 vs 98.3); AA-CLIP reports higher ISIC AP (87.7 vs 84.7); and Bayes-PFL reports higher CVC-ClinicDB AP (53.2 vs 39.8). The claim should be revised to 'on average' or supported by a significance test over datasets.","section":"§4.2, Tables 1 and 2"},{"comment":"The ablation labeled 'w/o Hierarchical Group Merging' removes only the assigned token F_assign, but the group tokens G_N and G_A that are the direct outputs of the merging hierarchy are still used by the Group-Gated Token Refiner and the Dynamic State Prompt. Therefore this ablation does not isolate the contribution of the hierarchical merging mechanism. A cleaner ablation would remove the entire group-token generation path while retaining the same number of parameters where possible, so that the role of the hierarchical grouping itself is measured.","section":"§4.3, Table 3(a)"}],"minor_comments":[{"comment":"In the 'Update' paragraph, the text begins 'Update: Update:' with the word 'Update' duplicated; this appears to be a typo.","section":"§3.3"},{"comment":"The denominator in Eq. 15 is typeset incorrectly; it should read the product of the L2 norms of t_refined_cls and T.","section":"§3.5, Eq. 15"},{"comment":"The affiliation 'NA VER Cloud' appears to be a typo for 'NAVER Cloud'; please correct it.","section":"Author affiliations"},{"comment":"The notation [1/L sum_l F_i^(l), F_assign] should specify whether concatenation is along the token dimension or the feature dimension; as written, the resulting matrix dimension is ambiguous even apart from the dimension mismatch with T.","section":"§3.5, Eq. 13"},{"comment":"Please state the exact temperature values for the Gumbel-Softmax assignment (Eq. 2) and the cosine similarity (Eq. 15), as well as the loss weighting coefficients in Eq. 19; these are omitted from the main text and the supplementary material is not available in the reviewed version.","section":"§4.1, Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The dimensional inconsistency in Eqs. 12, 13, and 15 is severe enough that I would not accept the paper without a corrected and reproducible method description. If the missing projection cannot be provided, the authors should withdraw the dynamic-state-prompt results rather than leave the paper in its current form. The overclaimed 'consistently outperforms' statement should be corrected in any revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The idea is worth taking seriously, but the math is not. The stress test is right, and broader than the reader's note: DINOv3 group tokens are 1024-dimensional, CLIP text embeddings are 768-dimensional, and the paper feeds the group tokens directly into the text encoder (Eq. 12), multiplies them against text embeddings in the anomaly map (Eq. 13), and computes cosine similarity between differently sized vectors in the final score (Eq. 15). No projection, adapter, or shared embedding space is described anywhere, and none of the losses in Eqs. 16-19 train one. As written, the method cannot run. If the authors intended a projection layer and omitted it from the writeup, that is a serious reproducibility failure; if they did not, the empirical results are unsupported. Either way, the central claim fails as presented.\n\nWhat is genuinely new: the dual-anchor formulation, using hierarchical group merging to build image-side normal/anomaly anchors, the gated CLS refiner, and the dynamic state prompt. The components are established in the cited literature, but the specific integration is not present in prior ZSAD work, and the broad evaluation across 14 industrial and medical benchmarks is a real plus. The ablations in Table 3 are internally consistent and point to each module contributing. The qualitative figures are plausible, though not evidence.\n\nSoft spots beyond the dimensional gap: no code, no error bars, no seeds. The group token configuration in Table 4 appears to be selected after seeing target-benchmark results, which makes the reported gains look optimistic. The t-SNE visualization is illustrative, not quantitative. The limitation section is honest about logical defects but does not mention the missing projection.\n\nWho this is for: researchers working on CLIP-based zero-shot anomaly detection. They will find the dual-anchor idea stimulating and the benchmark coverage useful, but no one can reproduce the method from the text, so the value right now is as an idea, not as a result.