SAGE: A Visual Language Model for Anomaly Detection via Fact Enhancement and Entropy-aware Alignment

Dechen Zhan; Donglin Di; Guoxin Zang; Lanshun Nie; Lei Fan; Xue Li; Yang Song

REVIEW 4 major objections 4 minor 80 references

SAGE: A Visual Language Model for Anomaly Detection via Fact Enhancement and Entropy-aware Alignment

T0 review · 4 major / 4 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 The paper claims that a VLM can be adapted to industrial anomaly reasoning by fusing self-generated facts into a test-vs-reference visual comparison and aligning outputs with expert preferences via entropy-aware DPO.

desk verdict Solid engineering contribution with a new preference dataset, but the MPDD QA evaluation has a train/eval separation question that needs answering before the headline numbers are trustworthy. read the letter →

arxiv 2507.07939 v2 pith:A3K2W22Q submitted 2025-07-10 cs.CL

Guoxin Zang , Xue Li , Donglin Di , Lanshun Nie , Dechen Zhan , Yang Song , Lei Fan This is my paper

classification cs.CL

keywords anomalydetectionvision-languagemodelsdirectpreferenceoptimizationfactenhancementindustrialinspectionzero-shotreasoningcomparisonlearningmulti-image

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 claims that a vision-language model can be made to reason about industrial anomalies by pairing each test image with a normal reference image, extracting text facts about the object offline, and fusing those facts into visual attention through a cross-attention layer. It then fine-tunes the model with an entropy-aware preference optimization that selects answer pairs by their uncertainty difference. The resulting system, SAGE, is reported to beat GPT-4o and several open VLMs on the MANTA and MPDD QA benchmarks in both zero-shot and one-shot settings. The paper also introduces AD-PL, a dataset of 28,415 expert-ranked QA instances, and MLE, a GPT-4o-based scoring rubric for identification, localization, and reasoning. The underlying claim is that explicit domain facts plus expert preference alignment close the gap between generic VLMs and industrial inspection needs.

What carries the argument

The load-bearing mechanism is the Fusion Attention layer $\phi_{FA}$ of Eq. (2): fact-sentence embeddings $t_g$ act as queries, the $K$ test patches least similar to the reference (selected by Eq. (1) as the argtop-$K$ of minimum cosine similarity to any reference patch) act as keys, and the full test image patches act as values. This lets each fact sentence pull out visual evidence from the places where test and reference disagree, producing a comparison-enhanced prompt that is concatenated with the test image and question embeddings before the LLM generates the answer. The second mechanism is E-DPO's entropy-difference bias $\eta \Delta H$ inserted into the DPO margin, which tells the optimizer whether the winning/losing pair is a subtle near-tie or an obvious distinction.

What would settle it

A decisive check would be to feed SAGE fact sentences that are deliberately wrong—describing a different defect than the one present—while keeping the image pair and training otherwise unchanged; the SFE pathway is carrying the claimed benefit only if accuracy drops materially compared with the correct-facts condition.

Watch

Extended reading notes

Core claim

SAGE is a two-stage fine-tuning framework built on InternVL2. It treats anomaly detection as comparison-based reasoning: given a test image and a normal reference image, it selects the test patches least similar to any reference patch, then uses a fusion attention layer in which offline-generated fact sentences query those discrepant patches while reading from the full test image. The resulting comparison-enhanced prompt is fed to the LLM alongside the test image and question. The first training stage, Self-Guided Fact Enhancement (SFE), teaches the model to use these fused facts; the second stage, Entropy-aware DPO (E-DPO), scores candidate answers with GPT-4o, selects the adjacent ranked pair with minimal entropy difference and the extreme pair with maximal entropy difference, and biases the DPO margin by that difference. On MANTA the paper reports average accuracy 69.6% zero-shot and 73.5% one-shot; on MPDD QA it reports 73.8% and 74.1%, ahead of GPT-4o and an InternVL2 SFT baseline in average accuracy and reasoning-quality scores.

Load-bearing premise

The load-bearing premise is that the fact sentences generated offline by a pretrained VLM, with no external supervision, are accurate enough to guide visual attention; the paper's own Limitations section grants that inaccurate facts may impair reasoning or misguide attention.

