🎬 VideoExplorer: Thinking with Video for Long-Form Understanding

🎬 Demo

👉 Introduction

VideoExplorer is a novel framework for long-video understanding that moves beyond single-pass reasoning. Inspired by the "thinking with video" principle, it performs faithful, efficient, and interpretable reasoning by dynamically exploring video content.

🎉 News

2025.10.16 - We released the newest version of VideoDeepResearch called VideoExplorer! It's smaller, cheaper, but just as effective in long video understanding. Details refer to our updated paper. ✨

2025.06.10 - We released the first version of VideoDeepResearch. 🎬

🚀 Overview

Long-video understanding is challenging. Existing methods often sacrifice detail by downsampling or rely on task-agnostic representations, limiting their perception.

VideoExplorer solves this by intertwining planning, temporal grounding, and scalable perception into a coherent, iterative loop:

1. Formulates a sub-question.
2. Locates the relevant moments.
3. Performs task-oriented, fine-grained perception.
4. Repeats until the final answer is reached.

💡 Key Features

• Iterative Reasoning: Dynamically explores video content instead of relying on a static context.
• Task-Oriented Perception: Focuses computational resources on relevant moments, enabling scalable analysis.
• Interpretable Trajectories: Each step of the reasoning process is transparent and traceable.

🏛️ Framework & Training

To overcome the lack of LVU training data, we constructed a high-quality dataset using difficulty-adaptive sampling. Our training pipeline consists of:

1. Supervised Trajectory Initialization
2. Trajectory-level Preference Optimization

This two-stage approach encourages adaptive temporal grounding and iterative information integration guided by downstream rewards.

📈 Results

Extensive evaluations on popular long-video benchmarks show that VideoExplorer achieves significant performance advantages over existing baselines, demonstrating its robustness, adaptability, and efficiency.

🚀 Quick Start

1. Clone & Install

# Clone repository
git clone https://github.com/yhy-2000/VideoDeepResearch.git
cd VideoDeepResearch

# Install dependencies
pip install -r requirements.txt

Project Layout:

VideoDeepResearch/
├── requirements.txt              # Python dependencies
├── eval/                         # Code for evaluating benchmarks
├── train/                        # Code for supervised finetuning (SFT) and trajectory-based direct preference optimization (TDPO)
├── asset/                        # Assets used in the demo
├── data/
│   ├── videos/                   # Raw video files
│   ├── clips/                    # Generated video clips
│   ├── dense_frames/             # Extracted key frames
└── README.md                     # This documentation

Launch Demo

Evaluation on Benchmarks

Training

Our training dataset is available at https://huggingface.co/datasets/avery00/VideoExplorer-Dataset/tree/main. To set up:

1. Place dpo_marathon.json in train/LLaMA-Factory-dpo/data.

2. Place the remaining two files in train/LLaMA-Factory-sft/data.

Environment Setting

mv train/LLaMA-Factory-sft train/LLaMA-Factory-main
cd train/LLaMA-Factory-main
pip install -e ".[torch,metrics]" --no-build-isolation
mv train/LLaMA-Factory-main train/LLaMA-Factory-sft

Supervised Finetuning

cd train

# load the right code
mv train/LLaMA-Factory-sft train/LLaMA-Factory-main

# finetuning planner
bash sft_planner.sh

# finetuning temporal grounder
bash sft_temporal_grounding_agent.sh

mv train/LLaMA-Factory-main train/LLaMA-Factory-sft

Trajectory-based Direct Preference Optimization

# load the right code
mv train/LLaMA-Factory-dpo train/LLaMA-Factory-main

# Trajectory-based DPO
bash train/dpo_planner.sh

mv train/LLaMA-Factory-main train/LLaMA-Factory-dpo

📬 Contact

Encounter issues or have questions? Reach out to:

H.Y. Yuan Email: hyyuan@ruc.edu.cn

📄 Citation

If you find this work helpful, please cite our paper:

@misc{yuan2025thinkvideosagenticlongvideo,
      title={Think With Videos For Agentic Long-Video Understanding}, 
      author={Huaying Yuan and Zheng Liu and Junjie Zhou and Hongjin Qian and Yan Shu and Nicu Sebe and Ji-Rong Wen and Zhicheng Dou},
      year={2025},
      eprint={2506.10821},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.10821}, 
}