Towards Explainable Industrial Anomaly Detection via Knowledge-Guided Latent Reasoning

The gist The proposed Reason-IAD framework is a knowledge-guided dynamic latent reasoning framework for explainable industrial anomaly detection.

How it works

Reason-IAD comprises two core components: first, a retrieval-augmented knowledge module that incorporates category-specific textual descriptions into the model input, enabling context-aware reasoning over domain-specific defects and second, an entropy-driven latent reasoning mechanism that 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.

Retrieval-Augmented Knowledge Integration

This module addresses the limitation of models pretrained on general-domain data by incorporating category-specific knowledge to enrich the model input prompt. Specifically, it constructs a category-oriented knowledge repository K = (ci, desci) where desci provides a detailed textual description of the category, encompassing both normal characteristics and defect patterns. The process involves encoding the query image into a visual embedding v and each category label ci into a textual embedding ki, projecting both modalities into a shared semantic space. Semantic alignment is then evaluated by computing the cosine similarity si between the query image and each category label. The top-k categories with the highest scores are retrieved, and their descriptions are incorporated into an enriched contextual representation to provide category-aware guidance for reasoning.

Entropy-Driven Latent Reasoning

The entropy-driven latent reasoning mechanism performs iterative exploration in a compact latent space using optimizable latent think tokens. To facilitate dynamic exploration, a set of latent think tokens Z = (z1, z2,..., zm) are introduced as optimizable vector embeddings and inserted into the original multimodal input sequence. During each iteration n, controlled stochasticity is incorporated by defining the perturbed state as Z'(n) = Z(n) + ξ(n), where ξ(n) ∼ N (0, σ2 I), to encourage diverse latent trajectories.

Entropy-Guided Reward Optimization

The objective is to identify and localize anomalous regions given a query image along with its corresponding textual prompt. The model generates a latent reasoning sequence G = (g1,..., gN) through an autoregressive factorization πθ(GQ, I) = Y N n=1 πθ(gnG<n, Q, I), to support anomaly detection. To regulate uncertainty, an entropy-guided reward mechanism leverages predictive uncertainty by computing the entropy of its probability distribution as H(P(n)i) = −X C j=1 P(n)i(j) log P(n)i(j), where P(n)i denotes the predicted probability distribution over answers by the i-th latent think token at iteration n. The global reward at iteration n is defined as R(Z'(n)) = 1 − (1/m) Xm i=1 H(P(n)i), where higher reward values incentivize the model to reduce uncertainty in its latent reasoning process. This optimization is performed using a REINFORCE-based policy gradient method, updating the latent tokens according to Z'(n) ← Z(n) + η∇Z(n)J(Z(n)), where J(·) denotes the expected reward objective.

Dynamic Visual Injection

Inspired by human cognition, this mechanism selectively incorporates the most informative image patches into the latent sequence alongside the latent think tokens. This is achieved by leveraging the attention weights of the latent think tokens to identify t highly relevant image patches, which are then resampled according to an attention distribution and injected into the latent sequence. The optimal patch set Vbest is updated based on a reward signal R(n) = 1 − (1/m) Xm i=1 H(P(n)i), where Vbest = Vcand, if R > Rbest, otherwise.

Experimental Results and Analysis

Extensive experiments demonstrate that Reason-IAD consistently outperforms existing methods across multiple tasks. On Qwen3-VL-8B, it achieves an average 79.73% accuracy in the one-shot setting, improving 4.92% over the base model and even surpassing the commercial model GPT-4O (74.92%). The ablation study shows that incorporating knowledge guidance consistently improves detection accuracy across all datasets, with accuracy increasing from 85.75% to 88.49% on MVTec when RAKI is included. Qualitative analysis shows that Reason-IAD generates the correct prediction while accurately identifying the key visual evidence of industrial defects.

Conclusion and Limitations

Reason-IAD provides a promising paradigm for training-free, expert-level anomaly detection by integrating a retrieval-augmented knowledge module, an entropy-driven latent reasoning mechanism, and a dynamic visual injection strategy. However, limitations exist; first, the inference efficiency of Reason-IAD is lower compared to single-pass MLLM-based methods due to the multi-step latent reasoning process. Second, performance is partially dependent on the capability of the underlying MLLM to follow retrieval-augmented knowledge and reasoning instructions.

Open Access and Ethics

The paper provides open access to the data and code, with sufficient instructions to reproduce the main experimental results. The work conforms to the NeurIPS Code of Ethics throughout our research process, including data usage, experimental design, and reporting. The paper discusses potential broader impacts of the proposed method on industrial anomaly detection systems and automation processes.

LLM Usage Declaration

The paper utilizes a large language model as part of the reasoning component in the proposed framework, and its usage is clearly described in the method section. The LLM is used only for inference-time module to assist in reasoning for anomaly detection, not for training.

References

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

[17] Huajie Tan, Yuheng Ji, Xiaoshuai Hao, Xiansheng Chen, Pengwei Wang, Zhongyuan Wang, and Shanghang Zhang. Reason-rft: Reinforcement fine-tuning for visual reasoning of vision language models. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

[18] Yuhao Chao, Jie Liu, Jie Tang, and Gangshan Wu. Anomalyr1: A grpo-based end-to-end mllm for industrial anomaly detection. arXiv preprint arXiv:2504.11914, 2025

[37] Peng Chen, Chao Huang, Yunkang Cao, Chengliang Liu, Wenqiang Wang, Mingbo Yang, Li Shen, Wenqi Ren, and Xiaochun Cao. Reason-iad: Knowledge-guided dynamic latent reasoning for explainable industrial anomaly detection. arXiv preprint arXiv:2602.09850

[43] Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space. arXiv preprint arXiv:2412.06769, 2024 <ref:2602.

1. Knowledge-guided retrieval for domain specificity can be extended to handle highly specialized industrial sub-domains by dynamically querying a larger, hierarchical knowledge base instead of just top-k category descriptions. This would enable the system to reason over extremely fine-grained defects that are not well represented in the initial category repository.

2. Entropy-driven latent reasoning can be adapted to incorporate uncertainty estimates from multiple independent model runs, allowing the reward mechanism to penalize predictions based on high epistemic uncertainty across different latent trajectories, leading to more robust decision boundaries.

3. Dynamic visual injection could be enhanced by using a learned attention mechanism over the entire latent sequence instead of just sampling candidate patches based on current token attention weights, allowing for a more holistic integration of relevant visual evidence into the reasoning process.