Think, Then Verify: A Hypothesis–Verification Multi-Agent Framework for Long Video Understanding (CVPR'2026)

The official implementation of CVPR 2026 paper: Think, Then Verify: A Hypothesis–Verification Multi-Agent Framework for Long Video Understanding.

TL;DR: We propose VideoHV-Agent, a multi-agent framework for long-form VideoQA built on a hypothesis–verification paradigm. Unlike prior single-agent or retrieval-based methods, it decomposes reasoning into explicit, verifiable steps executed by specialized agents.By testing structured hypotheses instead of aggregating noisy evidence, VideoHV-Agent filters spurious information, focuses on decision-relevant observations, and enables transparent agent coordination. This results in more robust, interpretable, and logically consistent video understanding.Experiments on EgoSchema, NextQA, and IntentQA show that VideoHV-Agent achieves state-of-the-art accuracy with strong efficiency and interpretability.

📌 Citation

If you find this paper useful, please consider starring 🌟 this repo and citing 📑 our paper:

@misc{wang2026thinkverifyhypothesisverificationmultiagent,
      title={Think, Then Verify: A Hypothesis-Verification Multi-Agent Framework for Long Video Understanding}, 
      author={Zheng Wang and Haoran Chen and Haoxuan Qin and Zhipeng Wei and Tianwen Qian and Cong Bai},
      year={2026},
      eprint={2603.04977},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.04977}, 
}

🌟 Overview

we propose VideoHV-Agent, a multi-agent framework that recasts long-video understanding as a hypothesis–verification process, thus achieving thinking then verify principle. By shifting from correlation-based retrieval to hypothesis–verification, VideoHV-Agent enables evidence-based, logically consistent, and interpretable reasoning for long-form VideoQA.VideoHV-Agent introduces a novel approach that:

• Hypothesis–Verification paradigm for long-form VideoQA
• Implements this paradigm as a multi-agent framework

😍 Case Study

Qualitative study of event understanding in long videos. VideoHV-Agent uses hypothesis–verification to locate decisive evidence, highlighting its ability to avoid search purposefully and ground conclusions in explicit visual proof:

🔄 Updates

• [2026/04/16]: Code released! 🎉
• [2026/03/5]: Initial version submitted to arXiv.
• [2026/02/21]: Our paper is accepted to CVPR 2026!

📚 License

This repository is released under the Apache License 2.0. This permissive license allows users to freely use, modify, distribute, and sublicense the code while maintaining copyright and license notices.

✨ Acknowledgement

We gratefully acknowledge all authors of VideoAgent-related work for their open-source code and meaningful contributions to the community.