Artificial Intelligence with Reasoning Ability: Vision-R1 – A New Approach for Multimodal-Logical AI Models

The world of Artificial Intelligence (AI) is developing rapidly. A particularly exciting field is the development of multimodal Large Language Models (MLLMs), which can process not only text but also images and other modalities. A promising approach to improving the logical capabilities of these models is the use of Reinforcement Learning (RL). Vision-R1, a new MLLM, impressively demonstrates how RL can improve reasoning in multimodal contexts.

The Challenge of Multimodal Thinking

Conventional MLLMs have difficulty performing complex thought processes like questioning and reflection in multimodal scenarios. This is primarily due to the lack of high-quality training data that fosters such abilities. Training with RL is therefore challenging, as the models lack a sufficient basis for learning complex thought patterns.

Vision-R1: An Innovative Approach

To overcome this challenge, Vision-R1 pursues a novel approach. First, a high-quality multimodal Chain-of-Thought (CoT) dataset with 200,000 examples was created – the Vision-R1-Cold dataset. This dataset serves as the basis for the initial training of the model and was generated without human annotations. Instead, it uses an existing MLLM and DeepSeek-R1 to extract relevant information through modality bridging and data filtering.

Progressive Thinking Suppression Training (PTST)

Another problem in training MLLMs with RL is so-called "overthinking." To avoid this, the Progressive Thinking Suppression Training (PTST) strategy was developed. This method uses Group Relative Policy Optimization (GRPO) with a special reward function to train the model progressively and promote the ability to learn correct and complex thought processes. This training takes place on a smaller dataset with 10,000 multimodal mathematical tasks.

Impressive Results

The results of Vision-R1 are promising. In various benchmarks for multimodal mathematical reasoning, the model was able to achieve an average improvement of about 6%. Particularly noteworthy is the performance on the established MathVista benchmark, where Vision-R1-7B achieved an accuracy of 73.5%. This is only 0.4% below the leading model, OpenAI O1.

Future Prospects

Vision-R1 demonstrates the potential of RL to improve multimodal thinking in AI models. The publication of the datasets and the code allows the research community to build on these results and achieve further progress in this important area. The development of MLLMs with improved logical capabilities opens up new possibilities for AI applications in various fields, from medical diagnostics to robotics.

The Significance for Mindverse

For Mindverse, a German company specializing in AI-powered content creation, image generation, and research, these developments are of great importance. The advances in the field of multimodal LLMs open up new possibilities for the development of innovative solutions, such as chatbots, voicebots, AI search engines, and knowledge systems. Mindverse can leverage these technologies to offer its customers even more powerful and intelligent AI solutions.

Quellenverzeichnis: Huang, W., Jia, B., Zhai, Z., Cao, S., Ye, Z., Zhao, F., Hu, Y., & Lin, S. (2025). Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models. arXiv preprint arXiv:2503.06749. https://huggingface.co/papers/2503.06749 https://arxiv.org/abs/2501.12948 https://arxiv.org/abs/2502.19634 https://huggingface.co/papers https://www.researchgate.net/publication/389398151_MedVLM-R1_Incentivizing_Medical_Reasoning_Capability_of_Vision-Language_Models_VLMs_via_Reinforcement_Learning https://github.com/mbzuai-oryx/Awesome-LLM-Post-training https://www.linkedin.com/posts/nimritakoul_the-paperdeepseek-r1-incentivizing-reasoning-activity-7289992761389289472-cokF https://paperswithcode.com/paper/r1-onevision-an-open-source-multimodal-large https://fireworks.ai/blog/deepseek-r1-got-eyes https://medium.com/@sahin.samia/deepseek-r1-explained-pioneering-the-next-era-of-reasoning-driven-ai-3eeb5ac4d4a0