Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

EMIT: Enhancing MLLMs for Industrial Anomaly Detection via Difficulty-Aware GRPO

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read EMIT shows that difficulty-aware GRPO, with response resampling and advantage reweighting, lifts an 8B multimodal language model's industrial anomaly detection accuracy by 7.77 points on MMAD.

desk verdict Genuine extension of GRPO for industrial anomaly detection with clean ablations, but the MMAD-constituent text-sample construction leaves a real leakage question hanging over the headline 7.77% gain. read the letter →

arxiv 2507.21619 v1 pith:PV5ET5YF submitted 2025-07-29 cs.CV

Wei Guan , Jun Lan , Jian Cao , Hao Tan , Huijia Zhu , Weiqiang Wang This is my paper
classification cs.CV
keywords industrialanomalydetectionmultimodallargelanguagemodelsgrouprelativepolicyoptimizationdifficulty-awaretrainingfew-shotcontrastiveembeddingsMMADbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

EMIT sets out to show that a multimodal large language model can be made competent at industrial anomaly detection by tuning it with a reinforcement-learning objective that deliberately targets hard examples. The core proposal is difficulty-aware GRPO: when a sampled group of model responses contains no correct answer, the policy resamples a fresh group, and the advantage of every response is reweighted by the fraction of incorrect responses so that difficult questions produce stronger learning signals. Around that mechanism, EMIT builds a multi-task question dataset from public industrial-inspection annotations, uses GPT-4-generated descriptive text as a stand-in for defective images that are missing, and adds a soft prompt plus heatmap-guided contrastive embeddings for one/few-shot comparisons. On MMAD, a multimodal benchmark for industrial anomaly detection, the full recipe raises InternVL3-8B's average accuracy from 74.18% to 81.95%, a gain of 7.77 percentage points across seven tasks.

What carries the argument

The load-bearing mechanism is the difficulty-aware GRPO, a modification of group relative policy optimization in which the policy samples $G$ responses per question and, if none matches the ground-truth answer, resamples until a correct response exists; the advantage of each response is then multiplied by $w=(I/G)+1$, with $I$ the number of incorrect responses, so difficult questions dominate the gradient. A second mechanism carries the few-shot part: heatmap-guided contrastive embeddings. For each selected ViT layer, every patch of the query image is compared by cosine distance with patches of the reference image in a local window, the minimum distance forms a 2D heat map, and a projector flattens the aggregated map into embeddings that the MLLM consumes alongside a learnable soft prompt. These two mechanisms are what the paper credits for the average accuracy gain and for the gains on anomaly discrimination and defect localization specifically.

What would settle it

Train EMIT's full recipe with normal-image and object-type sampling restricted to industrial datasets disjoint from MMAD, then evaluate on MMAD; if the accuracy gain over the base model collapses toward zero while a model trained with MMAD-sourced normal images keeps the 7.77-point gain, the reported improvement is largely evaluation-distribution familiarity rather than transferable anomaly detection.

Watch

Extended reading notes

Core claim

The paper's central discovery claim is that the failure mode of plain GRPO on hard anomaly samples is fixable, and fixing it is what unlocks MLLM performance on industrial anomaly detection. In plain GRPO, a question whose sampled responses are all wrong offers no correct reference for the model to learn from; EMIT's difficulty-aware GRPO resamples the response group until at least one correct answer appears, and then scales the standardized advantage of every response by a difficulty weight $w=(I/G)+1$, where $I$ is the number of incorrect responses in the group. This makes hard questions contribute larger gradient updates instead of being wasted or misleading. EMIT combines this with a two-stage recipe: first align a learnable soft prompt and a projector for heatmap-guided contrastive embeddings under SFT, then fine-tune the full model with the difficulty-aware GRPO, using a reward that mixes format adherence, answer correctness, a cosine reward tied to reasoning length, and a repetition penalty. The reported result is an average accuracy of 81.95% on MMAD, 7.77 points over the InternVL3-8B base model and ahead of the GRPO-tuned AnomalyR1 baseline.

