--- license: mit datasets: - array/SAT language: - en metrics: - accuracy base_model: - Qwen/Qwen2-VL-2B tags: - r1 pipeline_tag: image-text-to-text --- # VisualThinker-R1-Zero

Paper Link👁️

## 🚀 Introduction The recent DeepSeek-R1 demonstrated how reinforcement learning with simple rule-based reward can enable autonomous development of complex reasoning in large language models, characterized by the "aha moment", in which the model manifest self-reflection and increased response length during training. However, attempts to extend this success to multimodal reasoning often failed to reproduce these key characteristics. In this report, we present the first successful replication of these emergent characteristics for multimodal reasoning on only a non-SFT 2B model. Starting with Qwen2-VL-2B and applying reinforcement learning directly on the SAT dataset, our model achieves 59.47% accuracy on CVBench, outperforming the base model by approximately ~30% and exceeding both SFT setting by ~2%. In addition, we share our failed attempts and insights in attempting to achieve R1-like reasoning using RL with instruct models, aiming to shed light on the challenges involved. Our key observations include: (1) applying RL on instruct model often results in trivial reasoning trajectories, and (2) naive length reward are ineffective in eliciting reasoning capabilities. The project code is available at https://github.com/turningpoint-ai/VisualThinker-R1-Zero ## 🔮 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. . . . ``` ## ⚙️ Requirements and Installation * Python >= 3.10 * Pytorch == 2.0.1 * CUDA Version >= 11.7 * Install required packages: ```bash # install transformers pip install git+https://github.com/huggingface/transformers # install qwen-vl utils pip install qwen-vl-utils ``` ## 💻 Model Downloads and Usage ``` from PIL import Image import requests from io import BytesIO from transformers import AutoProcessor, AutoModelForImageTextToText # Load model directly processor = AutoProcessor.from_pretrained("AIcell/reproduce-1200") model = AutoModelForImageTextToText.from_pretrained("AIcell/reproduce-1200" ,torch_dtype="auto", device_map="auto") model.eval() # Prepare image input image_url = "https://multimodal-r1.s3.us-west-1.amazonaws.com/demo_image.jpg" # Prepare text input question = "Considering the relative positions of the sofa and the picture in the image provided, where is the sofa located with respect to the picture? Select from the following choices.\n(A) above or \n(B) below" prompt = f"A conversation between User and Assistant. The user asks a question about the image, and the Assistant solves it. The assistant first thinks about the reasoning process in the mind and then provides the user with the answer.\nUser: {question} \nAssistant: Let me solve this step by step.\n" # Create Message message = [ { "type": "image", "image": image_url, }, {"type": "text", "text": "" + prompt}, ] # Process input response = requests.get(image_url) image = Image.open(BytesIO(response.content)) text = processor.apply_chat_template(message, tokenize=False, add_generation_prompt=True) input = processor( text=text, image=image, padding=True, return_tensors="pt", ) input = input.to("cuda") # Generation of the output generated_ids = model.generate(**input, use_cache=True, max_new_tokens=1024, do_sample=True) generated_ids_trimmed = [ out_ids[len(in_ids):] for in_ids, out_ids in zip(input.input_ids, generated_ids) ] batch_output_text = processor.batch_decode( generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False ) # Get output output_text = batch_output_text[0] print(output_text) ``` ## 🙌 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](https://www.turningpoint-ai.com/)'s homepage for contact information. ## 📖 Acknowledgements We sincerely thank [DeepSeek](https://github.com/deepseek-ai/DeepSeek-R1), [Open-R1](https://github.com/huggingface/open-r1), [QwenVL](https://github.com/QwenLM/Qwen2.5-VL), [Open-R1-Multimodal](https://github.com/EvolvingLMMs-Lab/open-r1-multimodal), [R1-V](https://github.com/Deep-Agent/R1-V), [SAT](https://arxiv.org/abs/2412.07755), and [CV-Bench](https://cambrian-mllm.github.io/) for providing open source resources that laid the foundation of our project. ## 🤝 Contributors Here are the key contributors from [TurningPoint AI](https://www.turningpoint-ai.com/) to this project: [Hengguang Zhou](https://hengguangzhou.github.io/)¹^*, [Xirui Li](https://xirui-li.github.io/)¹^*, [Ruochen Wang](https://ruocwang.github.io/)¹^†, [Minhao Cheng](https://cmhcbb.github.io/)², [Tianyi Zhou](https://tianyizhou.github.io/)³ and [Cho-Jui Hsieh](https://web.cs.ucla.edu/~chohsieh/)¹⁴ ^* Project Leads, ^† Main Advisor ¹University of California, Los Angeles, ²Penn State University, ³University of Maryland and ⁴Google Research ## ✏️ Citation ``` @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}, } ```