Reason-IAD: Latent Reasoning for Anomaly Detection

• Reason-IAD is a training-free, knowledge-guided dynamic latent reasoning framework designed for explainable, multimodal industrial anomaly detection.
• It integrates Retrieval-Augmented Knowledge Integration, Entropy-Driven Latent Reasoning, and Dynamic Visual Injection to enhance defect detection, localization, and explanation.
• Empirical evaluations on benchmarks like MMAD demonstrate significant accuracy gains in one-shot and zero-shot settings compared to conventional MLLMs.

Reason-IAD is a training-free, knowledge-guided dynamic latent reasoning framework for explainable industrial anomaly detection. It is formulated for multimodal industrial anomaly detection QA: given an industrial image $I$ and a question $Q$, the model answers whether there is a defect, where it is, what it looks like, or how it affects the product. The framework supports both one-shot inference, in which a reference normal image from the same domain is provided, and zero-shot inference, in which only the query image is available. Its design combines Retrieval-Augmented Knowledge Integration (RAKI), Entropy-Driven Latent Reasoning (EDLR), and Dynamic Visual Injection, and it is evaluated on the MMAD benchmark over MVTec-AD, MVTec-LOCO-AD, VisA, and GoodsAD (Chen et al., 10 Feb 2026).

1. Research setting and motivation

Reason-IAD arises in a phase of industrial anomaly detection research in which classical anomaly scoring and segmentation pipelines have become insufficient for multimodal reasoning and explanation. Real-IAD sharpened this shift by arguing that mainstream benchmarks such as MVTec AD and VisA are small, largely saturated, and mostly single-view, whereas realistic inspection requires larger scale, multiple shooting angles, and product-level decision making. Real-IAD therefore introduced 151,050 high-resolution images over 30 object categories, captured from five cameras, together with sample-level evaluation via S-AUROC and a fully unsupervised setting with anomaly noise ratio $\alpha$ in training (Wang et al., 2024).

Against this background, Reason-IAD targets a more specific problem than unsupervised anomaly scoring. Its stated motivation is that existing MLLMs often fail on industrial anomalies because industrial defects are fine-grained and category-specific, general-domain pretraining lacks industrial knowledge, explicit chain-of-thought is inefficient and unstable, and static visual-text matching is insufficient. The method therefore replaces long textual reasoning chains with a compact latent-space process, while still grounding the model in category-specific industrial descriptions and iteratively revisiting image evidence (Chen et al., 10 Feb 2026).

This placement is significant because it situates Reason-IAD between two established traditions. One tradition is benchmark-driven anomaly detection, exemplified by Real-IAD, with emphasis on acquisition realism, multi-view inspection, and harder evaluation. The other is multimodal explainable anomaly analysis, in which the objective is not only anomaly discrimination but also defect classification, localization, description, and analysis.

2. Task formulation and system organization

The operational setting is multimodal industrial anomaly detection QA. The model receives an image $I$, a question $Q$, and, in the one-shot case, a reference normal image from the same domain. It is evaluated on seven MMAD subtasks: anomaly discrimination, defect classification, defect localization, defect description, defect analysis, object classification, and object analysis. The MMAD test set covers 38 product categories and 244 defect types (Chen et al., 10 Feb 2026).

Reason-IAD is organized around two principal modules and one visual control mechanism: Retrieval-Augmented Knowledge Integration, Entropy-Driven Latent Reasoning, and Dynamic Visual Injection. The high-level pipeline is stated as follows: encode the query image and category descriptions into a shared CLIP space; retrieve the most relevant category-specific knowledge snippets; insert the retrieved knowledge into the prompt; initialize a small set of latent think tokens; iteratively refine those latent tokens in hidden space; at each iteration sample perturbations, compute predictive entropy, use entropy as a reward signal, update latent tokens via policy-gradient style optimization, and dynamically inject the most informative image patches; then decode the final answer (Chen et al., 10 Feb 2026).

A central property of the framework is that it is described as training-free with respect to the base MLLM. It does not require supervised fine-tuning on anomaly labels. However, it still performs inference-time optimization over latent think tokens through reward-driven updates. This distinguishes it from post-training methods that adapt the backbone parameters directly.

