VisualThinker-R1-Zero: First ever R1-Zero's Aha Moment on just a 2B non-SFT Model

VisualThinker-R1-Zero is a replication of DeepSeek-R1-Zero in visual reasoning. We are the first to successfully observe the emergent “aha moment” and increased response length in visual reasoning on just a 2B non-SFT models.

For more details, please refer to the notion report.

Training dynamics of our VisualThinker-R1-Zero training starting from the Qwen-VL-2B, without SFT or reward models. An aha moment and increasing response length is ever observed at a multimodal model.

🔮 Highlights

1. We are the first to successfully produce the emergent “aha moment” and increased response length for multimodal reasoning on just a non-SFT 2B model.
2. We showed that vision-centric tasks could also benefit from improved reasoning capabilities.

Similar to DeepSeek R1, self reflection behavior is also observed during our RL training on vision-centric reasoning tasks. The model exhibits an emergent ability to rethink and correct its mistakes:

. . .
Therefore, dark brown wooden bed with white blanket is not above the doorway.
But wait! I can think of something else.
Maybe it's just higher than above the doorway, but slightly lower than above the doorway.
. . .

📢 Updates

• 2025-03-16: 🤗We released the model checkpoint at huggingface!
• 2025-02-26: 🔥We share our main findings in this notion blog.
• 2025-02-26: 🔥We release the VisualThinker R1 Zero repo.

💻 Hardware Requirements

* estimated

🧱 Setup

bash setup.sh

🤗 Prepare Dataset

cd src/data/SAT
bash prepare_dataset.sh

🏋️ Training

GRPO Training

To reproduce the multimodal aha moment, run the following code to train the non-SFT model with GRPO on SAT:

cd src/open-r1-multimodal
bash run_grpo_SAT.sh # Adjust open-r1-multimodal/configs/zero3.yaml or zero2.yaml accordingly

SFT Training

To obtain SFT model for comparison, run the following code to train the non-SFT model on SAT:

cd src/open-r1-multimodal
bash run_sft.sh # Adjust open-r1-multimodal/configs/zero3.yaml or zero2.yaml accordingly

📈 Evaluation

CVBench Evaluation

We provide following commands to reproduce our evaluation results on the CVBench. First change to evaluation directory:

cd src/eval 

To evaluate Base + GRPO (VisualThinker R1 Zero) model:

python evaluate_Qwen2_VL_CVBench-base.py --model_path  \
    --bs 8 \
    --use_reasoning_prompt

To evaluate Base model:

python evaluate_Qwen2_VL_CVBench-base.py --model_path  \
    --bs 8 \
    --no-use_reasoning_prompt

To evaluate Instruct + GRPO model:

python evaluate_Qwen2_VL_CVBench.py --model_path  \
    --bs 8 \
    --use_reasoning_prompt

To evaluate Instruct model:

python evaluate_Qwen2_VL_CVBench.py --model_path  \
    --bs 8 \
    --no-use_reasoning_prompt

🔍 Resources

Full experiment log: Upcoming

Models CKPT: 🤗VisualThinker-R1-Zero at huggingface

:coffee: Stay Connected!

We are always open to engaging discussions, collaborations, or even just sharing a virtual coffee. To get in touch or join our team, visit TurningPoint AI's homepage for contact information.

📖 Acknowledgements

We sincerely thank DeepSeek, Open-R1, QwenVL, Open-R1-Multimodal, R1-V, SAT, and CV-Bench for providing open source resources that laid the foundation of our project.

🤝 Contributors

Here are the key contributors from TurningPoint AI to this project:

Hengguang Zhou1* , Xirui Li1* , Ruochen Wang1† , Minhao Cheng2, Tianyi Zhou3 and Cho-Jui Hsieh14

• Project Leads, † Main Advisor 1University of California, Los Angeles, 2Penn State University, 3University of Maryland and 4Google Research

:white_check_mark: Cite

If you find our work useful for your projects, please kindly cite the following BibTeX:

@misc{zhou2025r1zerosahamomentvisual,
      title={R1-Zero's "Aha Moment" in Visual Reasoning on a 2B Non-SFT Model}, 
      author={Hengguang Zhou and Xirui Li and Ruochen Wang and Minhao Cheng and Tianyi Zhou and Cho-Jui Hsieh},
      year={2025},
      eprint={2503.05132},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2503.05132}, 
}