Pith. sign in

REVIEW 4 major objections 5 minor 3 cited by

AD-FM: Multimodal LLMs for Anomaly Detection via Multi-Stage Reasoning and Fine-Grained Reward Optimization

T0 review · 4 major / 5 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 This paper argues that GRPO-style reinforcement fine-tuning of multimodal LLMs for anomaly detection fails under binary rewards — and that forcing a three-stage reasoning protocol with localization-aware continuous rewards fixes it, yieldin

desk verdict A solid, well-ablated extension of GRPO fine-tuning for anomaly detection; the main weakness is unstated box parsing and missing error bars, not the core method. read the letter →

arxiv 2508.04175 v1 pith:IT2T7BOS submitted 2025-08-06 cs.CV

Jingyi Liao , Yongyi Su , Rong-Cheng Tu , Zhao Jin , Wenhao Sun , Yiting Li , Dacheng Tao , Xun Xu
show 1 more author
Xulei Yang
This is my paper
classification cs.CV
keywords anomalydetectionmultimodallargelanguagemodelsGRPOreinforcementfine-tuningfine-grainedrewardlocalizationmulti-stagereasoningindustrialdefect
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

The paper tries to establish that the weak spot in adapting general vision-language models to industrial anomaly detection is not model capacity but reward design inside GRPO: binary correct/incorrect feedback goes flat whenever all sampled responses give the same answer, killing the gradient signal. Its fix is to make the model first localize suspicious regions, then examine them, then decide — and to reward not just the final answer but also how many regions were found and how well the predicted boxes overlap the true defect areas. The authors report that this fine-grained, spatially grounded reward lifts average accuracy from 51.04% zero-shot to 73.15% across four industrial datasets, a gain of 22.11 points, and that the improvement is substantially larger than adding random reward noise. If correct, the method gives a practical recipe for squeezing usable anomaly detectors out of general MLLMs with only a fraction of the label budget.

What carries the argument

The load-bearing object is the fine-grained reward function $r_{acc} = r_{loc} + r_{cls}$, where $r_{loc}$ is computed from predicted and pseudo ground-truth bounding boxes via Hungarian matching, combining a box-count reward ($r_{count}$) and an average-GIoU reward, and $r_{cls}$ is a classification-accuracy term. Its role is to convert GRPO's binary reward surface into a continuous landscape that differentiates responses of varying analytical quality, restoring gradient signal when all sampled answers agree and filtering out 'spurious correctness' — responses that get the class right but point at the wrong region. The multi-stage reasoning protocol (think → rethink → answer) is what suppli

What would settle it

Take the MVTec training subset, hand-draw accurate bounding boxes for the same defect masks used to build pseudo ground-truth boxes, and rerun the full GRPO recipe with those hand-drawn boxes as supervision. If the +3.33% gain over the binary multi-stage baseline disappears or reverses, the benefit depends on the pseudo-label construction rather than on rewarding spatial reasoning itself. Alternatively, parse the think-stage outputs of a trained model on 100 test images and compare the parsed boxes to human-annotated boxes for the same images; a high mismatch rate would show the localization r

Watch

Extended reading notes

Core claim

The central claim is that the failure of GRPO-based anomaly-detection adaptation is caused by two coupled deficiencies — response/reward variance collapse and unconstrained reasoning — and that both are cured by the same design choice: make the model emit spatial predictions during a structured think-rethink-answer protocol, then grade those predictions with a continuous reward. Concretely, the model is prompted to output bounding boxes of suspicious regions in the first stage; the reward $r_{acc} = r_{loc} + r_{cls}$ combines a classification term with a localization term built from Hungarian-matched box pairs, using a count-accuracy measure and average GIoU against pseudo ground-truth boxe

Load-bearing premise

The fine-grained reward is only as good as the pseudo ground-truth boxes built from pixel masks via dilation and connected components, and as reliable as the parser that extracts predicted boxes from free-form reasoning text; if either is noisy, the reward is miscalibrated and the reported gains could shrink or vanish.

Editorial extensions

If this is right

  • Existing pixel-level anomaly masks can be recycled as pseudo-box supervision, so the method adapts MLLMs without requiring fresh box annotations.
  • Because the reward rewards localization quality, responses that identify the right defect count and location are preferentially reinforced, not just those that guess the binary label correctly.
  • The method's cross-dataset generalization results suggest that GRPO fine-tuning on one domain (MVTec) transfers to unseen defect types, which is the practical condition for factory deployment.
  • The zero-variance ablation (41.25% → 11.00%) indicates that the fine-grained reward converts GRPO from a sparse-signal process into one that learns from nearly every sampled response.
  • Ablations against random reward perturbation indicate the gain comes from the spatial meaning of the reward, not merely from injecting variance.

