RobustMAD: Evaluating Real-World Robustness of Multimodal Small Language Models for Deployable Anomaly Detection Assistants

by Anushiya Arunan, Xin Li, Yan Qin, U-Xuan Tan, Nhu Khue Vuong, Xiaoli Li, and Yuen Chau

• Paper: OpenReview / arXiv
• Project page: https://robustmad.github.io/
• Code: https://github.com/en-research/RobustMAD

Introduction

Multimodal industrial anomaly inspection assistants are a critical component of next-generation smart factories, enabling interactive vision–language–based querying. However, multimodal large language models remain impractical for on-site deployment due to prohibitive computational demands and privacy risks from cloud-based inference. Compact multimodal small language models (MSLMs) offer a deployable alternative, yet progress is constrained by the lack of comprehensive robustness analyses and meaningfully challenging benchmarks that reflect real-world industrial conditions.

RobustMAD is a practically realistic benchmark explicitly designed to evaluate real-world robustness through diverse open-ended queries spanning object understanding, anomaly detection, unanswerable problems, and visual quality degradations. Contrary to conventional assumptions, top-performing MSLMs exhibit promising capabilities, surprisingly outperforming even the larger GPT-5 Nano — but they remain far below industrial standards. RobustMAD exposes critical robustness gaps posing significant operational risks, driven by three recurring failure modes: (i) fragile multimodal grounding under fine-grained distinctions or degraded visual conditions, (ii) insufficiently comprehensive responses, and (iii) weak logical grounding on unanswerable or ill-posed queries, leading to hallucinated outputs.

Benchmark overview

RobustMAD is the first deployment-motivated benchmark for evaluating the real-world robustness of MSLMs in industrial anomaly inspection. It captures core challenges encountered in practice, including domain-intensive reasoning, non-standardized or ill-posed user queries, open-ended inspection reporting, and realistic visual quality variations. It comprehensively covers both major types of robustness:

• Knowledge-based robustness — model reasoning under diverse, imperfect, and domain-knowledge–intensive queries.
• Visual quality robustness — model sensitivity to image-quality perturbations common in dynamic assembly lines, such as motion blur and low lighting.

Dataset contents

This repository contains RobustMAD's question–answer annotations only. The underlying images are not redistributed here — see "Image data" below.

File	Description
RobustMAD_MCQ.json	Multiple-choice questions, original images
RobustMAD_MCQ_low.json	Multiple-choice questions, visually perturbed images
RobustMAD_OE.json	Open-ended questions, original images
RobustMAD_OE_low.json	Open-ended questions, visually perturbed images
selected_images_vetted.json	Metadata for the human-vetted image subset used in the benchmark

Image data — download required separately

RobustMAD is built on top of two external industrial image datasets, which you must download yourself and are not included in this repository:

• MVTec-AD: https://www.mvtec.com/company/research/datasets/mvtec-ad
• VisA: https://github.com/amazon-science/spot-diff

After downloading, place them at the repository root as:

• MVTec-AD/
• VisA/

matching the paths used in image_path, reference_image_path, query_txt_path, and reference_txt_path fields below (e.g. MVTec-AD/bottle/test/good/014.png). To generate the visually perturbed (motion blur / low-lighting) images used in the _low splits, see utils/generate_LoQ_images.py in the GitHub repo.

Full evaluation pipeline

Model-specific evaluation scripts (InternVL, Phi, Qwen, MiniCPM, GPT-5 Nano, Gemini 3 Flash) and LLM-judge scoring for open-ended responses are maintained in the GitHub repository. Briefly:

• Multiple-choice questions are scored by accuracy against the given answer letter.
• Open-ended questions are scored using an LLM-as-judge pipeline (utils/judge_OE_MSLM_response.py), requiring an OpenAI API key. Pre-judged responses for all evaluated models are provided in the repo's results/ folder if you'd rather skip re-running the judge.

See the GitHub README for exact commands to reproduce results per model.

License

The annotations in this VQA dataset (questions, answers, choices, metadata) are released under CC BY 4.0, matching our TMLR paper. You are free to use, share, and adapt them, including commercially, with attribution.

This license applies only to the annotations in this VQA dataset. The referenced images remain governed by their original licenses — MVTec-AD (CC BY-NC-SA 4.0) and VisA (see the spot-diff LICENSE) — and you are responsible for complying with those separately when you download and use the images.

Citation

If you find this benchmark useful, please consider citing our paper and giving the GitHub repo a ⭐:

@article{arunan2026robustmad,
  author    = {Arunan, Anushiya and Li, Xin and Qin, Yan and Tan, U-Xuan and Vuong, Nhu Khue and Li, Xiaoli and Yuen, Chau},
  title     = {{RobustMAD}: Evaluating Real-World Robustness of Multimodal Small Language Models for Deployable Anomaly Detection Assistants},
  journal   = {Transactions on Machine Learning Research},
  year      = {2026},
  url       = {https://openreview.net/forum?id=skrA9UYNIZ}
}

Acknowledgements

We sincerely thank the authors of the MVTec-AD and VisA datasets for providing the image data, and the authors of MMAD for the domain-knowledge text.

We also thank the individuals who contributed to human review of the RobustMAD questions and answers, and human validation of the GPT-5 LLM judge — see the full list in the GitHub README.