Editorial extensions

If this is right

The comparison-based paradigm removes the need for per-class normal-sample libraries: a single normal reference image at test time is enough for the model to spot deviations in unseen categories.
Fact sentences generated offline become inspectable artifacts, so an operator can audit why the model flagged an item as anomalous.
The entropy-aware pair selection in E-DPO should transfer to other preference-alignment tasks where candidate answers differ in quality by different amounts.
The AD-PL dataset and MLE rubric give later work a training resource and a three-axis evaluation protocol for anomaly reasoning.
Because SAGE outperforms both the base InternVL2 and its SFT variant in the reported evaluations, the two-stage training (SFE then E-DPO) is the source of the gain, not simply supervised fine-tuning.

Reading between the lines

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

Beyond the paper: since SFE facts are generated offline by the same VLM, one could make SFE a closed loop by regenerating facts from SAGE's own predictions at inference, potentially improving rare-anomaly coverage.
If facts are the active ingredient, then the fact-generation prompt matters as much as the fusion; swapping the prompt for a taxonomy of defect types should shift which anomalies are caught.
The entropy-difference heuristic suggests a curriculum: train on high-entropy-difference pairs first, then fine-tune on near-tie pairs; the paper trains jointly, so the curriculum effect is untested.
The comparison-patch selection of Eq. (1) is essentially unsupervised localization and could be evaluated standalone as an anomaly localizer, independent of the LLM.

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.

Desk Editor's Note

read the letter

Colleague,

The short version: this is a real engineering contribution—a comparison-based VLM for industrial anomaly reasoning, with a new 28k preference dataset (AD-PL) and an entropy-aware DPO variant. The ablations are clean and the MANTA numbers look credible. But the MPDD QA results, which are the least auditable part of the paper, come with a train/eval overlap question that the authors need to resolve.

What's actually new: SFE fuses VLM-generated fact sentences with the top-K most dissimilar patches via a fusion attention layer, which is a sensible way to inject domain priors into comparison reasoning. E-DPO uses GPT-4o MLE scores to select preference pairs with max/min entropy differences rather than just taking adjacent ranks; the entropy bias term in the Bradley-Terry objective is a small but original twist. AD-PL is a new resource, and the three-level question taxonomy (standard, identification, diagnostic) is useful. The ablation shows each component adds something: Fact[FE] beats hard concatenation, E-DPO beats vanilla DPO, and the full model is clearly better than the InternVL2 baseline.

Where it goes soft. First, the MPDD QA concern is real. AD-PL includes 448 MPDD QA test images, and the paper never states that those images, questions, or derived preference pairs are excluded from all training and validation. Since MPDD QA is proprietary and only the authors can audit this, the burden is on them to document a disjoint split. If the split is clean, the MPDD QA results stand; if not, Table 1's MPDD QA column is inflated. The paper does say stage-1 training uses MVTec QA and stage-2 uses VisA QA, but that does not rule out MPDD QA images appearing in fact generation, prompt selection, or the final evaluation set. This is a load-bearing documentation gap, not a proof of leakage.

Second, the GPT-4o circularity: the same MLE framework scores responses for E-DPO pair selection and evaluates final generation quality in Table 2. This inflates the reasoning-quality claims, though it does not directly affect the accuracy numbers in Table 1. Still, it needs a caveat or a human-annotated validation set.

Third, minor: no error bars, single runs, and AnomalyGPT is missing from the baselines even though it's the most directly comparable prior VLM for anomaly reasoning.

My take: the central mechanism is plausible and the MANTA results are backed by a reasonable ablation. The MPDD QA results deserve a serious referee who will ask for the split. It's a conditional accept, not a rejection. Worth engaging with.

Referee Report

4 major / 4 minor

Summary. The paper proposes SAGE, a VLM-based framework for industrial anomaly detection and reasoning that operates on test-reference image pairs. It introduces Self-Guided Fact Enhancement (SFE), which fuses text facts generated by a pretrained VLM with visual patch features through a cross-attention layer, and Entropy-aware Direct Preference Optimization (E-DPO), which uses GPT-4o-based Multiscale Logical Evaluation (MLE) scores to select informative preference pairs. The authors also construct AD-PL, a 28,415-instance preference-ranked QA dataset, and propose MLE as a scoring framework. Experiments on MANTA and a proprietary MPDD QA benchmark report accuracy, SBERT, and generated-response quality, claiming consistent superiority over seven baselines in zero- and one-shot settings.

