{"id":"4764a6e6-f560-483a-ae20-1bcef00e40e4","arxiv_id":"2509.10282","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"MCL-AD fuses point cloud, RGB, and text modalities with learnable decoupled prompts and a contrastive loss, reporting top scores on two 3D anomaly detection benchmarks.","lead":"This paper proposes MCL-AD, a framework for zero-shot 3D anomaly detection that combines point clouds, RGB images, and text prompts using CLIP encoders. The authors report state-of-the-art results on MVTec3D-AD and Eyecandies, but the training protocol and test-set hyperparameter tuning undermine the zero-shot claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is invalidated by protocol: MCL-AD trains on the test-split ground-truth masks and labels, so the reported zero-shot SOTA is unsubstantiated.","rationale":"The reader's weakest assumption exactly identifies the load-bearing flaw. The paper's own Section IV-A discloses that only the test sets are used, and Eqs. (20), (23), (26), and (28) show ground-truth masks and labels entering every loss. A method advertised as zero-shot cannot train on the images, point clouds, and annotations that later constitute its evaluation. This is not a disagreement with community consensus or a stylistic issue; it is a direct violation of the setting the paper claims to address. The ablation tables also reveal that the fusion weight eta and contrastive weight lambda3 are tuned on the test set, further fitting the evaluation data. In addition, the text contains multiple internal inconsistencies (e.g., Section IV-B reports an AUPRO gain of 36.1% versus the table's 56.0% baseline and 92.2% final, and Table I's PointAD pixel-level mean row is malformed), but the data-leakage issue alone is sufficient to invalidate the central SOTA claim. I therefore agree with the REJECT verdict; the stress test does not change it. The concrete test above would settle whether a corrected protocol preserves any of the reported gains, but as currently written the claim is unsupported.","tokens_in":20916,"tokens_out":3669,"duration_ms":31400,"concrete_test":"Re-run MCL-AD under a true zero-shot protocol: remove every ground-truth mask or label from the training objective (drop or replace the supervised Focal/Dice/cross-entropy terms in Eqs. (20), (23), (26), (28)), fix eta and lambda3 using only a held-out validation split or a predetermined rule, and then evaluate once on MVTec3D-AD and Eyecandies test sets. If I-AUROC and P-AUROC fall materially below the reported 89.0 and 97.7, the headline claim is an artifact of test-set supervision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A explicitly states that 'the training sets of both datasets only consist of normal samples, so we exclusively use their test sets, which contain point-level annotations for point clouds and pixel-level annotations for images.' The method therefore trains on the evaluation data. The losses in Eqs. (20), (23), (26), and (28) directly supervise the model with ground-truth masks M_i^(k), M_i, G_i^(k), and G_i, using Focal/Dice and cross-entropy terms. This contradicts the abstract's claim of detecting defects 'without relying on labeled training data' and makes the within-dataset numbers in Tables I and II measures of supervised fitting on the test set, not zero-shot generalization. Independently, hyperparameters eta and lambda3 are selected by ablations on the same test set (Tables VII and VIII; Section III-D sets lambda3=0.8 'based on the experiment in Section IV-C6'), so part of the reported gain is fitted to the evaluation data. Under a correct zero-shot protocol, the model would have no access to any test annotations and the prompt/hyperparameter choices would be fixed before seeing the test set; the paper provides no evidence that the 89.0 I-AUROC / 97.7 P-AUROC results survive that protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MCL-AD, a multimodal framework for zero-shot 3D anomaly detection that combines frozen CLIP encoders for RGB images and multi-view rendered depth images with learnable text prompts, a multimodal contrastive loss, and a collaborative modulation mechanism for inference-time fusion. Training optimizes point-cloud and RGB local/global losses plus the contrastive loss; testing uses the fused branch outputs. The authors report state-of-the-art results on MVTec3D-AD and Eyecandies under both within-dataset and cross-dataset settings, with image-level AUROC improving from 71.9% (MVP-PCLIP) to 89.0% and point-level AUROC from 87.5% to 97.7% on