Load-bearing premise

The load-bearing premise is that the training pipeline does not leak the evaluation distribution: five normal images per object type are sampled from MMAD's constituent datasets and turned into GPT-generated training text, so if that counts as training on the test classes, the measured 7.77-point gain could come substantially from object-class familiarity rather than from learned anomaly-detection skill.

Editorial extensions

If this is right

  • EMIT raises InternVL3-8B's average MMAD accuracy from 74.18% to 81.95%, and every one of the seven subtasks improves over the base model, with the largest gains on defect classification and localization.
  • Difficulty-aware GRPO beats plain GRPO by 2.79 points and SFT by 5.03 points on average, so the hard-sample resampling and reweighting, not just more fine-tuning, is what drives the improvement.
  • Omission of Stage 2 costs 6.62 points versus 0.89 points for Stage 1, making the difficulty-aware GRPO the dominant part of the pipeline.
  • Training with GPT-generated object-description text as a substitute query adds 2.58 points over image-only training, supporting the claim that text can compensate for missing defective images.
  • Removing the soft prompt and contrastive embeddings lowers accuracy on anomaly discrimination and defect localization, confirming their role in the one-shot setting.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Extension beyond the paper: the data-collection step draws five normal images per object type from MMAD's own subdatasets and converts them into training text, so part of the 7.77-point gain could come from exposing the model to the evaluation benchmark's object classes and normal-image statistics; a hold-out evaluation on object categories never seen in training would separate leakage from genuin
  • Extension beyond the paper: the ablation removes Stage 2 wholesale and never isolates response resampling from advantage reweighting, so the individual contribution of each mechanism to the 2.79-point edge over plain GRPO is untested; a four-way ablation would pin it down.
  • Extension beyond the paper: the text-as-image substitution could be tested by feeding GPT-generated descriptions of defective objects at inference time in place of the query image; comparable accuracy would indicate the model leans on semantic domain knowledge, while a sharp drop would show visual defect patches are indispensable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes EMIT, a framework for adapting multimodal large language models (MLLMs) to industrial anomaly detection (IAD). The method (i) constructs a multi-task training dataset from existing IAD datasets, (ii) uses GPT-4-generated object text descriptions to compensate for missing defective images, (iii) adds a soft prompt and heatmap-guided contrastive embeddings for one-/few-shot detection, and (iv) introduces difficulty-aware GRPO with response resampling and advantage reweighting. Experiments on the MMAD benchmark report an average improvement of 7.77 percentage points over the InternVL3-8B base model across seven tasks, along with comparisons to general-purpose MLLMs and to the GRPO-tuned AnomalyR1.

Significance. If the empirical claims are sustained, EMIT would be a useful contribution: it combines RL-based fine-tuning with visual contrastive embeddings for IAD, and the use of GPT-generated text to substitute for missing defective images is a practical idea that could help in data-scarce industrial settings. The release of code (GitHub link) is a strength. However, the current experimental evidence does not yet establish the central claim: the training protocol appears to include data from the evaluation benchmark's constituent datasets, the ablation isolating the proposed difficulty-aware mechanisms is missing, and all results come from single runs without variance estimates. These issues must be resolved before the claimed 7.77-point gain can be attributed to the method's intended components rather than to distributional overlap or noise.