Significance. If the empirical claims hold, SAGE offers a practical recipe for adapting general VLMs to domain-specific anomaly reasoning with limited supervision, and AD-PL is a substantial new resource for preference-based training. The ablations in Table 3 provide useful evidence that both fusion attention and entropy-aware preference selection contribute to the reported gains. However, the main contribution is empirical, and the evaluation has two serious threats: a potential overlap between the AD-PL training material and the MPDD QA evaluation set, and circular use of the same GPT-4o-based MLE judge for both training preferences and final scoring. These issues must be resolved before the claimed superior performance can be accepted. The code-and-data release is a strength, but the proprietary nature of MPDD QA limits independent auditability.

major comments (4)

[§4, §5.1] The MPDD QA evaluation is not auditable as reported. Section 4 states that AD-PL includes 448 MPDD QA test images, while Section 5.1 says only that MVTec QA is used for stage-1 training and VisA QA for stage-2 training; the paper never states whether the MPDD QA subset of AD-PL is excluded from all training, validation, offline fact generation, prompt/hyperparameter selection, and preference-pair construction. Since the reported MPDD QA results are on 2,240 instances from the same proprietary benchmark, any overlap would directly inflate the Table 1 numbers. Please document a disjoint split and release the evaluation set, or a sufficient random sample, for audit.
[§3.3, §5.2] The same GPT-4o MLE framework is used both to construct E-DPO training preferences (Eq. 3) and to score the final generated responses (Table 2). Optimizing the model against this judge and then measuring it with the same judge is circular; the reported MLE scores may reflect reward hacking rather than genuine anomaly-reasoning quality. I recommend reporting a correlation with human ratings, evaluating with an independent judge or explicit scoring criteria that were not used in training, or at minimum ablating the final scores with a held-out judge.
[§4] The dataset is described as 'expert-ranked' in the abstract and Section 3.3, but the construction pipeline in Section 4 says GPT-4o 'generate[s] structured question-answer pairs, each containing four preference-ranked responses,' and no human ranking step for the answers is described. Please clarify who produced the preference order and provide any inter-annotator agreement statistics. Without this clarification, E-DPO is aligning the model to LLM-generated preferences rather than to verified expert preferences.
[§5.2, Table 1] All comparative results are reported as single runs without error bars or significance tests. The headline one-shot MANTA gain over InternVL2(sft) is 2.2 accuracy points (73.5 vs. 71.3); with no variance estimate, this may not be statistically significant. Report means and standard deviations over multiple seeds, and state whether greedy decoding or sampling was used during evaluation.

minor comments (4)

[§3.3, Eq. (4)] The entropy notation is self-referential: Eq. (4) defines H_i for each answer i but sums over all i. H appears to be the entropy of the full candidate set, so please rewrite the definition to avoid an apparent dependence on the index being summed.
[§4] The image counts do not add up: 1,722 + 2,403 + 448 = 4,573, while the text reports 1,788 normal plus 2,775 anomaly images, which is 4,563. Please correct or explain the discrepancy.
[§5.2, Table 2] Table 2 does not include GPT-4o as a baseline for the MLE-based generation scores, even though GPT-4o is both the judge and a baseline in Table 1. Reporting GPT-4o's own MLE scores would help calibrate the scale and the scores' interpretability.
[§5.2, Figure 5] The Acc-G values for the baselines are presented only in a figure with small and overlapping labels. Include Acc-G as a column in Table 1 or provide a clean tabulation so readers can verify the claimed group-level gains.