MVTec3D-AD.","tokens_in":21143,"tokens_out":5145,"duration_ms":40334,"significance":"If the experimental protocol were valid, the reported gains would be substantial and practically relevant for zero-shot 3D anomaly detection. The paper offers a clearly described architecture, a sensible motivation for combining RGB and point-cloud modalities, and extensive ablations over prompt design and fusion hyperparameters. These strengths, however, do not offset the core problem: the method is trained on the ground-truth anomaly annotations of the test splits, and key hyperparameters are selected using the same test-set metrics. Consequently, the reported numbers do not demonstrate zero-shot generalization, and the central claim of the paper is unsupported.","major_comments":[{"comment":"The manuscript states that 'the training sets of both datasets only consist of normal samples, so we exclusively use their test sets, which contain point-level annotations for point clouds and pixel-level annotations for images.' The local and global losses in Eqs. (20), (23), (26), and (28) directly supervise the model with the ground-truth masks M_i^(k), M_i and labels G_i^(k), G_i. This means the method uses anomaly annotations from the evaluation data during training, contradicting the abstract's claim of detecting defects 'without relying on labeled training data.' Under this protocol, Tables I and II measure supervised fitting on the test set, not zero-shot generalization.","section":"IV-A, Eqs. (20), (23), (26), (28)"},{"comment":"The fusion weight η and the contrastive-loss weight λ3 are chosen by optimizing the same evaluation metrics on the MVTec3D-AD test set: λ3 is set to 0.8 'based on the experiment in Section IV-C6,' and η is set to 0.8 from the ablation in Table VII. This test-set-based selection of hyperparameters means part of the reported gain is fitted to the evaluation data. A credible zero-shot evaluation would fix these values before seeing the test set, and the paper provides no evidence that the reported results survive such a protocol.","section":"III-D, IV-C6, Tables VII and VIII"},{"comment":"The within-dataset evaluation protocol is underspecified. The text says the model is 'evaluated on test objects that belong to different categories but come from the same dataset,' but it does not state which categories are used for training the prompts versus held out for testing, nor whether the split is per category or per object. Without this information, the within-dataset numbers in Tables I and II are not reproducible, and it is impossible to assess whether the method generalizes across categories or merely memorizes category-specific test annotations.","section":"IV-A, Evaluation Setting"}],"minor_comments":[{"comment":"The definitions of L_point and L_rgb are written as L_point = F_local^p + L_global^p and L_rgb = F_local^rgb + L_global^rgb, where F_local^p and F_local^rgb are similarity maps rather than scalar losses; please clarify how these terms are reduced to scalar loss values.","section":"Eqs. (24) and (29)"},{"comment":"In the formula for Score_point, the global feature f_global^rgb appears in the numerator, which appears to be a typo for f_global^point; please correct.","section":"Eq. (35)"},{"comment":"The text states that 'the P-AUROC improving from 60.4% to 97.7%,' but in Table IV the value 60.4 is the I-AUROC of the baseline row; the P-AUROC of that row is not reported. Please correct the description.","section":"IV-C1, Table IV"},{"comment":"The sentence 'the training sets of both datasets only consist of normal samples, so we exclusively use their test sets' is confusing because the method is actually trained on the test sets; please rephrase to state precisely what data are used for training and evaluation.","section":"IV-A, Datasets"},{"comment":"References [13] and [46] refer to the same paper (AnomalyCLIP); please merge them or cross-reference them to avoid duplication.","section":"References [13] and [46]"},{"comment":"The pixel-level rows contain incomplete entries such as '(-, )' and appear misaligned; please reformat the table for clarity.","section":"Table I"},{"comment":"The rotation angles for rendering are given as five values for the X axis and four values for the Y axis, which does not obviously produce the stated 9 views; please clarify how the nine viewpoints are generated.","section":"IV-A, Implementation details"}],"recommendation":"reject","confidential_remarks":"The protocol issue is fundamental: the method trains on the test splits' ground-truth masks and labels and selects hyperparameters using test-set metrics, so the reported numbers cannot be interpreted as zero-shot performance. The paper would need a completely redesigned evaluation (e.g., prompt learning on normal samples only or on source categories only, with hyperparameters fixed on a validation split) to substantiate its claims. I cannot see a way to fix this within the scope of a revision; the experiments must be redone."