[CVPR 2026] Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm

🎊 News

• [2026.02] 🔥🔥Our work has been accepted by CVPR 2026! 🎉🎉🎉
• [2025.11] Our paper "Thinking with Video: Video Generation as a Promising Multimodal Reasoning Paradigm" has been released on arXiv! 📄 [Paper] On HuggingFace, it has achieved "#1 Paper of the Day" and also "#1 Paper of the Month"!
• [2025.11] 🔥We release "minitest" of our VideoThinkBench, including 500 test samples of vision-centric tasks and 250 test samples of text-centric tasks. This "minitest" can save your evaluation cost.
• [2025.12] 🔥We release VideoThinkBench Leaderboard that includes different models.

📜 Brief Introduction

Moving beyond the traditional paradigms of "Thinking with Text" (e.g., Chain-of-Thought) and "Thinking with Images", we propose "Thinking with Video"—a new paradigm that unifies visual and textual reasoning through video generation models. It naturally enables human-like dynamic reasoning through video generation, such as drawing and imagination.

💡 A New Unified Reasoning Paradigm     "Thinking with Video" leverages video generation models to visualize dynamic processes, represent temporal evolution, and embed text within video frames. This approach achieves unified multimodal understanding and generation, overcoming the static constraints of image-based reasoning and the modality separation in traditional approaches.

📊 VideoThinkBench: A Comprehensive Benchmark     We developed VideoThinkBench, the first reasoning benchmark specifically designed for evaluating video generation models. It comprises vision-centric tasks (eyeballing puzzles, visual puzzles, ARC-AGI-2, mazes) that leverage dynamic visual reasoning, and text-centric tasks adapted from established benchmarks (MATH, GSM8K, MMLU, MMMU, etc.) that test text-based reasoning capabilities within generated videos.

🚀 Surpassing VLMs on Several Tasks     Our evaluation shows that Sora-2 demonstrates competitive reasoning capabilities across both categories. Notably, Sora-2 surpasses state-of-the-art vision-language models on several vision-centric tasks, showcasing the unique advantages of dynamic visual reasoning. On text-centric tasks, Sora-2 achieves strong performance including 98.9% on GSM8K, 94.0% on MATH, and 75.5% on MMMU, demonstrating the potential of "Thinking with Video" as a unified multimodal reasoning paradigm.

📌 Contents

• 🛠️ Installation and Dataset Download
• 📚 VideoThinkBench
• 💻 Code and Evaluation
• 📈 Benchmark Results
• 💡 Takeaways
• 🔎 Citation

🛠️ Installation and Dataset Download

1. Clone this repository and navigate to Thinking-with-Video folder

   git clone --recursive https://github.com/tongjingqi/Thinking-with-Video.git
   cd Thinking-with-Video

2. Install dependencies

   conda create -y -n thinking_with_video python==3.12
   conda activate thinking_with_video
   pip install -r requirements.txt

3. Download benchmark datasets from Hugging Face

   hf download --repo-type dataset OpenMOSS-Team/VideoThinkBench --local-dir VideoThinkBench
   cd VideoThinkBench
   
   # upzip the zip datasets under the `Vision-Centric_Reasoning` and `Text-Centric_Reasoning` folders
   bash unzip_dir.sh Vision-Centric_Reasoning
   bash unzip_dir.sh Text-Centric_Reasoning
   # [Note] you can choose to use the minitest version for evaluation
   # bash unzip_dir.sh minitest_Vision-Centric_Reasoning
   # bash unzip_dir.sh minitest_Text-Centric_Reasoning
   
   # check the statistics of the datasets
   python check.py Vision-Centric_Reasoning > vision_centric_stats.txt
   python check.py Text-Centric_Reasoning > text_centric_stats.txt

📚 VideoThinkBench

VideoThinkBench is a comprehensive benchmark for evaluating video generation models' reasoning capabilities, consisting of two main categories:

Vision-Centric Tasks

• Eyeballing Puzzles: Spatial reasoning tasks requiring visual estimation and drawing
• Visual Puzzles: Pattern recognition and visual logic problems
• ARC-AGI-2: Abstract reasoning tasks requiring few-shot learning
• Mazes: Path-finding and navigation challenges

Text-Centric Tasks

Adapted from established benchmarks including:

• Math Reasoning: GSM8K, MATH-500, AIME24, AIME25
• General Knowledge Reasoning: BBH, MMLU, MMLU-Pro, GPQA-diamond, SuperGPQA-easy
• Multimodal Math Reasoning: MathVista, MathVision
• Multimodal Understanding: MMBench, MMMU

Dataset ("minitest"/full test version) is available on Hugging Face.