Circularity Check

2 steps flagged · score 6.0 of 10

GPT-4o MLE serves as both E-DPO training signal and final generation-quality judge; MPDD QA overlap not documented.

other [Sec. 3.3 (E-DPO, Eq. 3) and Sec. 5.1-5.2 (MLE evaluation)]
"We first perform the Score Calculation by leveraging an LLM (e.g., GPT-4o) to compute Multiscale Logical Evaluation (MLE) scores for each ranked answer, serving as a proxy for answer quality. ... we developed an evaluation pipeline using GPT-4o, which assessed model-generated responses and assigned a quantitative score on a scale of 1-5"

E-DPO's training pipeline computes MLE scores with GPT-4o (Eq. 3), uses them to select preference pairs via entropy (Eqs. 4-5), and biases the Bradley-Terry objective with delta-H (Eq. 6). The generation-quality evaluation in Table 2 (and the Score column in Table 3) then uses the same GPT-4o-based MLE pipeline to score SAGE and baselines. The model is therefore tuned with the same judge that measures its generation quality, so the reported reasoning-quality advantage is partly a by-construction alignment to the evaluation metric. This does not directly affect the objective accuracy numbers in Table 1.
other [Sec. 4 (AD-PL construction) and Sec. 5.1 (evaluation setup)]
"Test images are selected from MVTec QA (1,722), VisA QA (2,403), and MPDD QA (448)... We utilized the MVTec QA dataset for stage 1 training and the VisA QA dataset for stage 2 training. ... MPDD QA is an anomaly reasoning dataset derived from MPDD, featuring 2,240 QA instances"

AD-PL, the preference-optimized dataset used for E-DPO training, includes 448 images from MPDD QA, while Table 1 reports results on a 2,240-instance MPDD QA benchmark. The paper only states that the MVTec QA and VisA QA subsets of AD-PL are used for stages 1 and 2; it never documents that the 448 MPDD QA images are disjoint from the 2,240-instance evaluation set, or excluded from offline fact generation and validation or hyperparameter selection. Without a stated disjoint split, this leg of the central claim is not independently auditable and could reflect exposure to evaluation material.

full rationale

SAGE's principal accuracy claims on MANTA are externally grounded: MANTA is a published benchmark, and Table 1 uses answer-level accuracy, not the GPT-4o judge. The MANTA portion therefore does not reduce to the method's inputs. The circularity that does exist is in the generation-quality evaluation: the MLE metric is defined by GPT-4o scores, the same GPT-4o outputs steer E-DPO's preference-pair selection and entropy weighting, and Table 2 then reports MLE scores as evidence of superior reasoning. That is a partial by-construction alignment rather than a fully forced result, so the paper does not warrant the highest circularity scores. A second, undocumented risk is the MPDD QA split: AD-PL contains 448 MPDD QA images and the paper does not demonstrate that the 2,240-instance evaluation set excludes them. The Limitations section candidly notes that SFE's unverified facts may impair reasoning, but that is a correctness risk, not a circularity. No load-bearing self-citation or imported uniqueness theorem was found; the co-authored MANTA benchmark is used as an external dataset, which is not itself circular.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method adds no new physical entities. Its load-bearing assumptions are that a reference image is always available, unsupervised facts are accurate, GPT-4o MLE scores are valid quality proxies used both for training and evaluation, and entropy-difference pair selection is informative. Tuning parameters eta, K, and N are underspecified.

free parameters (3)

eta (η) entropy-difference scaling factor = not reported
Eq. (6) scales the entropy difference in the DPO objective; no value or sensitivity analysis is given, and it directly controls the strength of E-DPO.
K (top-K dissimilar patches) = not reported
Eq. (1) selects K patches with minimum similarity to the reference; Figure 7 studies its effect, but the K used for the final models is not stated in the setup.
N (number of fact sentences) = not reported
Sec. 3.2 generates N fact sentences and Figure 7 shows accuracy improves with more sentences; the final N is not given in the experimental setup.

assumptions (4)

domain assumption A normal reference image is available for every test image at inference time.
SAGE's comparison-based SFE and E-DPO require a paired normal reference image; the evaluation datasets provide one, but operational deployments may not.
domain assumption Offline VLM-generated fact sentences are accurate enough to guide visual attention.
Sec. 3.2 generates facts without supervision; the paper's Limitations states that inaccurate facts may impair reasoning or misguide attention.
ad hoc to paper GPT-4o MLE scores are a valid proxy for anomaly reasoning quality in both training and evaluation.
Eq. (3) uses GPT-4o scores to rank answers for E-DPO, and Sec. 5.2 uses the same scoring procedure to measure final output quality, creating a shared-judge loop.
ad hoc to paper Entropy differences between ranked answers identify the most informative preference pairs.
Eqs. (4)-(5) select minimal and maximal entropy-difference pairs; no experiment validates that this selection criterion improves learning beyond uniform pair selection.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SAGE: A Visual Language Model for Anomaly Detection via Fact Enhancement and Entropy-aware Alignment." pith.science (2026). https://pith.science/paper/A3K2W22Q

