MMR-AD: A Large-Scale Multimodal Dataset for Benchmarking General Anomaly Detection with Multimodal Large Language Models

Researchers introduced MMR-AD, a large-scale multimodal dataset designed to benchmark general anomaly detection using Multimodal Large Language Models (MLLMs). The study reveals that current state-of-the-art MLLMs fall short of industrial requirements for anomaly detection, though a proposed baseline model called Anomaly-R1 demonstrates significant improvements through reasoning-based approaches enhanced by reinforcement learning.

The development of MMR-AD addresses a critical gap in AI research: the lack of suitable benchmarks for training and evaluating general anomaly detection models using MLLMs. Traditional anomaly detection datasets remain image-focused and incompatible with the multimodal nature of modern language models, while MLLMs trained on web data lack domain-specific knowledge for industrial applications. This mismatch has hindered progress in a field with substantial real-world implications.

General anomaly detection represents a paradigm shift in industrial AI. Rather than building separate models for each product category or defect type, GAD aims to create universally applicable systems that detect novel anomalies without retraining. This approach dramatically reduces deployment costs and accelerates time-to-market for quality assurance applications across manufacturing, healthcare, and logistics sectors.

The introduction of Anomaly-R1 demonstrates that reasoning-based approaches combined with chain-of-thought training data yield measurable improvements over generalist MLLMs. By incorporating reinforcement learning, the model learns to better articulate anomaly detection rationale—a capability essential for high-stakes industrial environments where explainability matters as much as accuracy.

This research signals growing investment in AI benchmarking infrastructure. As MLLMs become the foundation layer for enterprise applications, standardized datasets become increasingly valuable. Organizations relying on anomaly detection for quality control face pressure to adopt more sophisticated models, potentially creating demand for specialized MLLM variants. However, widespread adoption requires bridging the performance gap between current systems and industrial requirements, a challenge this work directly tackles.

• →MMR-AD is the first comprehensive multimodal dataset specifically designed for benchmarking general anomaly detection with MLLMs
• →Current state-of-the-art MLLMs significantly underperform industrial requirements for anomaly detection tasks
• →Anomaly-R1's reasoning-based approach with reinforcement learning achieves substantial performance improvements over generalist models
• →General anomaly detection addresses the need for universal models that work across novel classes without retraining
• →The gap between web-pretrained MLLMs and domain-specific anomaly detection needs remains a critical research frontier