Reading between the lines

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

  • The same recipe — force the model to localize before it classifies, then reward the localization — could transfer to other fine-grained visual decision tasks (medical imaging, satellite inspection, document layout defects) where binary labels are cheap but spatial correctness matters.
  • The pseudo ground-truth box construction (dilation + connected components) is an unspoken hyperparameter; the reported gains could depend on how well those boxes match human-perceived defect extents, which suggests a stress test on datasets with fragmented masks.
  • The approach implicitly assumes that free-form text parsing of the predicted boxes is reliable; if a readable structured output format (e.g., JSON) replaced free-form <think> text, the reward would be more robust and the gains more reproducible.
  • The reported 73.15% still leaves substantial room on harder datasets like GoodsAD (56.74%), so the method improves but does not close the gap to specialized AD algorithms on the most difficult categories.
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 / 5 minor

Summary. The paper proposes AD-FM, a GRPO-based post-training framework for adapting a general MLLM (Qwen2.5-VL-7B) to industrial anomaly detection. The method has two main components. First, it replaces single-step prompting with a three-stage deliberation protocol (Think, Rethink, Answer) intended to increase response diversity and to create intermediate signals for supervision. Second, it defines a fine-grained reward that combines classification accuracy with localization terms, namely a count reward and an average-GIoU reward computed between predicted boxes and pseudo ground-truth boxes derived from pixel masks. Experiments are reported on a multi-domain setting (MVTec, VisA, MVTec-LOCO, GoodsAD) and a cross-dataset setting (trained on MVTec, tested on VisA, MPDD, DTD, DAGM). The main result is a multi-domain average accuracy of 73.15%, a +22.11% improvement over the zero-shot base model, with an ablation showing +3.33% from replacing a random reward perturbation with the localization-aware reward. The paper also reports a visualization of response-reward variance and a zero-variance analysis supporting the claim that the proposed reward reduces gradient collapse.

Significance. If the reported numbers are reproducible, the contribution is practically useful: it combines a simple reasoning format change with a continuous reward to address a known limitation of GRPO in binary decision tasks, and the empirical trajectory in Table 4 (68.07 -> 69.82 -> 70.22 -> 73.15) is coherent and supportive of the central claim. The paper also contains a falsifiable quantitative claim about reward variance (Fig. 5), which is a useful diagnostic. However, the current manuscript does not provide enough implementation detail to verify the mechanism: the parser used to extract predicted boxes from free-form reasoning text is not specified, the pseudo-box generation parameters are not given, and no error bars or repeated-seed results are reported. These omissions are load-bearing because the localization reward, the key contribution, is computed entirely from those two unstated processes. The comparisons against AnomalyR1 and OmniAD are additionally based on published averages without a controlled protocol, so the claimed advantage over concurrent methods is not yet established.

