Organizations: School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China · School of Artificial Intelligence and Robotics, Hunan University, Changsha, China · Department of Computer and Information Science, University of Macau, Macau, China
Industrial anomaly detection demands precise reasoning over fine-grained defect patterns. However, existing multimodal large language models (MLLMs), pretrained on general-domain data, often struggle to capture category-specific anomalies, thereby limiting both detection accuracy and interpretability. To address these limitations, we propose Reason-IAD, a knowledge-guided dynamic latent reasoning framework for explainable industrial anomaly detection. Reason-IAD comprises two core components. First, a retrieval-augmented knowledge module incorporates category-specific textual descriptions into the model input, enabling context-aware reasoning over domain-specific defects. Second, an entropy-driven latent reasoning mechanism conducts iterative exploration within a compact latent space using optimizable latent think tokens, guided by an entropy-based reward that encourages confident and stable predictions. Furthermore, a dynamic visual injection strategy selectively incorporates the most informative image patches into the latent sequence, directing the reasoning process toward regions critical for anomaly detection. Extensive experimental results demonstrate that Reason-IAD consistently outperforms state-of-the-art methods across multiple tasks. The code will be publicly available at https://github.com/chenpeng052/Reason-IAD.
Figures & tables
Figure 1: Comparison between existing reasoning methods and the proposed Reason-IAD. (a) Existing methods conduct reasoning through explicit chains of thought. (b) Reason-IAD retrieves domain-specific knowledge and identifies anomalies via iterative latent reasoning.
Figure 2: Overview of the proposed Reason-IAD. (a) Given a query image, Reason-IAD retrieves the most relevant category-specific descriptions and incorporates them into the model prompt to enhance anomaly awareness. (b) An entropy-guided latent reasoning module iteratively refines latent think tokens and dynamically injects visual evidence to improve reasoning accuracy. (c) Illustration of the iterative latent-space reasoning process, in which reward signals and visual cues progressively guide the model toward the final prediction.
Anomaly
Defect
Object
Type
Model
Scale
Discrimination
Classification
Localization
Description
Analysis
Classification
Analysis
Average
Random
Random Chance
-
50.00
25.00
25.00
25.00
25.00
25.00
25.00
28.57
Human (expert)
-
95.24
75.00
92.31
83.33
94.20
86.11
80.37
86.65
Human
Human (ordinary)
-
86.90
66.25
85.58
71.25
81.52
89.58
69.72
78.69
Claude-3.5-sonnet
-
60.14
60.14
48.81
67.13
79.11
85.19
79.83
68.36
Gemini-2.5-pro
-
83.07
73.86
67.20
79.97
86.27
94.88
83.08
81.19
Table 1: Performance comparison of commercial and open-source MLLMs on the MMAD benchmark under the standard one-shot setting. The best results are highlighted in bold, and all metrics are reported as accuracy (%).
Type
Model
Scale
Anomaly
Defect
Object
Average
Discrimination
Classification
Localization
Description
Analysis
Classification
Analysis
Open Source
Qwen3-VL
2B
63.27
47.71
52.48
63.48
77.82
92.16
80.03
68.14
Qwen2.5-VL
3B
60.63
47.03
54.44
63.51
79.25
87.48
82.02
67.77
Qwen3-VL
4B
65.83
53.97
58.16
66.35
77.82
91.98
83.38
71.07
Qwen3-VL thinking
4B
64.21
39.61
42.98
48.96
38.32
78.18
64.18
53.78
Qwen2.5-VL
7B
60.35
49.89
57.51
60.41
75.58
93.39
84.80
68.85
Table 2: Performance comparison of open-source MLLMs on the MMAD benchmark under the standard zero-shot setting. The best results are highlighted in bold, and all metrics are reported as accuracy (%).
Figure 5
Module
Datasets
RAKI
EDLR
Vpatch
MVTec
VisA
MVTec-LOCO
GoodsAD
85.75
75.44
67.45
69.38
✓
88.49
77.83
70.76
73.83
✓
✓
89.52
78.58
71.33
74.69
✓
✓
✓
89.84
78.93
71.65
75.00
Table 3: Ablation study of different components with Qwen3-VL-8B under the one-shot setting on four public anomaly detection datasets, all metrics are reported as accuracy (%).
Latent Tokens
Anomaly Discrimination
Defect Localization
Object Classification
2
73.91
61.12
95.64
3
74.68
62.01
96.02
4
75.24
62.78
96.31
5
74.97
62.43
96.18
6
74.63
62.05
95.91
Table 4: Ablation study on the number of latent tokens.
Figure 5: Comparison of model outputs for anomaly detection.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Image Source
Sampled Images
Generated Questions
Object Categories
Defect Categories
MVTec-AD Bergmann et al. (2019)
1691
8338
15
73
MVTec-LOCO-AD Bergmann et al. (2022)
1566
7624
5
89
VisA Zou et al. (2022)
2141
10622
12
67
GoodsAD Zhang et al. (2024)
2968
13088
6
15
SUM
8366
39672
38
244
Appendix
Table 5: Statistics on the composition and quantity of test data.
Model
Scale
Accuracy
Recall
Precision
F1
Human
Human (expert)
-
95.24
94.25
98.89
96.43
Human (ordinary)
-
86.90
87.07
94.35
89.30
Commercial
claude-3.5-sonnet
-
60.14
30.87
76.75
41.92
Gemini-1.5-flash
-
58.58
78.63
67.41
72.40
Gemini-1.5-pro
-
68.63
45.47
86.84
57.60
GPT-4o-mini
-
64.33
65.47
73.04
68.67
Appendix
Table 6: Performance comparison of different MLLMs in Anomaly Discrimination Tasks.
Figure 6: Performance gains of Reason-IAD over baseline models under the one-shot and zero-shot setting.
Model
RAKI
Anomaly
Defect
Object
Discrimination
Classification
Localization
Description
Analysis
Classification
Analysis
Qwen2.5-VL-7B
–
71.79
55.23
58.97
64.54
78.90
92.55
84.66
✓
70.74
67.97
60.82
72.78
82.62
97.17
84.99
Qwen3-VL-4B
–
74.02
59.25
62.74
70.09
79.22
92.47
83.59
✓
75.75
73.72
66.48
76.83
82.33
94.90
84.41
Appendix
Table 7: Performance comparison of different MLLMs with and without RAKI.
Figure 7: A case study of the defect location task on the Visa dataset.
Figure 8: A case study of the anomaly discrimination task on the MVTec-AD dataset.
Figure 9: A case study of the defect location task on the Visa dataset.
Figure 10: A case study of the defect description task on the GoodsAD dataset.
Figure 11: A case study of the object classification task on the MVTec-AD dataset.
Figure 12: A case study of the anomaly discrimination task on the MVTec-AD dataset.
Figure 13: A case study of the defect description task on the GoodsAD dataset.
Figure 14: A failure case study of the defect location task on the MVTec-AD dataset.
Figure 15: A failure case study of the anomaly discrimination task on the MVTec-LOCO dataset.
Department of Computer Science and Engineering, Southern University of Science and Technology, Shenzhen, China · Macau University of Science and Technology, Macau, China · Tencent YouTu Lab, Shenzhen and Shanghai, China +3