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read MCL-AD. The central claim—state-of-the-art zero-shot 3D anomaly detection—does not survive a close look at the protocol. Section IV-A says the training sets of MVTec3D-AD and Eyecandies consist only of normal samples, so the authors 'exclusively use their test sets, which contain point-level annotations for point clouds and pixel-level annotations for images.' The losses in Eqs. (20), (23), (26), and (28) directly supervise the model with ground-truth masks M_i^(k), M_i, G_i^(k), and G_i. That means the model is trained on the evaluation data. The abstract's 'without relying on labeled training data' is contradicted by the paper's own statement. This is not a minor oversight; it invalidates the within-dataset numbers in Tables I and II as zero-shot results. The cross-dataset experiment uses MVTec3D-AD's test set as the training source, so it is also not clean.\n\nA second, related problem: the fusion weight eta (Table VII) and the contrastive loss weight lambda3 (Table VIII) are chosen on the MVTec3D-AD test set. The reported 89.0 I-AUROC and 97.7 P-AUROC are therefore partly fitted to the evaluation data, not parameter-free predictions. The paper also does not specify the within-dataset category split, and no code is released, so the results cannot be independently checked.\n\nTo give credit where it's due: the specific combination—object-agnostic decoupled prompts for RGB and point clouds, a triplet contrastive loss between prompts, and the collaborative modulation fusion—is new relative to the cited MVP-PCLIP and AnomalyCLIP lines. The motivation is sensible: texture cues from RGB and geometric cues from depth are genuinely complementary for industrial inspection. If the authors re-run the experiments with a proper protocol (no test annotations used at any stage, hyperparameters fixed before seeing the test set, and a clear train/category split), the underlying method might be competitive. But as it stands, the claimed gains are not zero-shot results.\n\nThere are also smaller issues: some text numbers do not match the tables (e.g., the Eyecandies pixel-level phrasing), and the equations have typos (e.g., the first term in Eq. (24) should be L_local_p, not F_local_p). These are minor compared to the protocol problem.\n\nThis paper is for a knowledgeable reader who wants to see the current state of multimodal 3D anomaly detection, but only with the evaluation issues in mind. I would not cite it in its present form. If the editor asks, I'd suggest sending it to a serious referee because the method idea is worth examining, but the referee should be instructed to check the training protocol carefully. My own verdict is clear: reject in current form; a properly re-run version could be reconsidered.","headline":"The reported zero-shot SOTA is invalidated by the paper's own protocol: training on test-split ground-truth masks and tuning fusion weights on the same test set, so the gains are partly fitted, not predicted.","tokens_in":21709,"tokens_out":3628,"would_cite":false,"duration_ms":29593,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes MCL-AD, a zero-shot 3D anomaly detection framework that fuses point clouds, RGB images, and text prompts; it reports image-level AUROC of 89.0% and point-level AUROC of 97.7% on MVTec3D-AD.","keywords":["zero-shot 3D anomaly detection","multimodal learning","prompt learning","point cloud","RGB image","CLIP","MVTec3D-AD","Eyecandies"],"falsifier":"Run MCL-AD with the same architecture and hyperparameters but without access to any ground-truth masks, replacing the Focal, Dice, and cross-entropy losses in Eqs. (20), (23), (26), and (28) with normal-only objectives, and measure I-AUROC and P-AUROC on MVTec3D-AD. If the scores fall below the reported 89.0% and 97.7% and approach MVP-PCLIP's 71.9% and 87.5%, the gain is explained by test-set supervision rather than by multimodal collaboration.","tokens_in":20668,"feed_emoji":"🔍","tokens_out":13650,"duration_ms":98746,"temperature":0.7,"pith_summary":"Zero-shot 3D anomaly detection, finding defects