跳转至

文章背景与核心概要

当前的视觉语言模型（VLM）在视频理解领域取得了显著进展，但绝大多数现有的视频问答（QA）研究和基准仍局限于离线、单轮的静态范式，忽略了真实世界中用户常在AI生成答案时进行打断的动态交互场景。为了填补这一空白，本文正式提出了“中断场景下的在线视频问答（Online Video Question Answering under Interruption）”任务，并推出了首个标准化基准 OVIBench。

OVIBench 将用户中断精细划分为取消（Cancellation）、误触（False Trigger）和纠错（Correction）三种核心类型，支持开放式问答与多项选择评估。为了实现大规模、可复现的测试，作者开发了一套离线模拟协议，并在统一的时间设定下重现生成过程中的中断情况，同时配备了多维度的评估指标体系来衡量模型对中断的理解及响应生成能力。实验证明，该基准能有效区分不同模型的抗干扰与动态交互能力，尤其在处理纠错请求时表现明显。此外，作者还发布了专用于中断感知微调的训练集 OVI-Train，经其微调的模型在 OVIBench 上获得了显著的性能提升，充分验证了该基准与数据设计的有效性。

OVIBench: Benchmarking Online Video Question Answering under Interruption

OVIBench: Benchmarking Online Video Question Answering under Interruption

📌 Summary

📌 Summary

OVIBench 通过突破传统的离线、单轮视频问答（QA）模式，为视觉语言模型（VLM）引入了一种全新的评估范式。认识到真实世界中的用户经常会在AI生成内容的过程中对其进行打断，本研究建立了首个针对中断场景下的在线视频问答（Online Video QA under Interruption）的标准化基准。

OVIBench introduces a novel evaluation paradigm for Vision-Language Models (VLMs) by moving beyond traditional offline, single-round video question answering (QA). Recognizing that real-world users frequently interrupt AI during generation, this work establishes the first standardized benchmark for Online Video QA under Interruption.

该基准将用户中断归纳为三种截然不同的类型： 1. 取消（Cancellation） 2. 误触（False Trigger） 3. 纠错（Correction）

The benchmark categorizes user interruptions into three distinct types: 1. Cancellation 2. False Trigger 3. Correction

为了对模型进行严格测试，作者开发了一套离线模拟协议以及一套多维度的评估指标。此外，他们还发布了 OVI-Train——一个用于中断感知微调的专门训练集，实验证明，经过该数据集微调的模型在该基准上取得了显著的性能提升。

To test models rigorously, the authors developed an offline simulation protocol alongside a multi-dimensional evaluation metric suite. Furthermore, they release OVI-Train, a specialized training dataset for interruption-aware fine-tuning, which demonstrates significant performance gains for fine-tuned models on the benchmark.

📋 Metadata

📋 Metadata

  • arXiv ID: arXiv:2608.22279 [cs.CV]
  • 研究领域: 计算机视觉与模式识别 (cs.CV)；人工智能 (cs.AI)
  • 会议/期刊: EMNLP 2026
  • 提交日期: 2026年8月23日
  • 作者团队:
  • Naiming Liu
  • Zhiheng Wu
  • Shuning Wang
  • Tie Zhang
  • Bowen Liu
  • Tong Wang
  • arXiv ID: arXiv:2608.22279 [cs.CV]
  • Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
  • Conference/Venue: EMNLP 2026
  • Submission Date: August 23, 2026
  • Authors:
  • Naiming Liu
  • Zhiheng Wu
  • Shuning Wang
  • Tie Zhang
  • Bowen Liu
  • Tong Wang

🔍 Abstract

近期的视觉语言模型（VLM）在视频理解方面取得了强劲进展。然而，绝大多数现有的视频QA研究和基准仍遵循离线、单轮的范式，忽视了用户可能在模型生成答案时进行打断的现实交互场景。

Recent vision language models (VLMs) have achieved strong progress in video understanding. However, most existing video QA research and benchmarks still follow an offline, single-round paradigm, paradigm, overlooking realistic interactions where users may interrupt the model during answer generation.

为了填补这一空白，我们正式提出了中断场景下的在线视频问答任务，并推出了 OVIBench——首个用于在该设定下评估 VLM 的标准化基准。OVIBench 将中断分为三种类型：取消、误触和纠错，同时支持开放式和多项选择评估。

To address this gap, we formulate the task of Online Video Question Answering under Interruption and introduce OVIBench, the first standardized benchmark for evaluating VLMs in this setting. OVIBench categorizes interruptions into three types: Cancellation, False Trigger, and Correction, supporting both open-ended and multiple-choice evaluations.

为了实现大规模且可复现的测试，我们开发了一种离线模拟协议，在统一的时间设定下重现生成过程中的中断，并配备了一套多维度的指标体系来评估中断理解和响应生成能力。实验表明，OVIBench 能够有效区分模型处理中断的能力，特别是在遵循纠错请求方面表现尤为明显。

To enable large-scale and reproducible testing, we develop an offline simulation protocol that reproduces interruption during generation under a unified temporal setup, together with a multi-dimensional metric suite for assessing interruption understanding and response generation. Experiments demonstrate that OVIBench effectively distinguishes models' interruption-handling abilities, especially in following correction requests.

最后，我们构建了一个用于中断感知微调的训练集 OVI-Train。在此数据集上微调的模型在 OVIBench 上取得了显著的性能提升，验证了我们的基准和数据设计的有效性。OVIBench、OVI-Train 以及评估代码将向公众开源。

Finally, we construct a train set OVI-Train for interruption-aware fine-tuning. Models fine-tuned on this dataset achieve significant gains on OVIBench, validating the effectiveness of our benchmark and data design. OVIBench, OVI-Train, and the evaluation code will be publicly released.