MVX Benchmark
Overview
MVX Benchmark is a dataset designed to evaluate multi-video reasoning capabilities of modern multimodal large language models (MLLMs).
Most existing video benchmarks evaluate models on single-video understanding, where the model processes only one video at a time. However, many real-world scenarios require reasoning across multiple videos, such as comparing actions, matching events, or aggregating information from different video sources.
To address this gap, MVX introduces a benchmark that requires models to jointly analyze multiple videos and perform cross-video reasoning.
The benchmark contains 1,442 samples across 11 tasks, covering a range of perception and reasoning abilities.
Motivation
Current video understanding benchmarks primarily evaluate:
- single-video question answering
- temporal reasoning within one video
- object or action recognition in a single scene
However, real-world reasoning often requires comparing multiple videos simultaneously. For example:
- identifying which video shows the same action as a reference
- comparing events across videos
- aggregating information from multiple sources
MVX is designed to evaluate whether multimodal models can perform such cross-video reasoning tasks.
Benchmark Structure
MVX organizes tasks into three categories:
1. Low-Level Perception
These tasks test basic visual perception abilities across videos.
Examples include:
- object recognition
- counting
- attribute identification
Although simple individually, these tasks become challenging when models must compare multiple videos simultaneously.
2. Cross-Video Matching
These tasks require identifying relationships between different videos.
Examples include:
- action matching
- video similarity
- event correspondence
Models must determine whether different videos depict the same action or event.
3. High-Level Reasoning
These tasks require more complex reasoning across videos.
Examples include:
- cross-video comparison
- multi-video inference
- aggregation reasoning
These tasks test whether models can synthesize information from multiple videos to reach a conclusion.
Dataset Statistics
- Total Samples: 1,442
- Number of Tasks: 11
- Modality: Video + Question
- Output Format: Multiple-choice answers
Each example typically includes:
- a reference video
- multiple candidate videos
- a question requiring cross-video reasoning
- a multiple-choice answer
Example
Below is a simplified example from the benchmark.
Question
Which candidate video shows the same action as the reference video?
Videos
Reference Video
Candidate A
Candidate B
Candidate C
Candidate D
Answer
B
The model must analyze all videos and determine which candidate matches the reference action.
Access Policy
This dataset contains video content and is released under a controlled access policy.
Users requesting access to the dataset must agree to the following conditions:
- The dataset will be used for research purposes only
- Redistribution of the dataset is not permitted
- The dataset will not be used for commercial purposes
- Users must cite the dataset paper when publishing results
Access requests will be manually reviewed and approved.
Intended Use
The dataset is intended for:
- evaluating multimodal large language models
- studying multi-video reasoning
- benchmarking cross-video understanding
The dataset should not be used for commercial applications.
- Downloads last month
- 13