Thinking Beyond Videos:
Unifying Video Reasoning and Deep Research for Open-World Video Agents

Wenqi Liu^1,*, Shijie Ma^2,*, Yunxiao Wang^1,*, Meng Liu¹, Qile Su³, Han Liu⁴,
Bohan Hou⁵, Zeyu Wang¹, Xuanyu Zheng⁶, Changyi Liu⁶, Tianke Zhang⁶, Haonan Fan⁶,
Kaiyu Jiang⁶, Yingxin Li⁶, Jiankang Chen⁶, Xu Wang⁶, Hongyi Fu⁶, Jianxiong Wang⁶,
Bin Wen^6,‡, Tingting Gao⁶, Han Li⁶, Jianhua Yin¹, Yinwei Wei^1,†, Xuemeng Song^7,†

^*Equal contribution, ^†Corresponding author, ^‡Project Leader.
¹Shandong University, ²Institute of Automation, Chinese Academy of Sciences, ³Beihang University, ⁴City University of Hong Kong, ⁵Nanyang Technological University, ⁶Kuaishou Technology, ⁷Southern University of Science and Technology

arXiv Code 🤗 VideoRover Training Data VideoRover-Bench

Coming Soon

More experimental analyses, training code, data, and benchmarks will be released soon.

Interactive Case Study

01:47

Question

The player in this video mentions scoring in the 2006 and 2010 World Cups. What is the year of the third World Cup in which he scored a goal?

Agent trajectory

Ready

1 / 2

BibTeX

@article{liu2026thinking,
  title={Thinking Beyond Videos: Unifying Video Reasoning and Deep Research for Open-World Video Agents},
  author={Liu, Wenqi and Ma, Shijie and Wang, Yunxiao and Liu, Meng and Su, Qile and Liu, Han and Hou, Bohan and Wang, Zeyu and Zheng, Xuanyu and Liu, Changyi and Zhang, Tianke and Fan, Haonan and Jiang, Kaiyu and Li, Yingxin and Chen, Jiankang and Wang, Xu and Fu, Hongyi and Wang, Jianxiong and Wen, Bin and Gao, Tingting and Li, Han and Yin, Jianhua and Wei, Yinwei and Song, Xuemeng},
  journal={arXiv preprint arXiv:2608.23329},
  year={2026}
}

Thinking Beyond Videos:
Unifying Video Reasoning and Deep Research for Open-World Video Agents

Interactive Case Study

Illustration of the agentic pipeline in VideoTemp-o3. Given the video QA pair, it performs on-demand grounding and refines the initial rough segment. Finally, it produces a reliable answer grounded in the pertinent visual evidence.

Abstract

Data Construction

Multi-turn, multi-tool call data curation pipeline.

Data distribution of SFT data (left) and RL data (right).

Experiments

^† denotes reproduced results using official checkpoints. Bold and underline denotes the best and the second best results.

Main results on long video understanding benchmarks.

Results on temporal grounding (left) and video GQA (right) benchmarks.

Performance of temporal grounding (mIoU) and VideoQA (Acc) across various durations in VideoTemp-Bench.

Cases

BibTeX

Thinking Beyond Videos:Unifying Video Reasoning and Deep Research for Open-World Video Agents

Interactive Case Study

Illustration of the agentic pipeline in VideoTemp-o3. Given the video QA pair, it performs on-demand grounding and refines the initial rough segment. Finally, it produces a reliable answer grounded in the pertinent visual evidence.

Abstract

Data Construction

Multi-turn, multi-tool call data curation pipeline.

Data distribution of SFT data (left) and RL data (right).

Experiments

† denotes reproduced results using official checkpoints. Bold and underline denotes the best and the second best results.

Main results on long video understanding benchmarks.

Results on temporal grounding (left) and video GQA (right) benchmarks.

Performance of temporal grounding (mIoU) and VideoQA (Acc) across various durations in VideoTemp-Bench.

Cases

BibTeX

Thinking Beyond Videos:
Unifying Video Reasoning and Deep Research for Open-World Video Agents

^† denotes reproduced results using official checkpoints. Bold and underline denotes the best and the second best results.