in objects never seen during training, usually reads only point clouds, but the paper argues that the task is better solved by making three modalities talk to one another: point-cloud geometry, RGB appearance, and text semantics. The proposed MCL-AD renders point clouds into depth views, pushes RGB and depth features through a frozen CLIP encoder, and learns separate normal/anomaly text prompts for each branch. A multimodal contrastive loss aligns the two branches on semantic state, and a collaborative modulation mechanism reweights their score maps at test time so texture defects and geometric defects reinforce each other. On MVTec3D-AD the framework reports 89.0% image-level AUROC and 97.7% point-level AUROC, above the point-cloud-only MVP-PCLIP baseline, and it also improves on Eyecandies and in cross-dataset transfer. The claimed payoff is an industrial inspection method that adapts to novel objects without category-specific training.","feed_headline":"Multimodal fusion pushes zero-shot 3D defect detection to 89% AUROC","feed_subtitle":"MCL-AD aligns point clouds, RGB images, and text prompts to beat single-modality baselines on two 3D benchmarks.","key_machinery":"The load-bearing object is the object-agnostic decoupled text prompt (ODTP): two pairs of learnable prompt templates, one pair for RGB and one for point clouds, with the class label replaced by the generic word 'object' so the prompts carry no category bias. The multimodal contrastive loss (MCL) makes these prompts interoperable by treating the point-cloud prompt as an anchor, pulling the RGB prompt of the same semantic state closer and pushing the opposite state away. The collaborative modulation mechanism (CMM) is the inference-side counterpart: it fuses the two branches' anomaly score maps as $\\frac{\\eta\\,G_\\sigma(M_{\\mathrm{rgb}}) + (2-\\eta)\\,G_\\sigma(M_{\\mathrm{point}})}{2}$ with a dataset-dependent weight. Multi-view rendering and inverse rendering supply the shared feature space, since a frozen CLIP encoder sees only 2D views.","core_discovery":"The paper's central claim is that zero-shot 3D anomaly detection is a multimodal alignment problem, not a point-cloud-only problem. MCL-AD renders each point cloud into nine depth views, extracts RGB and depth features with a frozen CLIP image encoder, and learns four text-prompt embeddings: normal and anomalous prompts for the RGB branch, and normal and anomalous prompts for the point-cloud branch. An object-agnostic decoupled prompt template removes category priors, and a multimodal contrastive loss uses the point-cloud prompt as anchor to pull the same-state RGB prompt near while pushing the opposite-state prompt away. At inference, a collaborative modulation mechanism reweights the two branches' score maps with a balance weight. The authors report that this recipe lifts MVTec3D-AD image-level AUROC from 71.9% (MVP-PCLIP) to 89.0% and point-level AUROC from 87.5% to 97.7%, with consistent gains on Eyecandies and in cross-dataset transfer.","pith_inferences":["Retrain MCL-AD with the ground-truth masks removed from the local losses in Eqs. (20) and (26) and compare; the drop in AUROC would quantify how much of the 'zero-shot' gain comes from test-set supervision rather than multimodal collaboration.","The same decoupled-prompt idea could be tested on cheaper 2.5D inputs such as stereo depth or surface normals, where geometry cues are available without a full point-cloud rendering pipeline.","A fair cross-method comparison would hold the supervision budget constant; the baselines compared here are not described as training on test-set anomaly masks, so without that constraint the reported margins are not directly attributable to the architecture."],"forward_implications":["If the reported numbers are right, point-cloud-only zero-shot 3D detectors are leaving a large margin on the table, and fusing RGB and text at the prompt level becomes the effective baseline.","The object-agnostic decoupled prompts should transfer to unseen object categories without per-category prompt engineering, which is what an inspection system facing novel parts needs.","The cross-dataset results imply a model can be tuned on one industrial 3D dataset and deployed on another with no target-domain annotation, at a modest accuracy cost.","The dataset-dependent fusion weight $\\eta$ shows that no single static fusion rule dominates, so adaptive fusion is a necessary component for practical deployment."],"supporting_citations":[{"why":"Supplies the main MVTec3D-AD benchmark with