major comments (4)
  1. [Dataset Preparation, Data Collection (Fig. 7, Table 2)] The Data Collection paragraph states that, because evaluation is based on MMAD, the authors 'carefully handle its subdatasets' by randomly selecting five normal images per object type from MVTec AD, MVTec LOCO AD, VisA, and GoodsAD and generating GPT text samples from them. Since MMAD is built from exactly these subdatasets, this is training on the evaluation family at the level of object types and defect types. The ablation 'w/o Text Samples' in Table 2 shows a 2.58-point average drop (81.95 to 79.37), so a substantial portion of the headline 7.77-point gain over InternVL3-8B may come from evaluation-family familiarity rather than from generalizable IAD skill. The stated precaution of using only five normal images per type does not prevent this contamination. Please provide a version trained entirely without MMAD-constituent data (e.g., text samples generated only from Vision, Real-IAD, and MPDD), report the overlap between the sampled normal images and MMAD reference images, and quantify how many object types and defect types in the text samples coincide with the evaluation set.
  2. [Ablation Study, Training Strategy (Table 2)] The core methodological claim is that difficulty-aware GRPO (response resampling and advantage reweighting) improves over plain GRPO. However, the comparison in Table 1 between 'InternVL3 (GRPO)' and 'EMIT' is confounded: EMIT also adds the soft prompt, contrastive embeddings, and GPT text samples. The ablation study in Table 2 does not include 'w/o response resampling' or 'w/o advantage reweighting' rows, nor an 'EMIT with standard GRPO' condition. Consequently, the reported 2.79-point difference over standard GRPO cannot be attributed to the proposed difficulty-aware mechanisms. Please add ablations that isolate these two components while holding all other EMIT components fixed.
  3. [Experimental Setup and Results (Tables 1–3)] All results in Tables 1–3 are single runs with no standard deviations, confidence intervals, or repeated seeds. Several margins are small (e.g., Table 3: 73.87 vs 73.63 for anomaly discrimination; 76.39 vs 76.07 for defect localization), and without variance estimates it is unclear whether these differences are meaningful. Please report mean and standard deviation over at least three seeds, or otherwise justify the stability of the single-run numbers.
  4. [Ablation Study (Table 1 vs Table 2)] The 'w/o Stage2' row in Table 2 (75.33 average) is described as removing Stage 2, i.e., keeping only Stage 1 SFT. Yet Table 1 lists 'InternVL3 (SFT)' at 76.92 average. If these correspond to the same configuration, the discrepancy needs explanation; if they differ (e.g., Table 1's SFT does not include the soft prompt and contrastive embeddings), the difference must be stated explicitly. As written, the ablation is ambiguous and undermines the interpretation of the two-stage training contribution.
minor comments (6)
  1. [Methodology, Eq. (1)] Equation (1) has formatting issues: the layer superscript l is not rendered clearly, and 'cosinedist' should read 'cosine distance'.
  2. [Difficulty-Aware GRPO, Eq. (5)] The text says the '+1' in the weight formula prevents the weight from becoming zero, but when all responses are correct (zero incorrect responses) the weight is 1, not 0; the explanation should be corrected to say that it preserves a baseline weight for easy questions.
  3. [Abstract and Introduction] The phrase 'average improvement of 7.77%' refers to a difference in accuracy percentage points (74.18 to 81.95), not a relative percentage improvement; please use 'percentage points' for precision.
  4. [Implementation Details] In the sentence 'We config open-source InternVL3 8B as our base MLLM', 'config' should be 'configure'.
  5. [Verifiable Rewards] The cosine reward is stated to encourage a shorter reasoning process for correct answers and a more detailed reasoning process for incorrect answers; this direction is counterintuitive and is not motivated in the text. Please clarify the intended effect on the policy.
  6. [Table 1] The GLM-4.1V-Thinking row has a formatting error ('93.7983.26' appears without a separator between the Defect Analysis and Object Classification columns).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are empirical benchmark results, and no prediction reduces by construction to the training inputs.

full rationale

The claimed 7.77% improvement over InternVL3-8B is an empirical benchmark outcome, not an analytic consequence of the method's definitions. The difficulty-aware GRPO (Eqs. 2-6) is a standard RL update: it uses ground-truth correctness as a verifiable reward, and its 'difficulty' weight w = #Incorrect/G + 1 is a training heuristic that reweights gradients; nothing in these equations defines the MMAD accuracy in terms of the training loss, so no fitted parameter is renamed as a prediction. The soft prompt and heatmap-guided contrastive embeddings are also empirically ablated (Table 3), not derived from the target metric. The one potentially load-bearing overlap is the Data Collection statement that five normal images per object type from MMAD's constituent subdatasets are used to generate GPT text training samples; if those samples leak evaluation-family appearance or object-class information, the reported gains and the 2.58-point text-sample ablation could be inflated. However, this is a benchmark-contamination and generalization risk, not a circularity: the evaluation numbers are not equal to or implied by the training data by construction. No uniqueness theorem, forced ansatz via self-citation, or definitional equivalence is present. The possible shared authorship with the InternVL3 citation only supplies the base model, and the paper's central claim is the improvement over that base, so it is not a load-bearing self-citation loop.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The framework's gains rest on several hand-chosen hyperparameters and domain assumptions; the text-description substitution and the resampling guarantee are the two most consequential assumptions. No new physical entities are introduced.

