M4R: Measuring Massive Multimodal Understanding and Reasoning in Open Space

Website · Code · Leaderboard · Dataset · Dataset-Zip · Issue

Project Homepage:

https://open-space-reasoning.github.io/

About the Dataset:

This benchmark includes approximately 2,000 videos and 19,000 human-annotated question-answer pairs, covering a wide range of reasoning tasks (as shown in Figure 1). All annotations were performed by highly educated annotators, each holding at least a master's degree in engineering-related fields such as mathematics or computer science. The dataset features a variety of video lengths, categories, and frame counts, and spans three primary open-space reasoning scenarios: land space, water space, and air space. An overview of the dataset’s characteristics is shown in Figure 2, which illustrates the distributions of video duration, domain coverage, and reasoning styles. During annotation, we first design the hard-level tasks and label each question with the ground-truth answer. Based on these, we then construct the medium and easy tasks. The primary differences between difficulty levels lie in the number and types of answer choices. Details of the annotation procedure and difficulty levels are provided in our paper.

Dataset Format:

Figure 1. A question and answer example: For each open-space reasoning setting, we include three types of video lengths: short, medium, and long. Each video length includes tasks designed to evaluate temporal reasoning, spatial reasoning, and intent reasoning.

Dataset Distribution:

Figure 2. Distribution of video and task properties in the M4R benchmark.

Three Space Settings

Figure 3. Examples of multimodal Understanding and Reasoning in Open-Space Scenarios.

Reasoning Settings:

Figure 4. Examples of reasoning question settings in M4R across three key reasoning types: Temporal Reasoning, which involves understanding event sequences and motion over time; Spatial Reasoning, which focuses on relative positioning and orientation in space; and Intent Reasoning, which evaluates understanding of goal-directed behaviors and decision-making in dynamic environments..

One Example in Land Space Settings:

Figure 5. Land-space traffic accident scenarios for open-space video understanding and reasoning include intersection collisions, urban road accidents, nighttime incidents, rural road accidents, snow-covered road collisions, and freeway accidents.

Download Dataset

You can download the dataset directly from our Hugging Face repository via:

git lfs install
git clone https://huggingface.co/datasets/Open-Space-Reasoning/M4R

If you encounter any issues during the download, we also provide a zipped version for convenience: Download Dataset (ZIP)

Note: If you encounter any issues, please visit our GitHub page, where we provide more information about the project and detailed instructions for downloading and using the datasets.

Citation

If you find the repository useful, please cite the study

@article{gu2025m4r,
  title={Measuring Massive Multi-Modal Understanding and Reasoning in Open Space},
  author={Gu, Shangding and Wang, Xiaohan and Ying, Donghao and Zhao, Haoyu and Yang, Runing and Li, Boyi and Jin, Ming and Pavone, Marco and Yeung-Levy, Serena and Wang, Jun and Song, Dawn and Spanos, Costas},
  journal={Github},
  year={2025}
}