File size: 3,264 Bytes
b8a078c 728e348 b8a078c 7744f46 b8a078c 7744f46 b8a078c 728e348 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 | ---
dataset_info:
features:
- name: video_source
dtype: string
- name: video_id
dtype: string
- name: duration_sec
dtype: float64
- name: fps
dtype: float64
- name: question_id
dtype: string
- name: question
dtype: string
- name: choices
sequence: string
- name: correct_answer
dtype: string
- name: time_reference
sequence: float64
- name: question_type
dtype: string
- name: question_time
dtype: float64
splits:
- name: train
num_bytes: 291464
num_examples: 900
download_size: 98308
dataset_size: 291464
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
task_categories:
- video-text-to-text
---
<div align="center">
<h2>
RIVER: A Real-Time Interaction Benchmark for Video LLMs
</h2>
<img src="https://github.com/OpenGVLab/RIVER/raw/master/assets/RIVER%20logo.png" width="80" alt="RIVER logo">
[Yansong Shi<sup>*</sup>](https://scholar.google.com/citations?user=R7J57vQAAAAJ),
[Qingsong Zhao<sup>*</sup>](https://scholar.google.com/citations?user=ux-dlywAAAAJ),
[Tianxiang Jiang<sup>*</sup>](https://github.com/Arsiuuu),
[Xiangyu Zeng](https://scholar.google.com/citations?user=jS13DXkAAAAJ&hl),
[Yi Wang](https://scholar.google.com/citations?user=Xm2M8UwAAAAJ),
[Limin Wang<sup>†</sup>](https://scholar.google.com/citations?user=HEuN8PcAAAAJ)
[[💻 GitHub]](https://github.com/OpenGVLab/RIVER),
[[🤗 Dataset on HF]](https://huggingface.co/datasets/OpenGVLab/RIVER),
[[📄 ArXiv]](https://arxiv.org/abs/2603.03985)
</div>
## Introduction
This project introduces **RIVER Bench**, designed to evaluate the real-time interactive capabilities of Video Large Language Models through streaming video perception, featuring novel tasks for memory, live-perception, and proactive response.
[](https://github.com/OpenGVLab/RIVER/raw/master/assets/river.jpg)
Based on the frequency and timing of reference events, questions, and answers, we further categorize online interaction tasks into four distinct subclasses, as visually depicted in the figure. For the Retro-Memory, the clue is drawn from the past; for the live-Perception, it comes from the present—both demand an immediate response. For the Pro-Response task, Video LLMs need to wait until the corresponding clue appears and then respond as quickly as possible.
## Dataset Preparation
|Dataset |URL|
|--------------|---|
|LongVideoBench|https://github.com/longvideobench/LongVideoBench|
|Vript-RR |https://github.com/mutonix/Vript|
|LVBench |https://github.com/zai-org/LVBench|
|Ego4D |https://github.com/facebookresearch/Ego4d|
|QVHighlights |https://github.com/jayleicn/moment_detr|
## Citation
If you find this project useful in your research, please consider cite:
```BibTeX
@misc{shi2026riverrealtimeinteractionbenchmark,
title={RIVER: A Real-Time Interaction Benchmark for Video LLMs},
author={Yansong Shi and Qingsong Zhao and Tianxiang Jiang and Xiangyu Zeng and Yi Wang and Limin Wang},
year={2026},
eprint={2603.03985},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.03985},
}
``` |