major comments (4)
  1. [Sec. 3.3 / Table 4] The +3.33% improvement attributed to the localization reward r_loc rests on predicted boxes extracted from free-form <think> text. The manuscript never states how this extraction is implemented (e.g., regex, an LLM-based parser, or a structured output head), how unparseable outputs are treated, or what the parsing success/failure rate is. If parsing is noisy, r_loc becomes a partially zero or miscalibrated signal and the 70.22 -> 73.15 gap could be driven by variance injection rather than by genuine localization supervision. Please specify the parser, report its per-dataset success rate, and give the reward value assigned when parsing fails.
  2. [Sec. 4.1 / Eqs. (4)-(6)] The pseudo ground-truth boxes are described only as 'morphological dilation ... then connected component analysis'. The dilation kernel size, iteration count, and connected-component threshold are not reported. These parameters directly determine the number and shape of pseudo boxes, which in turn determines both r_count and the GIoU denominator in r_loc. Different reasonable choices could make the count reward trivially easy or impossible and could change the ablation result. Please report the exact parameter values, and ideally a sensitivity study over them, to establish that the reward measures what the paper claims.
  3. [Tables 1-4 / Sec. 4.5] All results are reported as single-point accuracies with no error bars, no number of random seeds, and no statistical significance test. The central ablation comparison 70.22 (r_cls+r_random) vs 73.15 (r_cls+r_loc) is a difference of 2.93 points with no uncertainty estimate; given the dataset sizes and the stochasticity of GRPO with LoRA, this difference may not be significant. Please run at least three seeds for the main tables and the ablation, and report mean +/- std. Also state the standard deviation of the random reward perturbation (cited as RRP) used in the r_random row; without it, the 'variance injection' baseline is underspecified.
  4. [Tables 1 and 3] The comparisons with AnomalyR1 and OmniAD use only the average numbers from their publications. Since the training data fraction, base backbone, prompt format, and evaluation protocol may differ from those used here, the reported superiority (12.53% and 4.35%) is not a controlled comparison. If the official code and checkpoints are unavailable, the paper should either reproduce those methods under the same protocol or explicitly list the protocol differences and frame the comparison as indicative rather than head-to-head. This is important because Table 3 also claims per-task superiority over OmniAD on categories where the aggregate differences are small.
minor comments (5)
  1. [Eqs. (4)-(7)] The equations are poorly rendered in the arXiv text (e.g., missing symbols, garbled subscripts). Please rewrite them with clear notation and define every term, including the matching procedure in Eq. (4) and the meaning of the constraint in Eq. (6).
  2. [Fig. 5] The y-axis label and the numbers in the figure are corrupted in the provided rendering. Since the zero-variance percentage is one of the paper's key empirical supports, the figure must be readable and the exact computation (e.g., variance across six responses per sample) should be stated in the caption.
  3. [Fig. 3] The qualitative examples contain corrupted Unicode and missing text (e.g., 'in �����stage' and 'in �������stage'). Please fix the encoding and ensure the quoted reasoning excerpts are legible.
  4. [Sec. 4.1 / Supplements] The paper repeatedly refers to the supplementary material for details (CLIP threshold protocol, pseudo-box generation, multi-task protocols), but no supplementary PDF or repository link is provided. Please either include the supplementary material or move the essential details into the main text.
  5. [Table 2] For CLIP baselines, both a fixed threshold (0.5) and a Youden-optimized threshold are reported. The paper acknowledges the test-set tuning issue, but it would be clearer to also report the standard 'no tuning' protocol for the proposed method and the CLIP methods without threshold optimization in the same table to make the deployment claim fully transparent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: fine-grained rewards are training objectives evaluated on held-out data, and no self-citation carries the derivation.

full rationale

The paper's central claims are empirical: multi-stage deliberation and fine-grained localization rewards improve GRPO fine-tuning of Qwen2.5-VL-7B on anomaly detection. The derivation chain is: sample 20% of each dataset, construct pseudo bounding boxes from pixel masks (Sec. 4.1), define rewards r_cls/r_loc from labels and localization (Sec. 3.3), train with GRPO, and measure binary classification accuracy on the held-out 80% (Tabs. 1-2) and on unseen datasets (Tab. 2). There is no fitted quantity that is relabeled as a prediction: the model is evaluated on data not used for training, and the CLIP threshold optimization (Youden's J) applies only to the baselines, making the comparison conservative. The rewards are supervised training signals based on the same annotations as the test metric; that is standard RL fine-tuning, not circularity. Citations to prior work by the authors (Cao et al. 2023/2025, Li et al. 2025b) appear only in background/related-work descriptions and are not load-bearing for the claimed +3.33% or +22.11% improvements. Reproducibility concerns about parsing predicted boxes from free-form text and pseudo-box generation hyperparameters are correctness risks, not self-reference; they do not make the derivation equivalent to its inputs.

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

The method's main assumptions are the validity of pseudo-boxes from masks, the parseability of model-generated boxes, and the hypothesis that localization quality is a proxy for genuine visual understanding. The GRPO variance collapse premise is standard RL math. No new physical or conceptual entities are postulated.