The implementation details reported for the framework are specific. The base models are Qwen2.5-VL-7B and Qwen3-VL-2B/4B/8B; the retrieval encoder is CLIP ViT-L/14-336; top-$k=2$ categories are retrieved; the number of latent think tokens is $m=4$; the number of reasoning iterations is $N=10$; the perturbation magnitude is 10%; and experiments use 4 NVIDIA A100 80GB GPUs (Chen et al., 10 Feb 2026).

3. Retrieval-Augmented Knowledge Integration

RAKI introduces explicit category-oriented industrial priors into the prompt. The framework maintains a knowledge repository

$K = \{(c_i, desc_i)\}_{i=1}^{N},$

where $c_i$ is a category label such as cable, breakfast box, screw, or bottle, and $Q$0 is a textual description containing normal visual characteristics, typical defect patterns, and domain cues useful for inspection (Chen et al., 10 Feb 2026).

Retrieval operates by encoding the image with a visual encoder $Q$1 and the category names with a text encoder $Q$2: $Q$3 Similarity is then computed by cosine similarity,

$Q$4

and the top-$Q$5 categories are selected. Their corresponding descriptions are inserted into the prompt (Chen et al., 10 Feb 2026).

The prompt structure explicitly frames the model as an industrial inspector. The reported template includes the form: “You are an expert industrial inspector... The following domain knowledge describes typical defect types and normal object characteristics: \n {Domain knowledge}.” In effect, the MLLM receives the query image, the question, and retrieved domain knowledge text. The intended consequence is context-aware reasoning over domain-specific defects rather than reasoning exclusively from general visual semantics (Chen et al., 10 Feb 2026).

RAKI is important because it converts industrial knowledge from an implicit background assumption into a first-class input. This matters in categories where the same local appearance can be either normal or anomalous depending on the product. The reported examples include settings in which defect classification, defect description, defect analysis, and object classification all benefit from category-specific priors.

4. Entropy-Driven Latent Reasoning and dynamic visual injection

The framework’s main reasoning mechanism is EDLR. Instead of producing a long explicit chain of textual thoughts, Reason-IAD introduces a small set of optimizable latent think tokens,

$Q$6

which are continuous embeddings inserted into the multimodal input sequence. The paper sets $Q$7 and the number of reasoning iterations to $Q$8 (Chen et al., 10 Feb 2026).

At iteration $Q$9, the latent tokens are perturbed with Gaussian noise: $\alpha$0 The paper also gives a local update form,

$\alpha$1

where $\alpha$2 is the learning rate and $\alpha$3 is the expected reward objective (Chen et al., 10 Feb 2026).

The reward is based on predictive entropy. For the predicted answer distribution $\alpha$4 associated with latent token $\alpha$5, entropy is defined as

$\alpha$6

and the global reward is

$\alpha$7

Lower entropy corresponds to higher confidence, so the latent reasoning trajectory is optimized toward confident and stable predictions. The policy-gradient update is written as

$\alpha$8

and, under Gaussian perturbation,

$\alpha$9

This establishes a REINFORCE-style inference-time optimization loop over the latent tokens (Chen et al., 10 Feb 2026).

Dynamic Visual Injection keeps this latent process grounded in image evidence. Visual patches are extracted, and the attention weights of the latent think tokens are used to identify the most relevant patches. Candidate patches $I$0 are sampled according to the attention distribution induced by the current latent sequence, and the update rule is

$I$1

The selected patches are then injected into the latent sequence. The stated intuition is that the model, like a human inspector, revisits suspicious regions when uncertain (Chen et al., 10 Feb 2026).

Taken together, EDLR and Dynamic Visual Injection define a compact reasoning process that is latent, iterative, entropy-regularized, and visually re-grounded at each iteration.

5. Empirical behavior, interpretability, and ablations

Reason-IAD is evaluated on MMAD in both one-shot and zero-shot settings. In the one-shot setting, the strongest reported result is for Qwen3-VL-8B + Reason-IAD, which achieves 79.73% average accuracy. This is reported as +4.92% over the base model and higher than GPT-4o at 74.92%. The paper also states that Reason-IAD outperforms AnomalyR1 by 2.43% average accuracy despite being training-free (Chen et al., 10 Feb 2026).

In the zero-shot setting, reported improvements include +5.05% for Qwen3-VL-4B and +4.53% for Qwen3-VL-8B. The Qwen3-VL-8B configuration is also reported to outperform InternVL2-76B by 7.30%, which the paper presents as evidence that the method improves reasoning more than simply scaling the backbone (Chen et al., 10 Feb 2026).

