📂 UniAR-Bench: Benchmarking Evidence-Grounded Anomaly Reasoning Across Modalities in MLLMs

UniAR-Bench is a comprehensive cross-modal anomaly reasoning benchmark with extensive tasks and credible evaluation. Built from publicly available datasets, UniAR-Bench comprises over 40,199 question-answer pairs covering coarse-to-fine reasoning across video, image and text modalities organized into four shared challenges:

• Global and Referring Anomaly Discrimination
• Fine-Grained Anomaly Classification
• Context Understanding
• Holistic Anomaly Analysis

🔍 Data Usage

We provide pre-processed Video-AR, Image-AR and Text-AR Tracks. All samples are used for evaluation. Origin MCQs are provided in "json" directories with basic sample info, evidence annotation and MCQs. We alse provide .tsv format evaluation input, following the VLMEvalKit data structure.

🎬 Video-AR Track:

• The videos directory contains full video for glob.acc. and prom.acc. evaluation.
• The event_evi directory contains trimmed videos with complete event from full videos for trim.acc. evaluation.
• The frame_evi directory contains extracted frame with evidence annotation for acc.@evi. evaluation.

📝 Text-AR Track:

• The long_gec directory contains fully annotated anomalous spans for all-task evaluation.
• The ref_hades directory contains partially annotated anomalous spans for evaluation excluding glob.acc. and acc.@evi..

Place all track datasets under the benchmark/ directory. The dataset structure is as follows:

benchmark
    ├── ImageAR
    │   ├── images
    │   └── jsons
    ├── TextAR
    │   ├── long_gec
    │   └── ref_hades
    └── VideoAR
        ├── event_evi
        ├── frame_evi
        ├── jsons
        └── videos

We provide detailed usage instructions in our paper and open-source code repository.

🙏 Acknowledgements

This codebase is based on VLMEvalKit.

UniAR-Bench builds upon seven publicly available anomaly detection datasets:

• UCF-Crime
• TAD
• ShanghaiTech
• MVTec AD
• VisA
• GEC WI-LOCNESS
• Hades.

Thanks for open-sourcing!