free parameters (7)
  • Number of sampled responses G = 8
    Set for GRPO; no sensitivity analysis is reported.
  • KL divergence coefficient beta = 0.01
    Chosen for stage-2 training; affects training stability and final accuracy.
  • Reward weights = classification 3, all others 1
    Hand-chosen; no ablation of the weighting is given, yet this directly controls the learning signal.
  • Advantage reweight denominator = G+1, via the added 1 in the weight formula
    The +1 prevents zero weight for easy questions but is arbitrary; it changes gradient scale for easy and hard questions.
  • Local range k for patch comparison = not specified
    Eq. (1) depends on k, but its value and the set of ViT layers used for the heatmaps are not reported, blocking exact reproduction.
  • LoRA rank and alpha = rank 8, alpha 32
    Set in implementation details; no sensitivity analysis is provided.
  • Learning rate and epochs = 1e-5, two epochs stage 1, one epoch stage 2
    Hand-chosen and not swept; the central result depends on them.
assumptions (5)
  • domain assumption The text-image semantic alignment of MLLMs is strong enough that GPT-generated object descriptive text can substitute for missing defective images as training input.
    Section 'Object Descriptive Text as Query Image' relies on this to replace absent defective images with text; the paper provides no direct measurement of how much text-based training transfers to visual queries.
  • domain assumption A patch-wise minimal cosine distance between query and reference features within a local window is a valid signal for defect presence and location.
    Eq. (1) defines the heatmap; the paper assumes these distances align with true defects, with no calibration or failure analysis.
  • domain assumption Resampling until a correct answer appears will eventually produce a correct answer and is a valid group-based RL signal.
    Section 'Response Resampling' states this ensures correct answers, but no termination condition or proof is given; for extremely hard samples resampling may never succeed.
  • domain assumption Verifiable rewards (format, classification, cosine, repetition) collectively incentivize correct and well-formatted IAD answers.
    Section 'Verifiable Rewards' assigns weights by hand; no sensitivity analysis or validation that these rewards align with downstream IAD quality.
  • domain assumption Training on Vision, Real-IAD, and MPDD generalizes to MMAD's object and defect distribution.
    Section 'Data Collection' uses these public datasets for the four training tasks; the paper does not measure domain shift beyond the final MMAD accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EMIT: Enhancing MLLMs for Industrial Anomaly Detection via Difficulty-Aware GRPO." pith.science (2026). https://pith.science/paper/PV5ET5YF

@misc{pith2026250721619,
  author       = {Pith},
  title        = {Pith review of: EMIT: Enhancing MLLMs for Industrial Anomaly Detection via Difficulty-Aware GRPO},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PV5ET5YF}},
  note         = {Machine review of arXiv:2507.21619}
}
read the original abstract

Industrial anomaly detection (IAD) plays a crucial role in maintaining the safety and reliability of manufacturing systems. While multimodal large language models (MLLMs) show strong vision-language reasoning abilities, their effectiveness in IAD remains limited without domain-specific adaptation. In this work, we propose EMIT, a unified framework that enhances MLLMs for IAD via difficulty-aware group relative policy optimization (GRPO). EMIT constructs a multi-task IAD dataset and utilizes GPT-generated object text descriptions to compensate for missing defective images. For few-shot anomaly detection, it integrates a soft prompt and heatmap-guided contrastive embeddings derived from patch-level comparisons. To better handle difficult data samples, i.e., cases where the MLLM struggles to generate correct answers, we propose a difficulty-aware GRPO that extends the original GRPO by incorporating a response resampling strategy to ensure the inclusion of correct answers in the sampled responses, as well as an advantage reweighting mechanism to strengthen learning from such difficult data samples. Extensive experiments on the MMAD benchmark demonstrate that EMIT significantly enhances the IAD performance of MLLMs, achieving an average improvement of 7.77\% over the base model (InternVL3-8B) across seven tasks.