The ablation study on Qwen3-VL-8B across four datasets attributes gains to each module. The baseline scores are 85.75 / 75.44 / 67.45 / 69.38. Adding RAKI yields 88.49 / 77.83 / 70.76 / 73.83. Adding EDLR yields 89.52 / 78.58 / 71.33 / 74.69. The full model achieves 89.84 / 78.93 / 71.65 / 75.00. The reported trend is that RAKI gives a clear jump, EDLR further improves accuracy, and patch injection yields additional gains (Chen et al., 10 Feb 2026).

Iteration count analysis is reported to show that increasing reasoning iterations consistently improves performance, with 10 iterations fixed as a balance between accuracy and efficiency. Qualitative examples are also emphasized. The paper states that Reason-IAD correctly identifies defects such as holes, scratches, corrosion, bent pins, or displaced caps; localizes likely defect regions; and gives human-readable explanations aligned with visual evidence. The listed examples include hazelnut defect discrimination, candle defect localization, screw damage detection, leather surface defect localization, coca-cola can scratch description, three-core cable classification, and ultrasonic sensor bent-pin analysis (Chen et al., 10 Feb 2026).

These results indicate that the method is evaluated not only as a classifier but also as a reasoning system whose outputs are expected to remain tied to concrete visual cues.

6. Relationship to adjacent reasoning-based IAD methods, limitations, and significance

Reason-IAD belongs to a broader research movement that treats industrial anomaly detection as a reasoning problem rather than only a scoring or segmentation problem. Within this literature, Triad modifies LLaVA-style AnyRes tokenization with an expert-guided region-of-interest tokenizer and manufacturing-driven Chain-of-Thought data, using InstructIAD and CoT-M to connect defects to manufacturing steps (Li et al., 17 Mar 2025). EIAD separates semantic reasoning and mask grounding through an Explainable Defect Detection Module and a Multi-modal Defect Location Module, trained with DDQA built from VISION, PR-REAL, Real-IAD, and BSData (Zhang et al., 18 Mar 2025). LR-IAD uses Qwen2-VL 7B with mask-free > and <answer> outputs and GRPO, treating localization as reasoning-based textual explanation rather than pixel-mask supervision (Zeng et al., 28 Apr 2025). AnomalyR1 applies GRPO with the Reasoned Outcome Alignment Metric and outputs bounding boxes that are later converted to masks by SAM2 (Chao et al., 16 Apr 2025). EMIT uses difficulty-aware GRPO with response resampling and advantage reweighting (Guan et al., 29 Jul 2025). IAD-R1 combines PA-SFT on Expert-AD with SC-GRPO to move from “Anomaly Perception” to “Anomaly Interpretation” (Li et al., 7 Aug 2025). ZSG-IAD extends the agenda further by combining RGB images, sensor images, and 3D point clouds with language-guided two-hop grounding and Executable-Rule GRPO (Chen et al., 20 Apr 2026). PB-IAD, by contrast, emphasizes prompt templates, semantic instructions, and domain user centricity with GPT-4.1 in dynamic manufacturing settings (Hofmann et al., 20 Aug 2025).

Within that landscape, Reason-IAD is distinctive in being explicitly training-free with respect to the base MLLM and in relocating reasoning from textual CoT into an iterative latent process. This suggests a different trade-off: instead of post-training the whole model for industrial reasoning, it performs inference-time optimization over latent think tokens while augmenting the prompt with retrieved industrial knowledge and selectively revisiting image patches.

The practical limitations stated for Reason-IAD are also clear. Retrieval depends on the quality of category labels and descriptions; top-$I$2 retrieval is based on CLIP cosine similarity and may miss nuanced category relations; inference-time latent optimization and patch search add runtime compute; performance still depends on the base MLLM; and the framework does not use explicit pixel-level segmentation supervision, so localization is inferred from reasoning and textual descriptions rather than trained masks (Chen et al., 10 Feb 2026).

The significance of the method lies in its formulation of explainable industrial anomaly detection as knowledge-guided latent reasoning. Rather than choosing between open-ended language reasoning and anomaly-specific visual processing, Reason-IAD combines retrieved category priors, entropy-based latent self-optimization, and iterative visual re-grounding. In the current literature, this makes it a representative example of a broader shift from anomaly detection as isolated visual scoring toward industrial anomaly analysis as multimodal, evidence-seeking, and explicitly reasoned decision making.