IAD-R1: Universal Industrial Anomaly Detection

• IAD-R1 is a universal post-training framework that leverages vision-language models for industrial anomaly detection via structured reasoning.
• It employs a two-stage pipeline using PA-SFT on the Expert-AD dataset and SC-GRPO with reinforcement learning to enhance anomaly perception and interpretation.
• Empirical results demonstrate significant gains—with up to 86.1% accuracy—validating its robustness across diverse industrial benchmarks and VLM backbones.

IAD-R1 is a universal post-training framework for industrial anomaly detection with vision-LLMs (VLMs). It is designed for zero-shot and one-shot generalization, and is explicitly motivated by a gap between “Anomaly Perception” and “Anomaly Interpretation”: the former denotes noticing suspicious visual patterns, whereas the latter denotes producing coherent, structured reasoning that links visual evidence, anomaly location, anomaly type, and the final anomaly decision. The framework combines Perception Activation Supervised Fine-Tuning (PA-SFT) on a structured chain-of-thought dataset called Expert-AD with Structured Control Group Relative Policy Optimization (SC-GRPO), a GRPO-based reinforcement-learning stage that rewards consistency, answer correctness, localization, and type prediction (Li et al., 7 Aug 2025).

1. Concept and problem setting

IAD-R1 addresses industrial anomaly detection under the practically important settings of zero-shot and one-shot inference. In the formulation described by the authors, the input consists of an industrial image $I_{\text{test}}$, optionally a normal reference image $I_{\text{ref}}$ from the same category in the one-shot setting, and a text prompt $p$. The system must determine whether the image contains a defect or anomaly, and the intended structured output may also specify where the anomaly is and what type it is (Li et al., 7 Aug 2025).

The framework is motivated by several constraints of industrial inspection. Defective samples are scarce, anomalies exhibit large inter-class variance, and traditional methods are often scenario-specific. The paper further argues that, although VLMs have significant generalization capability, their anomaly-detection performance remains limited because expert-assisted pipelines are bottlenecked by external anomaly experts, while end-to-end post-training often lacks high-quality reasoning supervision and suffers from weak RL objectives. In that diagnosis, ordinary supervised fine-tuning can induce “cognitive rigidity,” and coarse RL rewards can produce inconsistency between reasoning and final answer (Li et al., 7 Aug 2025).

IAD-R1 is not introduced as a new backbone architecture. Rather, it is presented as a universal post-training framework applicable to VLMs of different architectures and parameter scales. This places it within the broader R1-style post-training lineage associated with reinforcement-learning-based reasoning specialization. A plausible implication is that IAD-R1 transfers the reasoning-oriented post-training logic of DeepSeek-R1 into industrial anomaly detection, while specializing both supervision and rewards for anomaly reasoning (DeepSeek-AI et al., 22 Jan 2025).

2. Expert-AD and the structured anomaly reasoning schema

The first pillar of IAD-R1 is Expert-AD, a 5.9K-QA-pair industrial anomaly detection dataset described as the first industrial anomaly detection dataset containing high-quality chain-of-thought reasoning. Its image source is Real-IAD, and it is split into 2.9K QA pairs for PA-SFT and 3K QA pairs for SC-GRPO (Li et al., 7 Aug 2025).

Expert-AD is constructed through a five-step pipeline. First, the authors collect reference/query image pairs, normal/abnormal labels, specific anomaly types, coarse-grained type classes, and precise anomaly location information from Real-IAD. Second, Qwen-VL-Max is used to compare the query image and a normal reference image and generate detailed structured descriptions; for anomalous images, highlighted anomaly regions, coarse labels, and location information from masks are also supplied. Third, those descriptions are transformed into structured reasoning chains and fine-grained anomaly types. Fourth, human experts manually filter low-quality, inconsistent, or inaccurate entries. Fifth, the final samples are integrated into standardized tagged formats (Li et al., 7 Aug 2025).

The output schema is asymmetric between normal and anomalous cases. For a normal image, the target output is $O=(T,A)$, where $T$ is the chain-of-thought reasoning and $A$ is the final answer. For an anomalous image, the target is $O=(T,L,t,A)$, where $L$ is anomaly location and $t$ is anomaly type. In tag form, the normal template contains > and <answer>, whereas the anomalous template contains <think>, <location>, <type>, and <answer> (Li et al., 7 Aug 2025).

