HLVid Dataset

Project Page | Paper | GitHub

HLVid (High-resolution, Long-form Video QA) is a benchmark introduced in the paper "Attend Before Attention: Efficient and Scalable Video Understanding via Autoregressive Gazing".

It is designed to evaluate Multi-modal Large Language Models (MLLMs) on long-form, high-resolution video understanding. The benchmark features 5-minute videos at 4K resolution, challenging models to handle significant spatiotemporal redundancy while preserving critical information.

Dataset Details

The dataset contains question-answering pairs based on high-fidelity video content. Each entry in the test split includes:

• question_id: A unique identifier for the sample.
• category: The specific domain or reasoning category of the video/question.
• video_path: The path or reference to the source video file.
• question: The text-based question regarding the video.
• answer: The ground-truth text answer.

Citation

@article{shi2026attend,
  title={Attend Before Attention: Efficient and Scalable Video Understanding via Autoregressive Gazing},
  author={Shi, Baifeng and Fu, Stephanie and Lian, Long and Ye, Hanrong and Eigen, David and Reite, Aaron and Li, Boyi and Kautz, Jan and Han, Song and Chan, David M and others},
  journal={arXiv preprint arXiv:2603.12254},
  year={2026}
}