Figures

Figures reproduced from arXiv: 2507.21619 by the authors.

Figure 1
Figure 1. Comparison of models fine-tuned using diverse [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The architecture of EMIT which can integrate components such as the query image, text prompt, domain knowledge, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The process for producing contrastive embeddings. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Examples of four tasks from our training dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The prompt for constructing domain knowledge. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 7
Figure 7. Figure 7: Statistical overview of the training dataset. The [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: The framework of difficulty-aware GRPO for IAD. [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Global Logic and Local Search: Dual-Stream Multimodal In-Context Learning for Verifiable Industrial Anomaly Detection

    cs.CV 2026-07 conditional novelty 6.0 of 10

    A training-free dual-stream multimodal framework (PVLA + SAM 3 global logic + MCTS local search) improves verifiable industrial anomaly QA without defective training samples.

Reference graph

Works this paper leans on

56 extracted references · 22 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al

    Achiam, J.; Adler, S.; Agarwal, S.; Ahmad, L.; Akkaya, I.; Aleman, F. L.; Almeida, D.; Altenschmidt, J.; Altman, S.; Anadkat, S.; et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    B.; Bout, B.; Chaplot, D.; Chudnovsky, J.; Costa, D.; Monicault, B

    Agrawal, P.; Antoniak, S.; Hanna, E. B.; Bout, B.; Chaplot, D.; Chudnovsky, J.; Costa, D.; Monicault, B. D.; Garg, S.; Gervet, T.; Ghosh, S.; Héliou, A.; Jacob, P.; Jiang, A. Q.; Khandelwal, K.; Lacroix, T.; Lample, G.; Casas, D. L.; Lavril, T.; Scao, T. L.; Lo, A.; Marshall, W.; Martin, L.; Mensch, A.; Muddireddy, P.; Nemychnikova, V.; Pellat, M.; Platen...

    arXiv 2024

  5. [5]

    G.; Tuzel, O.; Huang, P.; Shan, J.; Shi, J.; and Cao, M

    Bai, H.; Mou, S.; Likhomanenko, T.; Cinbis, R. G.; Tuzel, O.; Huang, P.; Shan, J.; Shi, J.; and Cao, M. 2023. Vision datasets: A benchmark for vision-based industrial inspection. arXiv preprint arXiv:2306.07890

    arXiv 2023

  6. [6]

    Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y.; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y.; Ye, J.; Zhang, X.; Xie, T.; Cheng, Z.; Zhang, H.; Yang, Z.; Xu, H.; and Lin, J. 2025. Qwen2.5-VL Technical Report. arXiv:2502.13923

    arXiv 2025

  7. [7]

    Bergmann, P.; Batzner, K.; Fauser, M.; Sattlegger, D.; and Steger, C. 2022. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision, 130(4): 947--969

    2022

  8. [8]

    Bergmann, P.; Fauser, M.; Sattlegger, D.; and Steger, C. 2019. MVTec AD--A comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9592--9600

    2019

Show all 56 references
  1. [9]

    Cao, Y.; Zhang, J.; Frittoli, L.; Cheng, Y.; Shen, W.; and Boracchi, G. 2024. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly detection. In European Conference on Computer Vision, 55--72. Springer

    2024

  2. [10]

    Chao, Y.; Liu, J.; Tang, J.; and Wu, G. 2025. Anomalyr1: A grpo-based end-to-end mllm for industrial anomaly detection. arXiv preprint arXiv:2504.11914

    2025 arXiv

  3. [11]

    Chen, H.; Tu, H.; Wang, F.; Liu, H.; Tang, X.; Du, X.; Zhou, Y.; and Xie, C. 2025 a . Sft or rl? an early investigation into training r1-like reasoning large vision-language models. arXiv preprint arXiv:2504.11468

    2025 arXiv

  4. [12]

    Chen, Z.; Chen, H.; Imani, M.; and Imani, F. 2025 b . Can multimodal large language models be guided to improve industrial anomaly detection? arXiv preprint arXiv:2501.15795

    2025 arXiv

  5. [13]