\n\nRecommendation: if this crosses my desk, I would not desk-reject it outright—the idea deserves a serious referee—but I would send it back with the expectation of major revision. The authors need to specify the projection, correct the dimensional inconsistency, release code or at least detailed hyperparameters, and report variance. In its current form, I cannot endorse acceptance.","headline":"The dual-anchor idea is plausible and worth watching, but the forward pass as written is dimensionally impossible—1024-d image tokens fed into a 768-d text encoder with no projection—so the reported numbers are unsupported until that is fixed.","tokens_in":12996,"tokens_out":3060,"would_cite":false,"duration_ms":33359,"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":"The paper proposes a Dual-Anchor framework that adds hierarchical image anchors to text prompts and reports state-of-the-art zero-shot anomaly detection across 14 benchmarks.","keywords":["zero-shot anomaly detection","dual-anchor framework","hierarchical group merging","dynamic state prompt","CLIP","DINOv3","anomaly localization","industrial and medical benchmarks"],"falsifier":"Implement the dynamic state prompt exactly as written in Eq. (12): the 1024-dimensional $G_N$ token cannot pass through the 768-dimensional token embedding layer of the CLIP text encoder, so the missing projection must be specified before the reported gains can be attributed to the prompt. Then run the prompt ablation with the group tokens replaced by random 1024-dimensional vectors; if the image-level AUROC on MVTec-AD stays near the reported 92.7, the visual conditioning is not carrying the claimed signal.","tokens_in":11792,"feed_emoji":"🔍","tokens_out":9431,"duration_ms":85906,"temperature":0.7,"pith_summary":"This paper tries to establish a new recipe for zero-shot anomaly detection—spotting defects in domains never seen during training: instead of anchoring normal and abnormal semantics only in text prompts, it builds visual anchors by hierarchically merging image features, then fuses both kinds of anchors. The proposed Dual-Anchor framework forms normal and abnormal group tokens from DINOv3 visual features, uses those tokens to gate and refine the global image token, and inserts them into CLIP prompts so the text branch becomes image-dependent. Across 8 industrial and 6 medical benchmarks, the paper reports the best image-level and pixel-level results among the compared zero-shot anomaly detection methods, including gains over the strongest existing baseline. If the claim is right, prompt sensitivity and text bias are not unavoidable weaknesses of CLIP-based anomaly detection, because structured visual grouping can carry part of the semantic load.","feed_headline":"Image anchors beat text-only prompts on 14 anomaly benchmarks","feed_subtitle":"Hierarchical visual anchors reduce prompt sensitivity and generalize from factory floors to medical scans.","key_machinery":"The load-bearing objects are the two learnable image anchors $G_N$ and $G_A$ produced by Hierarchical Group Merging. Each Merge Block runs four stages—initialization by cross-attention with previous group tokens, Gumbel-Softmax assignment of patch tokens to groups, bipartite soft matching that merges the most similar group pairs, and a final update that assigns merged tokens back to the patches—so the two anchors accumulate local-to-global visual semantics. The Group-Gated Token Refiner computes cosine gating weights between the [CLS] token and the two anchors, passes the concatenated weights through normalization, a linear layer, and a sigmoid, then adds the result to [CLS] as a residual connection. The Dynamic State Prompt concatenates the visual group tokens into text prompts before CLIP encoding, which is the step that makes the text embedding image-dependent and directly reduces reliance on prompt wording.","core_discovery":"The central claim is that anchoring semantics on both modalities—text prompts plus hierarchical image anchors—yields more stable and more generalizable zero-shot anomaly detection than text-only anchoring. The paper constructs image anchors with a top-down Hierarchical Group Merging process that assigns DINOv3 patch tokens to learnable groups, merging the most similar groups until two final group tokens $G_N$ and $G_A$ stand for normal and abnormal states. These two tokens gate the [CLS] token through a Group-Gated Token Refiner and are inserted into dynamic state prompts of the form $[V_1]\\cdots[V_E][W_1][G_N][\\text{class}]$, making the text branch depend on the input image. The paper reports that this design outperforms all compared baselines on 8 industrial and 6 medical benchmarks, with the full model reaching 92.7 image-level AUROC and 92.4 pixel-level AUROC on MVTec-AD, and that removing any of the three modules lowers performance.","pith_inferences":["The dual-anchor recipe is not tied to CLIP's text encoder: the same hierarchical image anchors could be paired with other vision-language encoders, since the anchors are learned from auxiliary data rather than from target labels.","Because the image anchors are built without target-domain labels, the approach should extend naturally to few-shot anomaly detection, where a few normal examples could refine the group tokens before deployment.","A concrete stress test the paper does not run is a prompt-sensitivity sweep: varying the wording of the class prompt while keeping the visual anchors fixed would quantify how much of the reported stability comes from the image side."],"forward_implications":["Zero-shot