💻 Code and Evaluation

Vision-Centric Tasks

• Eyeballing Puzzles, Mazes, ARC-AGI-2: VisionCentric/ (submodule)
• Visual Puzzles: visual_puzzles/

Text-Centric Tasks

• All Text-Centric Tasks: TextCentric/

📈 Benchmark Results

Performance Comparison Across All Tasks

The table below summarizes the accuracy (%) of Sora-2 compared with SOTA vision-language models across the tasks in VideoThinkBench (full test):

Category	Task	Sora-2	Gemini 2.5 Pro	GPT5 high	Claude Sonnet 4.5
Vision-Centric	Eyeballing-Point	44.7	27.8	33.6	36.2
	Eyeballing-Line	38.0	21.0	24.0	26.3
	Eyeballing-Shape	34.5	34.5	32.5	50.5
	Visual-Symmetry	81.9	94.9	98.5	80.1
	Visual-Gradient	51.9	83.7	66.7	69.9
	Visual-Compositionality	57.5	67.0	85.0	82.0
	ARC-AGI-2	1.3	1.9	0.5	5.3
	Maze-Square	40.0	0.0	0.0	0.0
	Maze-Hexagon	0.0	0.0	0.0	0.0
	Maze-Labyrinth	0.0	0.0	0.0	0.0
	Average	35.0	33.1	34.1	35.0
Text-Centric	Text-Only Math	68.6	94.8	97.2	90.0
	Text-Only General Knowledge	65.3	84.5	85.2	86.3
	Multimodal Math	61.2	66.7	69.6	65.6
	Multimodal General Knowledge	79.1	83.0	80.6	82.3
	Average	68.6	82.3	83.2	81.1
Overall Average		44.6	47.1	48.1	48.2

Note: For Sora-2: Eyeballing Puzzles use Major Frame evaluation; Text-Centric Reasoning tasks use Audio evaluation results.

Leaderboard on VideoThinkBench (minitest) (or HERE)

Video Generation Models

#	Model	Average	Eyeballing Point	Eyeballing Line	Eyeballing Shape	Visual Symmetry	Visual Gradient	Visual Compositionality	ARC AGI 2	Maze-Square	Maze-Hexagon	Maze-Labyrinth
1	Sora 2	31.6	50	35	25	80	35	53	2.8	35.3	0.0	0.0
2	Veo 3.1	27.7	34	24	30	78	40	70	0.7	0.0	0.0	0.0
3	MiniMax Hailuo 2.3	26.0	37	34	28	73	45	43	0.0	0.0	0.0	0.0
4	doubao-seedance-1-0-pro-250528	12.4	22	24	35	25	10	8	0.0	0.0	0.0	0.0
5	Wan2.2-TI2V-5B	7.5	18	10	20	8	10	8	0.7	0.0	0.0	0.0

Image Generation Models

#	Model	Average	Eyeballing Point	Eyeballing Line	Eyeballing Shape	Visual Symmetry	Visual Gradient	Visual Compositionality	ARC-AGI-2	Maze-Square	Maze-Hexagon	Maze-Labyrinth
1	Nano Banana 2	29.8	24	30	35	85	50	73	0.71	0.0	0.0	0.0
2	Seedream 4.5	24.5	26	16	30	75	35	63	0	0.0	0.0	0.0
3	GPT image 1.5	19.3	24	15	18	38	50	48	0	0.0	0.0	0.0