@misc{pith2026250707939,
  author       = {Pith},
  title        = {Pith review of: SAGE: A Visual Language Model for Anomaly Detection via Fact Enhancement and Entropy-aware Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A3K2W22Q}},
  note         = {Machine review of arXiv:2507.07939}
}

read the original abstract

While Vision-Language Models (VLMs) have shown promising progress in general multimodal tasks, they often struggle in industrial anomaly detection and reasoning, particularly in delivering interpretable explanations and generalizing to unseen categories. This limitation stems from the inherently domain-specific nature of anomaly detection, which hinders the applicability of existing VLMs in industrial scenarios that require precise, structured, and context-aware analysis. To address these challenges, we propose SAGE, a VLM-based framework that enhances anomaly reasoning through Self-Guided Fact Enhancement (SFE) and Entropy-aware Direct Preference Optimization (E-DPO). SFE integrates domain-specific knowledge into visual reasoning via fact extraction and fusion, while E-DPO aligns model outputs with expert preferences using entropy-aware optimization. Additionally, we introduce AD-PL, a preference-optimized dataset tailored for industrial anomaly reasoning, consisting of 28,415 question-answering instances with expert-ranked responses. To evaluate anomaly reasoning models, we develop Multiscale Logical Evaluation (MLE), a quantitative framework analyzing model logic and consistency. SAGE demonstrates superior performance on industrial anomaly datasets under zero-shot and one-shot settings. The code, model and dataset are available at https://github.com/amoreZgx1n/SAGE.

Figures

Figures reproduced from arXiv: 2507.07939 by the authors.

**Figure 2.** The overview of SAGE. It consists of two training strategies: Self-Guided Fact Enhancement (SFE) and Entropy-aware [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** An example from the AD-PL dataset. Each instance [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Example and overview of AD-PL dataset. 4 AD-PL To support preference learning and enhance the logical consistency and robustness of anomaly reasoning, we construct a new dataset, AD-PL, based on existing benchmarks: MVTec AD [5], VisA [71], and MPDD [33]. We adopt a th…

**Figure 5.** Figure 5: Accuracy results by group on the MANTA dataset [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Qualitative comparison between GPT-4o and SAGE. The appropriate and inappropriate parts of the responses are [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Accuracy and score results on MANTA in the 1-shot [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

80 extracted references · 35 canonical work pages

[1]

Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al
[2]

Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. 2021. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision . 6836–6846

2021
[3]

Jaehyeok Bae, Jae-Han Lee, and Seyun Kim. 2023. Pni: industrial anomaly detec- tion using position and neighborhood information. InProceedings of the IEEE/CVF International Conference on Computer Vision . 6373–6383

work page 2023
[4]

Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-VL: A Frontier Large Vision- Language Model with Versatile Abilities. arXiv preprint arXiv:2308.12966 (2023)

arXiv 2023
[5]

Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. 2019. MVTec AD–A comprehensive real-world dataset for unsupervised anomaly de- tection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9592–9600

2019
[6]

Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. 2020. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4183–4192

work page 2020
[7]

Ralph Allan Bradley and Milton E Terry. 1952. Rank analysis of incomplete block designs: I. The method of paired comparisons.Biometrika 39, 3/4 (1952), 324–345

work page 1952
[8]

Yuxuan Cai, Xinwei He, Dingkang Liang, Ao Tong, and Xiang Bai. 2024. Anomaly Detection by Adapting a pre-trained Vision Language Model. arXiv preprint arXiv:2403.09493 (2024)

work page Pith review arXiv 2024

Show all 80 references

[9]

Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexan- der Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In European conference on computer vision . Springer, 213–229

2020
[10]

Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al . 2024. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Compu...