The paper organizes this supervision through a three-layer reasoning framework. The basic perception layer performs spatial scanning and key component positioning for precise anomaly localization. The knowledge-driven analysis layer combines industrial standards and knowledge to identify anomaly types from appearance integrity, surface quality, and structural integrity. The comprehensive decision layer produces the final anomaly judgment and impact evaluation. This schema is explicitly intended to connect perception, reasoning, and answer generation (Li et al., 7 Aug 2025).

3. Two-stage post-training pipeline

PA-SFT is the first training stage. It is intended to activate anomaly perception, establish reasoning-to-answer correlations, and provide a stable structured starting policy for reinforcement learning. The paper defines the supervised objective as

$$\mathcal{L}_{\text{PA-SFT}} = -\mathbb{E}_{(I,p,O)\sim \mathcal{D}_{\text{Expert-AD}}} \sum_{i=1}^{L}\log \pi_\theta(o_i \mid I,p,o_{<i})$$

where $I_{\text{ref}}$0 denotes the industrial image, text prompt, and target output sequence (Li et al., 7 Aug 2025).

PA-SFT is not presented as optional. The authors report that directly using the base model as the initial policy for RL causes severe reward bias, often collapsing toward prediction of a single class. PA-SFT reduces that collapse risk by supplying structured outputs and basic anomaly awareness before reinforcement learning begins (Li et al., 7 Aug 2025).

SC-GRPO is the second stage. It is described as the stage that enables a capability leap from “Anomaly Perception” to “Anomaly Interpretation.” For each input $I_{\text{ref}}$1, the model samples a group of candidate outputs,

$I_{\text{ref}}$2

and scores them with a structured reward. The group-relative advantage is

$I_{\text{ref}}$3

The PPO-style objective with KL regularization is written as

$I_{\text{ref}}$4

The method uses a reference policy initialized from PA-SFT, and the supplementary algorithm also introduces the importance ratio $I_{\text{ref}}$5 and a clipped surrogate objective of the usual PPO form (Li et al., 7 Aug 2025).

4. Reward design and reinforced consistency

The central design of SC-GRPO is a task-specific reward decomposition:

$I_{\text{ref}}$6

For normal images, only consistency and final-answer correctness are rewarded. For anomalous images, the reward additionally includes location and type accuracy (Li et al., 7 Aug 2025).

The consistency reward $I_{\text{ref}}$7 checks whether the output matches the correct structural template for normal or anomalous cases:

$I_{\text{ref}}$8

where Match is a regex-based pattern check. The answer accuracy reward is a binary correctness signal,

$I_{\text{ref}}$9

The location reward maps textual positions into a $p$0 spatial grid:

$p$1

and the type reward uses a graded semantic matching scheme:

$p$2

This design is explicitly meant to reduce reward sparsity and to align reasoning content, structured fields, and final answer. The paper contrasts it with an “Original” reward that uses only final-answer correctness and reports that naive answer-only RL degrades performance, whereas SC-GRPO improves it (Li et al., 7 Aug 2025).

In this respect, IAD-R1 is closely related to GRPO as used in DeepSeek-R1 (DeepSeek-AI et al., 22 Jan 2025), but its reward structure is specialized to anomaly reasoning rather than general verifiable reasoning. It also differs from AnomalyR1, which uses ROAM-enhanced GRPO for multimodal industrial anomaly detection and treats the task as multimodal QA with localization outputs (Chao et al., 16 Apr 2025). IAD-R1 instead emphasizes consistent reasoning, structured anomaly metadata, and universal VLM post-training.

5. Backbones, evaluation protocol, and empirical results

IAD-R1 is evaluated on seven VLM backbones spanning Qwen and LLaVA families: Qwen2-VL-2B, Qwen2.5-VL-Instruct-3B, Qwen2.5-VL-Instruct-7B, LLaVA-1.5-7B, LLaVA-1.6-8B, LLaVA-OneVision-SI-0.5B, and LLaVA-OneVision-SI-7B (Li et al., 7 Aug 2025). Reported implementation details include full-parameter fine-tuning on 4 Nvidia A100 80G GPUs, bf16 precision, 2 epochs for PA-SFT, and 1 epoch for SC-GRPO (Li et al., 7 Aug 2025).

