RIVER: A Real-Time Interaction Benchmark for Video LLMs
收藏资源简介:
该项目介绍了RIVER Bench,旨在通过流式视频感知评估视频大型语言模型的实时交互能力,具有记忆、实时感知和主动响应等新任务。根据参考事件、问题和答案的频率和时间,我们进一步将在线交互任务分为四个不同的子类。对于Retro-Memory,线索来自过去;对于live-Perception,线索来自现在——两者都要求立即响应。对于Pro-Response任务,视频大型语言模型需要等待相应的线索出现,然后尽快响应。
This project presents RIVER Bench, a benchmark developed to assess the real-time interactive capabilities of video large language models (video LLMs) through streaming video perception. This benchmark features novel tasks including memory, real-time perception, and proactive response. Based on the frequency and timing of reference events, questions, and answers, we further classify online interactive tasks into four distinct subclasses. For the Retro-Memory subclass, cues originate from past content; for the Live-Perception subclass, cues derive from the current moment—both demand immediate responses. For the Pro-Response task, video LLMs are required to await the corresponding cues before responding as promptly as possible.
RIVER 数据集概述
数据集基本信息
- 数据集名称:RIVER: A Real-Time Interaction Benchmark for Video LLMs
- 核心目标:评估视频大语言模型(Video LLMs)的实时交互能力,重点关注流式视频感知。
- 任务类型:包含记忆、实时感知和主动响应等新型任务。
任务分类
根据参考事件、问题和答案的频率与时间,在线交互任务被划分为四个子类:
- 回溯记忆:线索来自过去。
- 实时感知:线索来自当下。
- 主动响应:视频大语言模型需要等待对应线索出现,并尽快响应。
数据来源
数据集整合了以下多个公开数据集:
- 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
访问与引用
- GitHub仓库:https://github.com/OpenGVLab/RIVER
- Hugging Face数据集:https://huggingface.co/datasets/nanamma/RIVER
- 论文地址:https://arxiv.org/abs/2603.03985
- 引用格式: 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}, }