    Deng, H.; Luo, H.; Zhai, W.; Cao, Y.; and Kang, Y. 2024. Vmad: Visual-enhanced multimodal large language model for zero-shot anomaly detection. arXiv preprint arXiv:2409.20146

    2024

  6. [14]

    Dettmers, T.; Pagnoni, A.; Holtzman, A.; and Zettlemoyer, L. 2024. Qlora: Efficient finetuning of quantized llms. Advances in Neural Information Processing Systems, 36

    2024

  7. [15]

    Fan, L.; Huang, J.; Di, D.; Su, A.; Pagnucco, M.; and Song, Y. 2024. Revitalizing Reconstruction Models for Multi-class Anomaly Detection via Class-Aware Contrastive Learning. arXiv preprint arXiv:2412.04769

    2024 arXiv

  8. [16]

    Fang, Z.; Wang, X.; Li, H.; Liu, J.; Hu, Q.; and Xiao, J. 2023. Fastrecon: Few-shot industrial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 17481--17490

    2023

  9. [17]

    Gu, Z.; Zhu, B.; Zhu, G.; Chen, Y.; Tang, M.; and Wang, J. 2024. Anomalygpt: Detecting industrial anomalies using large vision-language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 1932--1940

    2024

  10. [18]

    Guo, J.; Lu, S.; Zhang, W.; Chen, F.; Li, H.; and Liao, H. 2025. Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. In Proceedings of the Computer Vision and Pattern Recognition Conference, 20405--20415

    2025

  11. [19]

    X.; Nguyen, A.-N.; Tran, D.-T.; Duong, V.-H.; Mai, A.-T.; Pham, D.-L.; Phan, K.-T.; Do, M.-Q.; Duong, T

    Hoang, D.-C.; Tan, P. X.; Nguyen, A.-N.; Tran, D.-T.; Duong, V.-H.; Mai, A.-T.; Pham, D.-L.; Phan, K.-T.; Do, M.-Q.; Duong, T. H. A.; et al. 2025. Unsupervised visual-to-geometric feature reconstruction for vision-based industrial anomaly detection. IEEE Access

    2025

  12. [20]

    Hu, H.; Wang, X.; Fan, J.; Zeng, Z.; Lu, J.; Hong, O.; and Zhang, J. 2025. DSMBAD: Dual-Stream Memory Bank Framework for Unified Industrial Anomaly Detection. Electronics, 14(14): 2748

    2025

  13. [21]

    H.; Bae, K.; and Kang, B

    Hyun, J.; Kim, S.; Jeon, G.; Kim, S. H.; Bae, K.; and Kang, B. J. 2024. Reconpatch: Contrastive patch representation learning for industrial anomaly detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2052--2061

    2024

  14. [22]

    Jeong, J.; Zou, Y.; Kim, T.; Zhang, D.; Ravichandran, A.; and Dabeer, O. 2023. Winclip: Zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19606--19616

    2023

  15. [23]

    Jezek, S.; Jonak, M.; Burget, R.; Dvorak, P.; and Skotak, M. 2021. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (I...

    2021

  16. [24]

    Jiang, X.; Li, J.; Deng, H.; Liu, Y.; Gao, B.-B.; Zhou, Y.; Li, J.; Wang, C.; and Zheng, F. 2024 a . Mmad: A comprehensive benchmark for multimodal large language models in industrial anomaly detection. arXiv preprint arXiv:2410.09453

    2024 arXiv

  17. [25]

    Jiang, X.; Liu, J.; Wang, J.; Nie, Q.; Wu, K.; Liu, Y.; Wang, C.; and Zheng, F. 2022. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35: 15433--15445

    2022

  18. [26]

    Jiang, Y.; Lu, X.; Jin, Q.; Sun, Q.; Wu, H.; and Zhuo, C. 2024 b . Fabgpt: An efficient large multimodal model for complex wafer defect knowledge queries. In Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design, 1--8

    2024

  19. [27]

    Jin, E.; Feng, Q.; Mou, Y.; Lakemeyer, G.; Decker, S.; Simons, O.; and Stegmaier, J. 2025. Logicad: Explainable anomaly detection via vlm-based text feature extraction. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 4129--4137

    2025

  20. [28]