2024
[11]

Gonzalez, Ion Stoica, and Eric P

Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality

2023
[12]

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Se- bastian Gehrmann, et al. 2023. Palm: Scaling language modeling with pathways. Journal of Machine Learning Research 24, 240 (2023), 1–113

2023
[13]

Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and An- drea Vedaldi. 2014. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition . 3606–3613

2014
[14]

Niv Cohen and Yedid Hoshen. 2020. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357 (2020)

2020 arXiv
[15]

Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier
[16]

Smith, Hannaneh Hajishirzi, Ross Girshick, Ali Farhadi, and Aniruddha Kembhavi

Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, Jiasen Lu, Taira Anderson, Erin Bransom, Kiana Ehsani, Huong Ngo, YenSung Chen, Ajay Patel, Mark Yatskar, Chris Callison-Burch,...
[17]

Hanqiu Deng and Xingyu Li. 2022. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 9737–9746

2022
[18]

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv prepri...

2020 arXiv
[19]

Okwudili M Ezeme, Qusay H Mahmoud, and Akramul Azim. 2019. Dream: deep recursive attentive model for anomaly detection in kernel events. IEEE Access 7 (2019), 18860–18870

2019
[20]

Lei Fan, Yiwen Ding, Dongdong Fan, Yong Wu, Maurice Pagnucco, and Yang Song
[21]

Lei Fan, Yiwen Ding, Maurice Pagnucco, and Yang Song. 2024. Patch-wise augmentation for anomaly detection and localization. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 5425–5429

2024
[22]

Lei Fan, Dongdong Fan, Yiwen Ding, Yong Wu, Donglin Di, Maurice Pagnucco, and Yang Song. 2025. GrainBrain: Multiview Identification and Stratification of Defective Grain Kernels. IEEE Transactions on Industrial Informatics (2025)

2025
[23]

Lei Fan, Dongdong Fan, Zhiguang Hu, Yiwen Ding, Donglin Di, Kai Yi, Maurice Pagnucco, and Yang Song. 2024. MANTA: A Large-Scale Multi-View and Visual- Text Anomaly Detection Dataset for Tiny Objects.arXiv preprint arXiv:2412.04867 (2024)

2024 arXiv
[24]

Lei Fan, Junjie Huang, Donglin Di, Anyang Su, Tianyou Song, Maurice Pag- nucco, and Yang Song. 2024. Salvaging the Overlooked: Leveraging Class- Aware Contrastive Learning for Multi-Class Anomaly Detection. arXiv preprint arXiv:2412.04769 (2024)

2024 arXiv
[25]

Honghao Gao, Wangyang Jiang, Qionghuizi Ran, and Ye Wang. 2024. Vision- language interaction via contrastive learning for surface anomaly detection in consumer electronics manufacturing. IEEE Transactions on Consumer Electronics (2024)

2024
[26]

Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. 2019. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF international...

2019
[27]

Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. 2023. AnomalyGPT: Detecting Industrial Anomalies using Large Vision- Language Models. arXiv preprint arXiv:2308.15366 (2023)

2023 arXiv
[28]

Tanmay Gupta, Amita Kamath, Aniruddha Kembhavi, and Derek Hoiem. 2022. Towards general purpose vision systems: An end-to-end task-agnostic vision- language architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 16399–16409

2022
[29]

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition . 770–778

2016
[30]

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

2024 arXiv
[31]

Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. 2021. OpenCLIP. If you use this software, please cite it as below

2021
[32]

Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichan- dran, and Onkar Dabeer. 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
[33]

Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. [n. d.]. 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 ...

2021
[34]

Xi Jiang, Jian Li, Hanqiu Deng, Yong Liu, Bin-Bin Gao, Yifeng Zhou, Jialin Li, Chengjie Wang, and Feng Zheng. 2024. Mmad: The first-ever comprehensive benchmark for multimodal large language models in industrial anomaly detection. arXiv preprint arXiv:2410.09453 (2024)

2024 arXiv
[35]

Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al
[36]

Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. 2021. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 9664–9674

2021
[37]

Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning . PMLR, 19730–19742

2023
[38]

Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning . PMLR, 12888–12900

2022
[39]

Wenqiao Li, Xiaohao Xu, Yao Gu, Bozhong Zheng, Shenghua Gao, and Yingna Wu. 2024. Towards scalable 3d anomaly detection and localization: A benchmark via 3d anomaly synthesis and a self-supervised learning network. In Proceedings of the IEEE/CVF Conference on Computer Vision a...