free parameters (5)
  • Reward weighting between localization and classification = 1:1 (implicit sum in Eq. 7)
    No ablation or tuning reported for the relative weight of r_loc and r_cls; the paper implicitly fixes both weights to 1.
  • Pseudo-box generation parameters (dilation kernel, connected component thresholds)
    Not specified in the paper; these directly determine the pseudo ground-truth boxes used in the localization reward.
  • GRPO KL regularization coefficient
    Not stated; standard GRPO hyperparameter that controls policy update strength and stability.
  • LoRA rank and alpha = rank 64, alpha 128
    Given in Sec 4.1 but not swept; standard LoRA hyperparameters with no sensitivity analysis.
  • Number of responses per prompt and epochs = 6 responses, 15 epochs
    Fixed choices; no ablation on these GRPO training settings.
assumptions (5)
  • standard math Uniform rewards in GRPO produce zero advantages and therefore no learning signal
    Follows from z-score normalization in GRPO; equation (1) and Sec 3.1.
  • domain assumption Pseudo bounding boxes derived from pixel masks are a valid proxy for anomalous regions
    Sec 4.1 Bounding Box Generation; if dilation or merging introduces errors, localization rewards are miscalibrated.
  • ad hoc to paper Localization quality correlates with genuine analytical insight, so downweighting correct but poorly localized responses improves classification
    Core hypothesis of the fine-grained reward design in Sec 3.3; not independently validated beyond the reported accuracy gains.
  • domain assumption The model reliably follows the multi-stage format and predicted boxes can be extracted from free-form text
    Sec 3.3 and Fig. 1; extraction method is not described.
  • domain assumption A 20% uniform sample per category is a representative training split for each dataset
    Sec 4.1 Datasets and Protocols; standard practice but a choice that affects both training and evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AD-FM: Multimodal LLMs for Anomaly Detection via Multi-Stage Reasoning and Fine-Grained Reward Optimization." pith.science (2026). https://pith.science/paper/IT2T7BOS

@misc{pith2026250804175,
  author       = {Pith},
  title        = {Pith review of: AD-FM: Multimodal LLMs for Anomaly Detection via Multi-Stage Reasoning and Fine-Grained Reward Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IT2T7BOS}},
  note         = {Machine review of arXiv:2508.04175}
}
read the original abstract

While Multimodal Large Language Models (MLLMs) demonstrate remarkable capabilities across diverse domains, their application to specialized anomaly detection (AD) remains constrained by domain adaptation challenges. Existing Group Relative Policy Optimization (GRPO) based approaches suffer from two critical limitations: inadequate training data utilization when models produce uniform responses, and insufficient supervision over reasoning processes that encourage immediate binary decisions without deliberative analysis. We propose a comprehensive framework addressing these limitations through two synergistic innovations. First, we introduce a multi-stage deliberative reasoning process that guides models from region identification to focused examination, generating diverse response patterns essential for GRPO optimization while enabling structured supervision over analytical workflows. Second, we develop a fine-grained reward mechanism incorporating classification accuracy and localization supervision, transforming binary feedback into continuous signals that distinguish genuine analytical insight from spurious correctness. Comprehensive evaluation across multiple industrial datasets demonstrates substantial performance improvements in adapting general vision-language models to specialized anomaly detection. Our method achieves superior accuracy with efficient adaptation of existing annotations, effectively bridging the gap between general-purpose MLLM capabilities and the fine-grained visual discrimination required for detecting subtle manufacturing defects and structural irregularities.

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  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.

  2. AD-Copilot: A Vision-Language Assistant for Industrial Anomaly Detection via Visual In-context Comparison

    cs.CV 2026-03 conditional novelty 6.0 of 10

    AD-Copilot trains an MLLM on a new curated industrial dataset Chat-AD with a Comparison Encoder that uses cross-attention on image pairs, reaching 82.3% accuracy on MMAD and 3.35x gains on MMAD-BBox while generalizing...

  3. AgentIAD: Agentic Industrial Anomaly Detection via Adaptive Memory Augmentation

    cs.CV 2025-12 unverdicted novelty 6.0 of 10

    AgentIAD introduces an agentic VLM with Perceptive Zoomer, Web Searcher, and Comparative Retriever tools plus two-stage SFT-then-RL training, achieving 5.92% higher classification accuracy than prior SOTA on the MMAD ...