    Kim, D.; Park, C.; Cho, S.; Lim, H.; Kang, M.; Lee, J.; and Lee, S. 2025. GenCLIP: Generalizing CLIP Prompts for Zero-shot Anomaly Detection. arXiv preprint arXiv:2504.14919

    2025

  21. [29]

    Li, B.; Zhang, Y.; Guo, D.; Zhang, R.; Li, F.; Zhang, H.; Zhang, K.; Zhang, P.; Li, Y.; Liu, Z.; and Li, C. 2024 a . LLaVA-OneVision: Easy Visual Task Transfer. arXiv:2408.03326

    2024 arXiv

  22. [30]

    Li, F.; Zhang, R.; Zhang, H.; Zhang, Y.; Li, B.; Li, W.; Ma, Z.; and Li, C. 2024 b . LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models. arXiv:2407.07895

    2024 arXiv

  23. [31]

    Li, M.; He, J.; Ying, Z.; Li, G.; and Zhou, M. 2024 c . MemADet: A Representative Memory Bank Approach for Industrial Image Anomaly Detection. In 2024 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 2261--2266. IEEE

    2024

  24. [32]

    Li, W.; Chu, G.; Chen, J.; Xie, G.-S.; Shan, C.; and Zhao, F. 2025. Lad-reasoner: Tiny multimodal models are good reasoners for logical anomaly detection. arXiv preprint arXiv:2504.12749

    2025 arXiv

  25. [33]

    Li, Y.; Wang, H.; Yuan, S.; Liu, M.; Zhao, D.; Guo, Y.; Xu, C.; Shi, G.; and Zuo, W. 2023. Myriad: Large multimodal model by applying vision experts for industrial anomaly detection. arXiv preprint arXiv:2310.19070

    2023 arXiv

  26. [34]

    Loshchilov, I. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101

    2017 arXiv

  27. [35]

    Mokhtar, S.; Mousakhan, A.; Galesso, S.; Tayyub, J.; and Brox, T. 2025. Detect, Classify, Act: Categorizing Industrial Anomalies with Multi-Modal Large Language Models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 4058--4067

    2025

  28. [36]

    Roth, K.; Pemula, L.; Zepeda, J.; Sch \"o lkopf, B.; Brox, T.; and Gehler, P. 2022. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 14318--14328

    2022

  29. [37]

    Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

    2017 arXiv

  30. [38]

    Shao, Z.; Wang, P.; Zhu, Q.; Xu, R.; Song, J.; Bi, X.; Zhang, H.; Zhang, M.; Li, Y.; Wu, Y.; et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  31. [39]

    M.; Goedeckemeyer, A.; Saade, A.; Feng, A.; Kolesnikov, A.; Bendebury, A.; Abdagic, A.; Vadi, A.; György, A.; Pinto, A

    Team, G.; Kamath, A.; Ferret, J.; Pathak, S.; Vieillard, N.; Merhej, R.; Perrin, S.; Matejovicova, T.; Ramé, A.; Rivière, M.; Rouillard, L.; Mesnard, T.; Cideron, G.; bastien Grill, J.; Ramos, S.; Yvinec, E.; Casbon, M.; Pot, E.; Penchev, I.; Liu, G.; Visin, F.; Kenealy, K.; B...

    2025 arXiv

  32. [40]

    Team, K.; Du, A.; Yin, B.; Xing, B.; Qu, B.; Wang, B.; Chen, C.; Zhang, C.; Du, C.; Wei, C.; Wang, C.; Zhang, D.; Du, D.; Wang, D.; Yuan, E.; Lu, E.; Li, F.; Sung, F.; Wei, G.; Lai, G.; Zhu, H.; Ding, H.; Hu, H.; Yang, H.; Zhang, H.; Wu, H.; Yao, H.; Lu, H.; Wang, H.; Gao, H.;...

    2025 arXiv

  33. [41]

    Team, V.; Hong, W.; Yu, W.; Gu, X.; Wang, G.; Gan, G.; Tang, H.; Cheng, J.; Qi, J.; Ji, J.; Pan, L.; Duan, S.; Wang, W.; Wang, Y.; Cheng, Y.; He, Z.; Su, Z.; Yang, Z.; Pan, Z.; Zeng, A.; Wang, B.; Shi, B.; Pang, C.; Zhang, C.; Yin, D.; Yang, F.; Chen, G.; Xu, J.; Chen, J.; Che...