2024
[40]

In Proceedings of the IEEE/CVF international conference on computer vision

Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision. 4015–4026
[41]

Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. 2024. LLaVA-NeXT: Improved reasoning, OCR, and world knowl- edge. MM’25, October 27–31, 2025, Dublin, Ireland Guoxin Zang, et al

2024
[42]

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual in- struction tuning. Advances in neural information processing systems 36 (2023), 34892–34916

2023
[43]

Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. 2021. VT-ADL: A vision transformer network for image anomaly detec- tion and localization. In 2021 IEEE 30th International Symposium on Industrial Electronics (ISIE). IEEE, 01–06

2021
[44]

Patrick Pérez, Michel Gangnet, and Andrew Blake. 2023. Poisson image editing. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2 . 577–582

2023
[45]

Yun Liang, Zhiguang Hu, Junjie Huang, Donglin Di, Anyang Su, and Lei Fan
[46]

Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. In Thirty-seventh Conference on Neural Infor- mation Processing Systems

2023
[47]

Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084 (2019)

2019 arXiv
[48]

Tal Reiss, Niv Cohen, Liron Bergman, and Yedid Hoshen. 2021. Panda: Adapting pretrained features for anomaly detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 2806–2814

2021
[49]

Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. 2022. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 14318–14328

2022
[50]

Bernhard Schölkopf, Robert C Williamson, Alex Smola, John Shawe-Taylor, and John Platt. 1999. Support vector method for novelty detection.Advances in neural information processing systems 12 (1999)

1999
[51]

Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

2021
[52]

Xiu Su, Shan You, Jiyang Xie, Mingkai Zheng, Fei Wang, Chen Qian, Changshui Zhang, Xiaogang Wang, and Chang Xu. 2022. ViTAS: Vision transformer archi- tecture search. In European Conference on Computer Vision . Springer, 139–157

2022
[53]

Xiu Su, Shan You, Mingkai Zheng, Fei Wang, Chen Qian, Changshui Zhang, and Chang Xu. 2021. K-shot nas: Learnable weight-sharing for nas with k-shot supernets. In International Conference on Machine Learning . PMLR, 9880–9890

2021
[54]

David MJ Tax and Robert PW Duin. 2004. Support vector data description. Machine learning 54 (2004), 45–66

2004
[55]

Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

2023 arXiv
[56]

Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. 2024. Qwen2-VL: Enhancing Vision-Language Mode...

2024 arXiv
[57]

Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade W Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa R Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. 2...

2022
[58]

Peng Wu, Xuerong Zhou, Guansong Pang, Lingru Zhou, Qingsen Yan, Peng Wang, and Yanning Zhang. 2024. Vadclip: Adapting vision-language models for weakly supervised video anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 6074–6082

2024
[59]

Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, Zhenda Xie, Yu Wu, Kai Hu, Jiawei Wang, Yaofeng Sun, Yukun Li, Yishi Piao, Kang Guan, Aixin Liu, Xin Xie, Yuxiang You, Kai Dong, Xingkai Yu, Haowei Zhang,...

2024
[60]

Zhongze Wu, Hongyan Xu, Yitian Long, Shan You, Xiu Su, Jun Long, Yueyi Luo, and Chang Xu. 2024. Detecting Any instruction-to-answer interaction relationship: Universal Instruction-to-Answer Navigator for Med-VQA. In Forty- first International Conference on Machine Learning

2024
[61]

Jiacong Xu, Shao-Yuan Lo, Bardia Safaei, Vishal M Patel, and Isht Dwivedi. 2025. Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models. arXiv preprint arXiv:2502.07601 (2025)

2025 arXiv
[62]

Hanyue Yang, Zhenfeng Zhu, Chen Lin, Wenjun Hui, Shenghui Wang, and Yao Zhao. 2023. Self-supervised surface defect localization via joint de-anomaly reconstruction and saliency-guided segmentation. IEEE Transactions on Instru- mentation and Measurement 72 (2023), 1–10

2023
[63]