Reference graph

Works this paper leans on

48 extracted references · 21 canonical work pages · cited by 3 Pith papers

  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]

    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

    work page 2022

  4. [4]

    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

    work page 2019

  5. [5]

    Cao, Y.; Xu, X.; Cheng, Y.; Sun, C.; Du, Z.; Gao, L.; and Shen, W. 2025. Personalizing Vision-Language Models With Hybrid Prompts for Zero-Shot Anomaly Detection. IEEE Transactions on Cybernetics, 55(4): 1917–1929

    work page 2025

  6. [6]

    Cao, Y.; Xu, X.; Sun, C.; Huang, X.; and Shen, W. 2023. Towards Generic Anomaly Detection and Understanding: Large-scale Visual-linguistic Model (GPT-4V) Takes the Lead. arXiv:2311.02782

    arXiv 2023

  7. [7]

    Chao, Y.; Liu, J.; Tang, J.; and Wu, G. 2025. AnomalyR1: A GRPO-based End-to-end MLLM for Industrial Anomaly Detection. arXiv:2504.11914

    arXiv 2025

  8. [8]

    Chen, L.; Li, L.; Zhao, H.; Song, Y.; and Vinci. 2025 a . R1-V: Reinforcing Super Generalization Ability in Vision-Language Models with Less Than \ 3. https://github.com/Deep-Agent/R1-V. Accessed: 2025-02-02

    work page 2025

Show all 48 references
  1. [9]

    Chen, Z.; Chen, H.; Imani, M.; and Imani, F. 2025 b . Can Multimodal Large Language Models be Guided to Improve Industrial Anomaly Detection? arXiv:2501.15795

    2025 arXiv

  2. [10]

    Chen, Z.; Wang, W.; Tian, H.; Ye, S.; Gao, Z.; Cui, E.; Tong, W.; Hu, K.; Luo, J.; Ma, Z.; et al. 2024. How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites. arXiv preprint arXiv:2404.16821

    2024 arXiv

  3. [11]

    Cimpoi, M.; Maji, S.; Kokkinos, I.; Mohamed, S.; ; and Vedaldi, A. 2014. Describing Textures in the Wild. In Proceedings of the IEEE Conf. on Computer Vision and Pattern Recognition ( CVPR )

    2014

  4. [12]

    Deng, H.; and Li, X. 2022. Anomaly Detection via Reverse Distillation from One-Class Embedding. arXiv:2201.10703

    2022 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:2409.20146

    2024

  6. [14]

    J.; Guan, X.; and Wang, X

    Fan, Y.; He, X.; Yang, D.; Zheng, K.; Kuo, C.-C.; Zheng, Y.; Narayanaraju, S. J.; Guan, X.; and Wang, X. E. 2025. GRIT: Teaching MLLMs to Think with Images. arXiv:2505.15879

    2025 arXiv

  7. [15]

    Gu, Z.; Zhu, B.; Zhu, G.; Chen, Y.; Tang, M.; and Wang, J. 2023. AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models. arXiv:2308.15366

    2023 arXiv

  8. [16]

    Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  9. [17]

    He, H.; Bai, Y.; Zhang, J.; He, Q.; Chen, H.; Gan, Z.; Wang, C.; Li, X.; Tian, G.; and Xie, L. 2024. MambaAD: Exploring State Space Models for Multi-class Unsupervised Anomaly Detection. arXiv preprint arXiv:2404.06564

    2024 arXiv

  10. [18]

    J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W

    Hu, E. J.; Shen, Y.; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2021. LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685

    2021 arXiv

  11. [19]

    P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Radford, A.; et al

    Hurst, A.; Lerer, A.; Goucher, A. P.; Perelman, A.; Ramesh, A.; Clark, A.; Ostrow, A.; Welihinda, A.; Hayes, A.; Radford, A.; et al. 2024. GPT-4o System Card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  12. [20]

    Jaech, A.; Kalai, A.; Lerer, A.; Richardson, A.; El-Kishky, A.; Low, A.; Helyar, A.; Madry, A.; Beutel, A.; Carney, A.; et al. 2024. Openai o1 system card. arXiv preprint arXiv:2412.16720

    2024 arXiv

  13. [21]

    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

  14. [22]

    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

  15. [24]

    Jiang, X.; Li, J.; Deng, H.; Liu, Y.; Gao, B.-B.; Zhou, Y.; Li, J.; Wang, C.; and Zheng, F. 2025. MMAD: A Comprehensive Benchmark for Multimodal Large Language Models in Industrial Anomaly Detection. arXiv:2410.09453

    2025 arXiv

  16. [25]

    C.; Lo, W.-Y.; Dollár, P.; and Girshick, R

    Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A. C.; Lo, W.-Y.; Dollár, P.; and Girshick, R. 2023. Segment Anything. arXiv:2304.02643

    2023 arXiv

  17. [26]