    2025 arXiv

  34. [42]

    Wang, C.; Zhu, W.; Gao, B.-B.; Gan, Z.; Zhang, J.; Gu, Z.; Qian, S.; Chen, M.; and Ma, L. 2024. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2024

  35. [43]

    Xiaomi, L.-C.-T. 2025. MiMo-VL Technical Report. arXiv:2506.03569

    2025 arXiv

  36. [44]

    M.; and Dwivedi, I

    Xu, J.; Lo, S.-Y.; Safaei, B.; Patel, V. M.; and Dwivedi, I. 2025. Towards zero-shot anomaly detection and reasoning with multimodal large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, 20370--20382

    2025

  37. [45]

    Yeo, E.; Tong, Y.; Niu, M.; Neubig, G.; and Yue, X. 2025. Demystifying Long Chain-of-Thought Reasoning in LLMs. arXiv:2502.03373

    2025 arXiv

  38. [46]

    Yu, Q.; Zhang, Z.; Zhu, R.; Yuan, Y.; Zuo, X.; Yue, Y.; Dai, W.; Fan, T.; Liu, G.; Liu, L.; et al. 2025. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476

    2025 arXiv

  39. [47]

    Yuan, J.; Jie, P.; Zhang, J.; Li, Z.; and Gao, C. 2025. MFP-CLIP: Exploring the Efficacy of Multi-Form Prompts for Zero-Shot Industrial Anomaly Detection. arXiv preprint arXiv:2503.12910

    2025 arXiv

  40. [48]

    Zeng, P.; Pang, F.; Wang, Z.; and Yang, A. 2025. LR-IAD: Mask-Free Industrial Anomaly Detection with Logical Reasoning. arXiv preprint arXiv:2504.19524

    2025 arXiv

  41. [49]

    Zhang, J.; Ding, R.; Ban, M.; and Dai, L. 2024 a . PKU-GoodsAD: A supermarket goods dataset for unsupervised anomaly detection and segmentation. IEEE Robotics and Automation Letters, 9(3): 2008--2015

    2024

  42. [50]

    Zhang, X.; Xu, M.; and Zhou, X. 2024. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16699--16708

    2024

  43. [51]

    Zhang, Y.; Cao, Y.; Xu, X.; and Shen, W. 2024 b . Logicode: an llm-driven framework for logical anomaly detection. IEEE Transactions on Automation Science and Engineering

    2024

  44. [52]

    Zhang, Z.; Ruan, J.; Gao, X.; Liu, T.; and Fu, Y. 2025. Eiad: Explainable industrial anomaly detection via multi-modal large language models. arXiv preprint arXiv:2503.14162

    2025 arXiv

  45. [53]

    Zhao, S.; Lin, Y.; Han, L.; Zhao, Y.; and Wei, Y. 2025. OmniAD: Detect and Understand Industrial Anomaly via Multimodal Reasoning. arXiv preprint arXiv:2505.22039

    2025 arXiv

  46. [54]

    Zhou, Q.; Pang, G.; Tian, Y.; He, S.; and Chen, J. 2023. Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961

    2023

  47. [55]

    Zhu, J.; Wang, W.; Chen, Z.; Liu, Z.; Ye, S.; Gu, L.; Tian, H.; Duan, Y.; Su, W.; Shao, J.; Gao, Z.; Cui, E.; Wang, X.; Cao, Y.; Liu, Y.; Wei, X.; Zhang, H.; Wang, H.; Xu, W.; Li, H.; Wang, J.; Deng, N.; Li, S.; He, Y.; Jiang, T.; Luo, J.; Wang, Y.; He, C.; Shi, B.; Zhang, X.;...

    2025 arXiv

  48. [56]

    Zou, Y.; Jeong, J.; Pemula, L.; Zhang, D.; and Dabeer, O. 2022. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In European conference on computer vision, 392--408. Springer

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.