Yue Wang, Jinlong Peng, Jiangning Zhang, Ran Yi, Yabiao Wang, and Chengjie Wang. 2023. Multimodal Industrial Anomaly Detection via Hybrid Fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8032–8041

2023
[64]

Qirui Yang, Fangpu Zhang, Yeying Jin, Qihua Cheng, Pengtao Jiang, Huanjing Yue, and Jingyu Yang. 2025. DSDNet: Raw Domain Demoir\’eing via Dual Color-Space Synergy. arXiv preprint arXiv:2504.15756 (2025)

2025
[65]

Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo
[66]

Yuan Yao, Ao Zhang, Zhengyan Zhang, Zhiyuan Liu, Tat-Seng Chua, and Maosong Sun. 2024. Cpt: Colorful prompt tuning for pre-trained vision-language models. AI Open 5 (2024), 30–38

2024
[67]

Zihao Zeng, Xuyao Huang, Boxiu Li, and Zhijie Deng. 2025. SIFT: Grounding LLM Reasoning in Contexts via Stickers

2025
[68]

Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. 2022. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605 (2022)

2022 arXiv
[69]

Qirui Yang, Qihua Cheng, Huanjing Yue, Le Zhang, Yihao Liu, and Jingyu Yang
[70]

IEEE Transactions on Image Processing (2025)

Learning to see low-light images via feature domain adaptation. IEEE Transactions on Image Processing (2025)

2025
[71]

Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer
[73]

In European Conference on Computer Vision

Follow the rules: reasoning for video anomaly detection with large language models. In European Conference on Computer Vision . Springer, 304–322
[77]

P Xing, Hao Zhang, Joseph E

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. Judging LLM-as-a-judge with MT- Bench and Chatbot Arena

2023
[78]

Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. 2023. Anoma- lyCLIP: Object-agnostic Prompt Learning for Zero-shot Anomaly Detection. In The Twelfth International Conference on Learning Representations

2023
[80]

In European Conference on Computer Vision

Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In European Conference on Computer Vision . Springer, 392–408
[2021]

In International conference on pattern recognition

Padim: a patch distribution modeling framework for anomaly detection and localization. In International conference on pattern recognition . Springer, 475– 489
[2022]

Advances in neural information processing systems 35 (2022), 23716–23736

Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems 35 (2022), 23716–23736

2022
[2023]

In ECAI 2023

Identifying the defective: Detecting damaged grains for cereal appearance inspection. In ECAI 2023. IOS Press, 660–667

2023
[2024]

arXiv preprint arXiv:2409.17146 (2024)

Molmo and PixMo: Open Weights and Open Data for State-of-the-Art Multimodal Models. arXiv preprint arXiv:2409.17146 (2024)

2024 arXiv
[2025]

IEEE Transactions on Instrumentation and Measurement 74 (2025), 1–9

Tocoad: Two-stage contrastive learning for industrial anomaly detection. IEEE Transactions on Instrumentation and Measurement 74 (2025), 1–9

2025

Pith tools

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

[1] [1]

Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al

[2] [2]

Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, and Cordelia Schmid. 2021. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision . 6836–6846

2021

[3] [3]

Jaehyeok Bae, Jae-Han Lee, and Seyun Kim. 2023. Pni: industrial anomaly detec- tion using position and neighborhood information. InProceedings of the IEEE/CVF International Conference on Computer Vision . 6373–6383

work page 2023

[4] [4]

Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-VL: A Frontier Large Vision- Language Model with Versatile Abilities. arXiv preprint arXiv:2308.12966 (2023)

arXiv 2023

[5] [5]

Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. 2019. MVTec AD–A comprehensive real-world dataset for unsupervised anomaly de- tection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 9592–9600

2019

[6] [6]

Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. 2020. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4183–4192

work page 2020

[7] [7]

Ralph Allan Bradley and Milton E Terry. 1952. Rank analysis of incomplete block designs: I. The method of paired comparisons.Biometrika 39, 3/4 (1952), 324–345

work page 1952

[8] [8]

Yuxuan Cai, Xinwei He, Dingkang Liang, Ao Tong, and Xiang Bai. 2024. Anomaly Detection by Adapting a pre-trained Vision Language Model. arXiv preprint arXiv:2403.09493 (2024)

work page Pith review arXiv 2024