    S.; Reid, M.; Matsuo, Y.; and Iwasawa, Y

    Kojima, T.; Gu, S. S.; Reid, M.; Matsuo, Y.; and Iwasawa, Y. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems

    2022

  18. [27]

    Lei, J.; Hu, X.; Wang, Y.; and Liu, D. 2023. PyramidFlow: High-Resolution Defect Contrastive Localization Using Pyramid Normalizing Flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 14143--14152

    2023

  19. [28]

    Li, X.; Li, Z.; Kosuga, Y.; and Bian, V. 2025 a . Optimizing Safe and Aligned Language Generation: A Multi-Objective GRPO Approach. arXiv preprint arXiv:2503.21819

    2025 arXiv

  20. [29]

    Li, Y.; Wang, H.; Yuan, S.; Liu, M.; Zhao, D.; Guo, Y.; Xu, C.; Shi, G.; and Zuo, W. 2025 b . Myriad: Large Multimodal Model by Applying Vision Experts for Industrial Anomaly Detection. arXiv:2310.19070

    2025 arXiv

  21. [30]

    Liu, H.; Li, C.; Li, Y.; Li, B.; Zhang, Y.; Shen, S.; and Lee, Y. J. 2024. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge

    2024

  22. [31]

    Liu, Z.; Sun, Z.; Zang, Y.; Dong, X.; Cao, Y.; Duan, H.; Lin, D.; and Wang, J. 2025. Visual-RFT: Visual Reinforcement Fine-Tuning. arXiv preprint arXiv:2503.01785

    2025 arXiv

  23. [32]

    Liu, Z.; Zhou, Y.; Xu, Y.; and Wang, Z. 2023. SimpleNet: A Simple Network for Image Anomaly Detection and Localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20402--20411

    2023

  24. [33]

    Ma, H.; Fu, G.; Luo, Z.; Wu, J.; and Leong, T.-Y. 2025. Exploration by Random Reward Perturbation. arXiv preprint arXiv:2506.08737

    2025 arXiv

  25. [34]

    Madaan, A.; Tandon, N.; Gupta, P.; Hallinan, S.; Gao, L.; Wiegreffe, S.; Alon, U.; Dziri, N.; Prabhumoye, S.; Yang, Y.; et al. 2023. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems

    2023

  26. [35]

    OpenAI. 2023. ChatGPT. https://chat.openai.com/

    2023

  27. [36]

    Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems

    2022

  28. [37]

    Qwen; :; Yang, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Li, C.; Liu, D.; Huang, F.; Wei, H.; Lin, H.; Yang, J.; Tu, J.; Zhang, J.; Yang, J.; Yang, J.; Zhou, J.; Lin, J.; Dang, K.; Lu, K.; Bao, K.; Yang, K.; Yu, L.; Li, M.; Xue, M.; Zhang, P.; Zhu, Q.; Men, R.; Lin,...

    2025 arXiv

  29. [38]

    Roth, K.; Pemula, L.; Zepeda, J.; Schölkopf, B.; Brox, T.; and Gehler, P. 2022. Towards Total Recall in Industrial Anomaly Detection. arXiv:2106.08265

    2022 arXiv

  30. [39]

    D.; Perkins, N

    Ruopp, M. D.; Perkins, N. J.; Whitcomb, B. W.; and Schisterman, E. F. 2008. Youden Index and optimal cut-point estimated from observations affected by a lower limit of detection. Biometrical Journal: Journal of Mathematical Methods in Biosciences, 50(3): 419--430

    2008

  31. [40]

    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

  32. [41]

    V.; Zhou, D.; et al

    Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Xia, F.; Chi, E.; Le, Q. V.; Zhou, D.; et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems

    2022

  33. [42]

    Wieler, M.; and Hahn, T. 2007. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, volume 6, 11

    2007

  34. [43]

    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

  35. [44]

    Zhang, J.; He, H.; Gan, Z.; He, Q.; Cai, Y.; Xue, Z.; Wang, Y.; Wang, C.; Xie, L.; and Liu, Y. 2024 b . ADer: A Comprehensive Benchmark for Multi-class Visual Anomaly Detection. arXiv preprint arXiv:2406.03262

    2024 arXiv

  36. [45]

    Zhang, X.; Xu, M.; and Zhou, X. 2024. RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection

    2024

  37. [46]

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

    2025 arXiv

  38. [47]

    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

  39. [48]

    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

  40. [49]

    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.