paired point clouds, RGB images, and ground-truth anomaly masks.","marker":"[20]"},{"why":"Supplies the Eyecandies benchmark and the cross-dataset transfer target.","marker":"[33]"},{"why":"Provides the frozen CLIP image and text encoders whose shared embedding space all modalities are projected into.","marker":"[29]"},{"why":"Establishes the prompt-based zero-shot anomaly detection paradigm that MCL-AD extends from 2D to 3D.","marker":"[9]"},{"why":"The point-cloud-only MVP-PCLIP baseline whose reported MVTec3D-AD numbers MCL-AD claims to beat.","marker":"[27]"},{"why":"The PointAD point-and-pixel zero-shot 3D baseline used in within-dataset and cross-dataset comparisons.","marker":"[31]"},{"why":"Contributes the object-agnostic prompt strategy that MCL-AD adapts into decoupled RGB and point-cloud prompts.","marker":"[46]"},{"why":"Supplies the projection-based rendering of point clouds into depth maps that the point-cloud branch relies on.","marker":"[45]"}],"fun_headline_variants":["MCL-AD: multimodal zero-shot 3D anomaly detection","Point clouds + RGB + texts lift zero-shot 3D defect detection","Zero-shot 3D anomaly detection gets a multimodal edge","Multimodal collaboration for zero-shot 3D anomaly detection","Fusing modalities boosts zero-shot 3D anomaly detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that training on the test splits of MVTec3D-AD and Eyecandies, including the ground-truth anomaly masks used in the loss functions, counts as zero-shot because no object-specific normal training split is used; if test-set annotations are labeled training data, the reported gains and the zero-shot claim stand or fall with that allowance.","fun_headline_variants_meta":{"raw":{"variants":["MCL-AD: multimodal zero-shot 3D anomaly detection","Point clouds + RGB + texts lift zero-shot 3D defect detection","Zero-shot 3D anomaly detection gets a multimodal edge","Multimodal collaboration for zero-shot 3D anomaly detection","Fusing modalities boosts zero-shot 3D anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1304,"prompt_tokens":965,"completion_tokens":339,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":252}},"tokens_in":581,"tokens_out":339,"duration_ms":3001,"temperature":1.0,"reasoning_tokens":252,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:55:32.113576+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MCL-AD with the same architecture and hyperparameters but without access to any ground-truth masks, replacing the Focal, Dice, and cross-entropy losses in Eqs. (20), (23), (26), and (28) with normal-only objectives, and measure I-AUROC and P-AUROC on MVTec3D-AD. If the scores fall below the reported 89.0% and 97.7% and approach MVP-PCLIP's 71.9% and 87.5%, the gain is explained by test-set supervision rather than by multimodal collaboration.","supporting_citations":[{"cited_title":"The mvtec 3d-ad dataset for unsupervised 3d anomaly de- tection and localization,","cited_arxiv_id":null,"evidence_quote":"Supplies the main MVTec3D-AD benchmark with paired point clouds, RGB images, and ground-truth anomaly masks."},{"cited_title":"The eyecandies dataset for unsuper- vised multimodal anomaly detection and localization,","cited_arxiv_id":null,"evidence_quote":"Supplies the Eyecandies benchmark and the cross-dataset transfer target."},{"cited_title":"Learning transferable visual models from natural language supervision,","cited_arxiv_id":null,"evidence_quote":"Provides the frozen CLIP image and text encoders whose shared embedding space all modalities are projected into."},{"cited_title":"Winclip: Zero-/few-shot anomaly clas- sification and segmentation,","cited_arxiv_id":null,"evidence_quote":"Establishes the prompt-based zero-shot anomaly detection paradigm that MCL-AD extends from 2D to 3D."},{"cited_title":"Pointad: Comprehending 3d anomalies from points and pixels for zero-shot 3d anomaly detection,","cited_arxiv_id":null,"evidence_quote":"The PointAD point-and-pixel zero-shot 3D baseline used in within-dataset and cross-dataset comparisons."},{"cited_title":"Anoma- lyclip: Object-agnostic prompt learning for zero-shot anomaly detection,","cited_arxiv_id":null,"evidence_quote":"Contributes the object-agnostic prompt strategy that MCL-AD adapts into decoupled RGB and point-cloud prompts."},{"cited_title":"Pointclip V2: prompting CLIP and GPT for powerful 3d open-world learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the projection-based rendering of point clouds into depth maps that the point-cloud branch relies on."}],"review_version":2}