anomaly detection no longer has to rely on carefully engineered or heavily tuned prompt wording; the image group tokens carry normal and abnormal semantics into the text branch.","Hierarchical merging of visual features provides a local-to-global representation that improves pixel-level localization, not just image-level classification, on both object-centric and texture datasets.","The dual-anchor design transfers across domain boundaries: the same auxiliary-trained recipe works on industrial defects and medical lesions without any target-domain labels.","Ablations show each component is load-bearing: removing Hierarchical Group Merging, the Group-Gated Token Refiner, or the Dynamic State Prompt each lowers both image-level and pixel-level metrics.","The number of group tokens per merge stage is a real hyperparameter, with [16, 8, 4, 2] outperforming both smaller and larger settings in the paper's ablation."],"supporting_citations":[{"why":"CLIP supplies the frozen text encoder and the image-text alignment space that the Dual-Anchor framework must stabilize.","marker":"[23]"},{"why":"DINOv3 supplies the multi-layer patch features and the 1024-dimensional representations from which the image anchors are built.","marker":"[25]"},{"why":"GroupViT provides the learnable group-token merging concept that Hierarchical Group Merging adapts to anomaly detection.","marker":"[28]"},{"why":"Token Merging supplies the bipartite soft matching procedure used to merge group tokens without learned parameters.","marker":"[4]"},{"why":"Bayes-PFL defines the benchmark configurations used in the experiments and is the strongest baseline the paper compares against.","marker":"[22]"},{"why":"AnomalyCLIP is the object-agnostic learnable-prompt baseline that motivates the paper's goal of reducing prompt dependence.","marker":"[30]"},{"why":"WinCLIP establishes the hand-crafted-prompt baseline and the multi-layer visual feature usage the paper builds on.","marker":"[15]"},{"why":"AdaCLIP supplies the hybrid learnable-prompt baseline for zero-shot anomaly detection.","marker":"[6]"}],"fun_headline_variants":["Dual anchors outperform text-only on 14 anomaly benchmarks","Hierarchical visual anchors cut prompt sensitivity in anomaly detection","Gated visual anchors stabilize zero-shot anomaly detection","Image-text dual anchors reduce prompt reliance in zero-shot detection","Hierarchical grouping enables dual anchors for robust anomaly detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's dynamic state prompt inserts 1024-dimensional visual group tokens directly into a CLIP text encoder that embeds 768-dimensional tokens, and no projection or adapter for that dimension mismatch is described.","fun_headline_variants_meta":{"raw":{"variants":["Dual anchors outperform text-only on 14 anomaly benchmarks","Hierarchical visual anchors cut prompt sensitivity in anomaly detection","Gated visual anchors stabilize zero-shot anomaly detection","Image-text dual anchors reduce prompt reliance in zero-shot detection","Hierarchical grouping enables dual anchors for robust anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001198,"raw_usage":{"total_tokens":4922,"prompt_tokens":912,"completion_tokens":4010,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":3933}},"tokens_in":528,"tokens_out":4010,"duration_ms":31894,"temperature":1.0,"reasoning_tokens":3933,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:22:11.364731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement the dynamic state prompt exactly as written in Eq. (12): the 1024-dimensional $G_N$ token cannot pass through the 768-dimensional token embedding layer of the CLIP text encoder, so the missing projection must be specified before the reported gains can be attributed to the prompt. Then run the prompt ablation with the group tokens replaced by random 1024-dimensional vectors; if the image-level AUROC on MVTec-AD stays near the reported 92.7, the visual conditioning is not carrying the claimed signal.","supporting_citations":[{"cited_title":"Learning transferable visual models from natural language supervi- sion","cited_arxiv_id":null,"evidence_quote":"CLIP supplies the frozen text encoder and the image-text alignment space that the Dual-Anchor framework must stabilize."},{"cited_title":"Bayesian prompt flow learning for zero-shot anomaly detec- tion","cited_arxiv_id":null,"evidence_quote":"Bayes-PFL defines the benchmark configurations used in the experiments and is the strongest baseline the paper compares against."},{"cited_title":"Winclip: Zero- /few-shot anomaly classification and segmentation","cited_arxiv_id":null,"evidence_quote":"WinCLIP establishes the hand-crafted-prompt baseline and the multi-layer visual feature usage the paper builds on."},{"cited_title":"Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly de- tection","cited_arxiv_id":null,"evidence_quote":"AdaCLIP supplies the hybrid learnable-prompt baseline for zero-shot anomaly detection."}],"review_version":1}