The evaluation datasets are MVTec-AD, MPDD, VisA, DAGM, DTD, and SDD. The main metric is balanced accuracy, defined as the arithmetic mean of normal-sample accuracy and abnormal-sample accuracy. The paper reports both 0-shot and 1-shot settings, with 0-shot using a test image and prompt, and 1-shot additionally supplying a normal reference image (Li et al., 7 Aug 2025).

Model	Average accuracy	Note
LLaVA-OneVision-SI-0.5B baseline	50.7	Raw 0.5B backbone
IAD-R1(LLaVA-OneVision-SI-0.5B)	83.8	Beats GPT-4.1 and Claude-Sonnet-4
GPT-4.1	78.3	Best commercial comparison in the table
Anomaly-OV (7B comparison)	78.9	Best open-source comparison in the table
IAD-R1(LLaVA-OneVision-SI-7B)	86.1	Best overall reported result

The headline result is that IAD-R1(LLaVA-OneVision-SI-7B) achieves 86.1 average accuracy, with 86.7 on MVTec, 70.9 on MPDD, 78.0 on VisA, 94.8 on DAGM, 96.2 on DTD, and 90.1 on SDD. The paper states that this is $p$3 over Anomaly-OV at 78.9 average and $p$4 over GPT-4.1 at 78.3 average (Li et al., 7 Aug 2025).

The most emphasized small-model result is IAD-R1(LLaVA-OneVision-SI-0.5B), which reaches 83.8 average accuracy, compared with 50.7 for the raw 0.5B backbone. On DAGM, its performance rises from 50.0 to 93.3, which the paper reports as a 43.3-point average-accuracy gain. The paper also notes that IAD-R1(Qwen2.5-VL-Instruct-3B) at 77.4 average outperforms the Qwen2.5-VL-Instruct-72B baseline at 73.2 average, underscoring the parameter efficiency of task-specific post-training (Li et al., 7 Aug 2025).

6. Ablations, significance, and limitations

The ablations attribute most of the gain to the structure of the training data and the two-stage pipeline. In PA-SFT, direct-answer tuning on the same images is weak, whereas Expert-AD reasoning supervision is strong. For example, on LLaVA-OneVision-SI-0.5B, the average goes from 50.4 for the base model to 50.5 with answer-only “Original” tuning, but to 76.1 with Expert-AD. On LLaVA-OneVision-SI-7B, the same comparison is 64.2, 65.4, and 80.4 (Li et al., 7 Aug 2025).

The RL ablations show that SC-GRPO improves upon PA-SFT, while answer-only RL harms performance. For LLaVA-OneVision-SI-7B, the averages are 80.4 for PA-SFT alone, 72.2 for the “Original” answer-only reward, and 81.0 for SC-GRPO. The paper also reports that the full reward set $p$5 performs best overall, that a $p$6 grid is the best location discretization, that fine-grained type labels outperform coarse type labels, and that full-parameter fine-tuning outperforms LoRA (Li et al., 7 Aug 2025).

The broader significance of IAD-R1 lies in its claim that industrial anomaly detection benefits from reasoning specialization rather than only larger generic VLMs. This suggests a general pattern already visible in R1-style work: reinforcement learning is not merely a polishing stage, but a mechanism for aligning extended reasoning with task-specific correctness criteria (DeepSeek-AI et al., 22 Jan 2025). In industrial anomaly detection, IAD-R1 operationalizes that principle through structured anomaly reasoning and anomaly-specific reward design (Li et al., 7 Aug 2025).

The paper identifies one explicit limitation: Expert-AD is still relatively small, which may limit generalization to broader industrial scenarios. It also notes that gains are stronger in 0-shot than in 1-shot, which the authors attribute to training on single-image data. Future work proposed in the paper includes collecting more real industrial data, constructing more chain-of-thought annotations, testing additional backbones and scales, and evaluating on more anomaly datasets (Li et al., 7 Aug 2025).