Vision-Language Models

#	Model	Average	Eyeballing Point	Eyeballing Line	Eyeballing Shape	Visual Symmetry	Visual Gradient	Visual Compositionality	ARC AGI 2	Maze-Square	Maze-Hexagon	Maze-Labyrinth
1	Claude Sonnet 4.5	37.3	40	34	60	75	75	83	5.7	0.0	0.0	0.0
2	Gemini 2.5 Pro	35.6	33	23	40	95	95	68	2.1	0.0	0.0	0.0
3	GPT5 high	35.5	39	30	23	98	80	85	0.0	0.0	0.0	0.0
4	Qwen3-VL-235B-A22B	30.2	24	17	30	93	55	83	0.0	0.0	0.0	0.0
5	Qwen3-VL-32B	29.6	33	21	20	85	55	78	4.1	0.0	0.0	0.0
6	Qwen3-VL-Plus	29.4	32	29	30	90	35	78	0.0	0.0	0.0	0.0

Note:

• "Eyeballing Point/Line/Shape" refer to Point Tasks, Line Tasks and Shape Tasks in Eyeballing Puzzles. The results of video generation models are Major Frame evaluation results.
• "Visual Symmetry/Gradient/Compositionality" refer to the Symmetry Tasks, Gradient Tasks and Compositionality Tasks in Visual Puzzles.

💡 Takeaways

Our systematic evaluation on VideoThinkBench reveals seven key findings:

1. Surpassing VLMs on Eyeballing Puzzles: Sora-2 generally surpasses SOTA VLMs on eyeballing puzzles, exhibiting strong geometric and physical reasoning abilities. It can simulate the extension and reflection of rays and manipulate geometric elements (e.g., points and lines) to support spatial reasoning.

2. Inductive Reasoning on Visual Puzzles: Sora-2's performance is comparable to Claude Sonnet 4.5 on Shape-Drawing puzzles, demonstrating inductive reasoning capabilities. Sora-2 can recognize and apply patterns of color, shape, and size, solving visual puzzles involving symmetry, gradients, and compositionality.

3. Few-Shot Learning Capabilities: Sora-2 is a few-shot learner. On ARC-AGI-2, which requires finding patterns in input-output pairs, while SOTA VLMs achieve less than 5% accuracy, Sora-2 can often make reasonable predictions, although they do not strictly match dataset annotations.

4. Unified Multimodal Reasoning: On text-centric tasks, Sora-2 shows surprising performance on text and multimodal reasoning benchmarks. The video generation model can embed text within video frames, enabling unified multimodal understanding and generation. This demonstrates that "Thinking with Video" is potentially a unified multimodal reasoning paradigm.

5. Improved In-Context Learning with More Examples: Sora-2 achieves better in-context learning by providing more examples. Experiments show that Sora-2 performs better when provided with all examples compared to only one example, revealing an underexplored direction for analyzing and improving the in-context learning abilities of video generation models.

6. Test-Time Scaling with Self-Consistency: Self-consistency can improve Sora-2's performance on verifiable video generation reasoning tasks. This reveals an underexplored direction: test-time scaling in video generation reasoning tasks.

7. Analysis of Capability Source: We systematically analyzed the source of Sora-2's capabilities. Sora-2 maintains performance comparable to the original test set on adapted math problems, reducing the likelihood of test set leakage. However, Sora-2 struggles to generate coherent reasoning processes in videos, even when providing correct final answers. Through comparative experiments with Wan 2.5, we speculate that Sora-2's text-centric reasoning ability originates from its prompt rewriter model.

⚖️ Licenses

This project is licensed under the MIT License - see the LICENSE file for details.

🔎 Citation

If you find our work helpful, please consider citing our paper 📝 and starring us ⭐️!

@article{tong2025thinking,
  title={Thinking with video: Video generation as a promising multimodal reasoning paradigm},
  author={Tong, Jingqi and Mou, Yurong and Li, Hangcheng and Li, Mingzhe and Yang, Yongzhuo and Zhang, Ming and Chen, Qiguang and Liang, Tianyi and Hu, Xiaomeng and Zheng, Yining and others},
  journal={arXiv preprint arXiv:2511.04570},
  year={2025}
}

---

⭐ Star History

---

Made with ❤️ for advancing multimodal reasoning research